What changed in how organisations work with AI over twelve months — across twenty studies from IBM IBV, PwC, KPMG, Microsoft, BCG and Stanford HAI. Eight dimensions, from the economics of return to employee experience, and what to measure yourself against in each. Every figure links to the report it came from; the “i” button opens the deep dive.
If you read only one paragraph
Over the year AI moved from a cost line to a return line — but the return landed extremely
unevenly. The top of the ROI range is about +250%, the bottom about −20%:
270 percentage points apart. The companies most exposed to AI added 33,5%
in productivity, and the top 20% within them — 163%. The bottleneck moved
from data to architecture and the operating model: in Microsoft's analysis organisational factors account for 67% of the total relative importance in predicting self-reported impact, individual factors for 32%; this is a statistical association, not a causal effect. Governance stopped
being the brake and became the condition for scaling. And the labour market did not collapse —
it split into two tracks, and the harshest of it is happening at the entry
to a profession.
If the year comes down to one shift: in 2025 the main question was “how do we get AI to people”, and in 2026 it is “how do we rebuild processes, metrics and accountability around AI”. The technology became markedly more capable and more autonomous; the main gap now runs not between “has AI / has no AI”, but between organisations that redesign the work and those that stop at new tools.
The limits of this data
This is big business. KPMG surveys companies with revenue of ≥ $1 bn, and the average revenue in IBM's technology-leader sample is $14.4 bn. Small and mid-sized business is almost absent here.
These are not the same companies year over year. Each edition surveys new people, so “26% → 76%” is a comparison of two market snapshots, not the path of one and the same group.
Return is measured mostly by self-report. No source computed ROI from management accounts; independent market series exist only at PwC and Stanford.
The question of the year — how much AI gave back — got answers that are meaningless to average: multiples for some, losses for others, on the same technology in the same year. What follows are seven statements about what stands behind that.
01
A year ago AI was measured by what it cost. Now it is measured by what it returns — and the return has split apart
In 2025 the reports recorded near-universal optimism alongside near-total absence of measurement. Over three years only 25% of AI initiatives delivered the expected ROI, only 16% were scaled to enterprise level. In marketing the picture was worse still: 19% of pilots delivered the expected ROI, 14% were rolled out enterprise-wide. And 68% of CAIOs admitted they launch AI projects even when they cannot assess their impact.
A year on, return is measured more often — and above all its spread became visible.
270 percentage points between the top and the bottomIBM Tech 2026
Extreme AI-ROI values across organisations' portfolios. The lower bound is real losses, not merely a weak result.
The same technology in the same year delivers both a multiple and a real loss: +250% at the top of the distribution and −20% at the bottom.
In 2026 the gap became measurable — and PwC independently confirms the same thing from another direction: companies where AI applies most broadly show productivity growth of 33,5%, but the top 20% within that group — 163%, five times more. Twenty per cent of companies take 74% of the gains.
The figure of the year
163%
productivity growth for the top 20% within the companies where AI applies most broadly — five times morethan the average of their own group (33.5%)
Productivity growth: the extreme quartiles by AI exposurePwC 2026
Cumulative growth in revenue per employee, 2018 → 2025. The unit of observation is the company.
The extreme quartiles of occupations by AI exposure: 33,5% versus 24% cumulative growth in revenue per employee. The unit of observation is the company, not the occupation.
Two different statements must not be confused here. This data does not let you say that “average ROI has risen”: IBM shows a wide spread of return across portfolios, and PwC strong heterogeneity in productivity growth. IBM's lower bound stays negative, and at PwC most of the gain is concentrated in the upper part of the distribution.
The practical conclusion is not that “the same companies keep winning” — annual samples cannot track that — but that the fact of adoption alone is not enough for predictable return. The effect of AI is distributed extremely unevenly.
Two independent methods — IBM's survey and PwC's market data — give a compatible picture: between organisations and within groups of organisations the effect differs enormously. That is stronger than trying to derive a single “average ROI” from them, but weaker than a claim of causality or of the same companies winning.
02
The bottleneck moved: from data to architecture and the operating model
In 2025 data was named as the main constraint. IBM devoted a separate CDO study to it: scattered stores, no shared taxonomy, six to twelve months of data cleaning per initiative.
In 2026 the wording changed. The problem is not that there is no data, but that systems are designed for human decision speed, while AI works at “machine speed” — at a different volume, with different autonomy, at a different decision frequency. 77% of technology leaders say AI adoption is outpacing their ability to govern it, 70% — that teams deploy technology faster than IT can track it.
Microsoft comes at it from the other end: it asks not “are we keeping up”, but what the effect depends on at all — and two thirds of the answer turns out not to be about the person.
What explains the effect of AI: the organisation versus the personMicrosoft WTI 2026
What the effect of AI is made of: 29 factors grouped into three families. 20,000 people surveyed across ten countries.
×2,5
how many times stronger an organisation's AI culture is than the single most significant individual factor
0,68–0,69
R² of the three families of models the result rests on
In Microsoft's model organisational factors carry roughly twice the total relative importance of individual ones. This is an association, not a proven cause: it may be that companies with a strong AI culture are simply better run in everything.
Two sources that barely overlap methodologically — IBM surveys CIOs and CTOs, Microsoft surveys AI users and supplements the survey with M365 telemetry — give a compatible picture from opposite ends of the organisation. That is rare. Looking for the lever only in an individual's skills is not enough. The organisational part — culture, rules, process redesign — is associated with the outcome markedly more strongly than anything relating to the individual. It cannot be bought from a vendor, which is why it is often taken on later. An important qualification: Microsoft speaks of the statistical importance of factors, not of a causal effect.
So the lever is not where people look for it. The organisational part — culture, rules, process redesign — is associated with the outcome markedly more strongly than anything relating to the individual. It cannot be bought from a vendor, and so it is taken on last.
One qualification, without which the conclusion becomes wrong: Microsoft says “explains variance”, not “causes”. It may be that organisations with a strong AI culture were simply better run in everything to begin with, and then the cause lies deeper still.
03
Governance stopped being the brake and became the condition for scaling
KPMG puts it literally: “governance is the condition for scaling”. That is not rhetoric — there is arithmetic behind it. By 2027 enterprises expect an average of 1,661 agents each, 38% more than today; at hundreds of decisions per agent per day that is hundreds of thousands of autonomous decisions daily. No approval committee will digest that.
IBM split organisations into those where control is built into the architecture itself and those where it stayed an approval procedure — and from there all three indicators move the same way.
×16
more agents are deployed by organisations with designed-in control
versus manual governance
×4
smaller share of the AI budget goes on control for them
cheaper, not dearer
+18%
operating margin relative to those who govern manually
a measurable effect in the accounts
This combination changes the very framing of the “governance slows innovation” debate. In IBM's sample organisations with control designed into the architecture simultaneously report greater agent scale, a smaller budget share on control and higher operating margin. That is a strong association, but not an experiment: it cannot prove that architectural control caused the difference. The practical meaning matters anyway — at hundreds and thousands of agents manual approvals scale worse, so control has to be designed as part of the system rather than as a separate committee after the fact.
×16 at a quarter of the budget share is not a process improvement, it is a change in the class of system. Manual gates have linear cost: every new agent needs a new approval. Designed-in control has almost fixed cost. At 1,661 agents the difference between those two curves becomes insurmountable.
04
The labour market did not collapse — it split into two tracks
The same fault line ran outside organisations too — across the labour market, where it is visible even without surveys.
PwC 2025 gave an optimistic overall picture: jobs and wages are growing in practically every occupation exposed to AI, including the most automatable ones. PwC 2026 takes that overall picture apart.
Three types of AI impact on an occupation — and the whole job market broken down by them.
Two of these three groups move in different directions, and they diverge not on one measure but on three at once.
The two tracks diverge on three measures at oncePwC 2026
Occupations where expertise requirements rise also grow faster in skills, in wages and in the number of openings.
Occupations with rising expertise requirements outpace those with a falling entry bar on all three measures: skills 68% versus 33%, openings 39% versus 17%, wages 37% versus 26%. Skills and openings are measured against 2018, wages against 2021.
The harshest of it is at the entry to a profession. Entry-level openings most exposed to AI are seven times more likely to demand skills once expected of an experienced specialist. Those that “grew up in their requirements” rose by 35%; the rest fell by 10%.
Entry to a profession: the market split into two halves with opposite signsPwC 2026
Change in the number of entry-level openings, 2019 → 2025.
×7
more often the entry-level openings most exposed to AI demand the skills of an experienced specialist
−16%
drop in entry-level employment in fields exposed to AI (Stanford)
−20%
employment of US developers aged 22–25 by 2024
49%
CEOs expect hiring of junior specialists to fall
The comparison with last year is more telling here than any single figure. PwC 2025 wrote honestly: growth everywhere, even in the most automatable occupations. PwC 2026 does not abandon that — the aggregate is still rising. It is just that the aggregate stopped being informative. This is the classic story of average temperature: the mean barely moves while the spread inside it grows.
In substance the two tracks really do diverge. In occupations where AI lowers the entry bar, requirements for prior expertise fall further and the premium for it is lower; where expertise requirements rise, wages and the number of openings grow faster. That does not make one track “good” and the other “bad” — they describe different ways of rebuilding work.
One possible mechanism: AI replaces part of the simple tasks on which beginners used to gain experience. Job-posting data on its own does not prove that causality. But if the pressure on early roles persists, companies will have to design the path of gaining experience separately — otherwise the funnel that produces mid-level specialists may narrow.
05
Individual gain is growing faster than the organisation's ability to convert it
This is BCG's most durable finding across two years. In 2025: 47% of employees save more than an hour a day, but only a third receive guidance on where to put that time. In 2026: 52% save a working day a week or more, and the guidance gap has not closed.
The guidance gap: the lower the level, the less guidanceBCG 2026
Share of employees receiving limited or no guidance on where to put the time AI saves, and the share who do not convert that time into strategic work.
Receive no guidanceDo not convert time into strategic work
Between frontline employees and leaders there are 14 points on guidance and 22 on converting time. The “overall” row is a separate figure across the whole sample.
Training is the same story. 88% of employees expect they will need major reskilling within five years; only 36% believe they have been trained adequately. Neither figure changed from 2025.
Two figures did not move a single point over the year: 88% expect reskilling, 36% consider themselves trained. In a year when almost everything else moved by tens of points, these two did not move at all. The 52-point gap reproduced twice on different samples of 10,635 and 11,749 people. This is not noise.
The guidance gap runs not down the whole ladder but across one rung: frontline employees and managers are equally left without guidance — 66% and 66% — while leaders are markedly better off at 52%. On converting time into strategic work the ladder is already monotonic: 58 / 43 / 36. A leader answering “do you give guidance” answers honestly “yes” — because they give it to themselves and their own circle.
The two years' surveys do not use fully identical measures, so it cannot be said strictly that “conversion has not risen”. But the tension persists: employees report large time savings, while a substantial share simultaneously receives no clear rules for reinvesting it and does not convert it into strategic work. Individual productivity may well be rising — it simply does not guarantee extra margin, revenue or strategic capacity. This is one of the main unaccounted gaps of 2026.
06
Agents came out of pilots into operation, but operating models did not keep up
Integration of agents into end-to-end workflows rose from 13% to 30% — more than double in a year. Microsoft records growth in the number of active agents in the M365 ecosystem 15× year over year, 18× in large enterprises.
Six waves KPMG, fielded in the US, companies with revenue ≥ $1bn. The only source with within-year dynamics.
DeploymentPilotExperiment
Across six waves pilot and deployment swapped places: 65% versus 11% in early 2025 and 35% versus 53% in mid-2026. In the Q4 2025 wave the question was asked differently, so the lines do not run through it.
But half of employees say their company has no clear rules for managing mixed “human × AI” teams, and the average organisation recorded 54 agent incidents over the year, 17% of them high severity. And the check from independent data: agent deployment stays in the single-digit percentages across almost every business function.
Here three sources give three non-matching readings of one phenomenon — and that, to me, is the most interesting place in the whole picture. KPMG: deployment above 50%. BCG: integration into end-to-end workflows — 30%. Stanford: single-digit percentages by function. A five- to sixfold spread.
There is no contradiction here — there are three different questions. “Have you deployed agents” (at least one, somewhere) is KPMG. “Are agents built into an end-to-end workflow” is BCG. “What share of functions actually runs on agents” is Stanford. The stricter the question, the smaller the figure. And that is exactly the scale on which to measure your own organisation: not “do we have agents”, but “what share of functions”.
Separately on telemetry Microsoft: ×15 year over year is the only telemetric reading here of organisational agent use specifically; Stanford's technical benchmarks belong to a different category of data. It does not ask an opinion, it counts objects. And it confirms the direction but says nothing about value: an agent created and launched twice counts twice too.
07
IBM changed its research focus: from individual functions to the business and technology perimeters
The six findings above rest on measured indicators. The seventh is an observation about how IBM's own research focus changed, so it needs to be read more carefully.
In 2025 IBM published four separate role-based studies: CEO, CMO, CDO, CAIO. In 2026 — two: CEO (the business perimeter) and Tech Leader for CIOs/CTOs (the technology perimeter). Meanwhile the CAIO role went from exotic to standard in a year.
Share of organisations that have introduced a CAIO role or its functional equivalent.
76%
of organisations in 2026 it was 26% a year earlier, 11% in 2023
A role that barely existed three years ago became standard faster than any other part of the org chart in this data.
The structure of the publications is interesting here as context, but not as a market metric in its own right. In 2025 IBM published separate CEO, CMO, CDO and CAIO studies; in 2026 the main cut became “CEO / Tech Leader”. IBM does not explain whether this means the functions are converging, an editorial change, or simply a different design of the research programme.
One possible reading is that the focus really did shift to the seam between the business and technology perimeters. But the structure of publications does not prove that on its own; such a conclusion needs to be backed by substantive indicators — the distribution of accountability, CAIO penetration and the blurring of boundaries between business and technology.
The 26% → 76% rise gives a more substantive signal: a sharp increase in the share of organisations with a CAIO or functional equivalent is consistent with the institutionalisation of AI and the appearance of a formal owner for the agenda. But here too these are two market snapshots, not a panel of the same companies.
2. The outside view
From the outside the year looks like this: private AI investment added 127.5%, the lead of the best American model over the Chinese one narrowed to 2.7%, and the public and the experts diverged by 50 points — growth, parity and distrust in one frame.
This is not a dimension but the context in which everything else happens. The basis is Stanford HAI, the primary source here for comparable macro and technology context.
2.1 Money
Private AI investment: the US versus ChinaAI Index 2026
Annual volume in billions of dollars. The US is ahead of China by a factor of twenty-three.
Private investment is not China's whole bill: by the source's own estimate, state funds directed roughly $184bn into AI companies over two decades.
+127,5%
growth in private AI investment over the year; it accounts for 60% of the entire global corporate volume
+200%
growth in generative AI — almost half of all private funding
+71%
newly funded AI companies; in the US 1,953 — ten times the nearest country
>$150bn
in annual capital expenditure by Google alone; corporate AI investment worldwide more than doubled
Two figures are worth placing side by side. Annual capital expenditure by Google alone — more than $150bn. All of China's private AI investment in the same year — $12.4bn. The order of magnitude differs by more than tenfold, while the gap in model quality between the countries stood at 2.7% by March 2026 and has effectively closed.
The scale of capital does not on its own explain the gap in model quality. These figures cannot be divided by each other: they have different denominators. But their proximity shows that the link between the volume of capital and the quality of frontier models is non-linear. For a corporate budget the conclusion is more cautious: more spend does not on its own guarantee a proportional rise in quality.
And one more thing: growth in private investment of 127.5% over the year shows that capital keeps flowing rapidly into AI and that the market remains in a phase of intensive expansion.
2.2 Adoption and use
Adoption hit the ceiling — for organisations, but not for the workAI Index 2026
Two different things under one word, “adoption”: how many organisations use it and how many people.
The upper panel hit the ceiling: 88% is practically saturation, an indicator that no longer tells anyone apart from anyone. Meanwhile agent deployment, around which the whole rhetoric of 2026 is built, measures in single-digit percentages for an independent observer. The gap between “we use AI” and “AI does the work” is the central fact of this year.
Among people the same technology spread faster than the personal computer or the internet — but spread so unevenly that the producing country ended up in the bottom half of the second ten.
The US leads on investment and models — and ranks 24th on penetrationAI Index 2026
Share of the population using generative AI. Penetration correlates with GDP per capita, but the US falls off the relationship.
The US is 24th in the world on consumer penetration, the bottom half of the global ranking, while first on investment and models.
This is one and the same reality measured at two different altitudes. “We use AI” in 2026 means roughly what “we use email” meant in 1999: a fact of presence, not a fact of work.
On the US at 24th: a country that produced more than 90% of the frontier models of 2025 and invested $285.9bn trails the UAE on consumer penetration by more than a factor of two. The most reasonable explanation, I think, is that the production of the technology and its consumption have drifted apart geographically. For a corporate reader something practical follows: the intuition that “our people are surely already using all this” can be badly wrong depending on the market.
But where people did pick AI up, they get something quite measurable from it — just not into their employer's till.
2.3 Value for the consumer
Consumer surplus from generative AI in the USAI Index 2026
Estimated value received by users above what they pay. Annual figure in billions of dollars.
$172bn
of value a year that users receive above what they pay
+54%
growth in total surplus over the year
×3
the factor by which median value per user grew
Most of these tools remain free or nearly free.
This raises an important question for corporate return calculations. $172bn is an estimate of value to users, not of employers' financial result. It illustrates a more general principle: value created by AI does not have to show up in an organisation's revenue or profit.
This figure does not directly explain why only 8% of CEOs saw a noticeable revenue gain from AI over the year, but it points to a possible gap: a user may get convenience, time saved or a better result, while the company sees no comparable revenue gain. So the effect on a person and the financial effect on an organisation have to be measured separately.
What remains is to look at the technology itself — how capable it actually is of doing what it is paid for.
2.4 The edge of what the technology can do
>90%
of the notable frontier models of 2025 were produced by industry, not academia
2,7%
the lead of the top US model over China by March 2026 — the gap has effectively closed
50,1%
accuracy of the best models at reading an analogue clock
~66%
agent success on OSWorld — it was 12%; every third attempt still fails
The jagged frontier is the source's own term. A model takes gold at the International Mathematical Olympiad — and reads an analogue clock correctly half the time. The most capable systems simultaneously became the least transparent: training code, parameter counts and dataset sizes are no longer disclosed.
That cannot be checked on trust — but what happened to the capabilities themselves over the year is counted across four benchmarks, and there is no opacity here.
The only series that already has 2026 data. On the left, what was known a year ago; on the right, what is known now.
Agents grew most: from 12% to 66% success on real computer tasks — and every third attempt still fails. The values AI Index 2026 are as of early 2026; FrontierMath is given at the top difficulty tier.
The benchmarks all rise together — but the return on that rise in real work is distributed nothing like evenly.
Productivity gains are extremely uneven by type of workAI Index 2026
The effect concentrates in structured, easily measurable work — exactly where it is easiest to count.
In tasks that require judgement the gain is markedly weaker or negative — the source gives no figures. Customer support is given as a range, 14–15%.
The judgement layer is precisely what most expensive employees are paid for. Hence the smeared average corporate ROI: the measurable gain lands where the work was already transparent and countable, and where judgement is needed it weakens or goes negative.
I would treat the jaggedness as the main planning risk. A model that takes gold at a mathematics olympiad intuitively seems capable of anything “simpler than mathematics”. The clock shows that this is wrong: capabilities have no single scale of difficulty. Planning adoption “from a general impression of the model” is the most expensive way to be wrong in 2026. Each specific task has to be checked.
Capabilities have one more limit, and it is not in software: electricity, one Taiwanese contractor, and people who do not believe any of it.
2.5 Physical and social constraints
29.6 GW
of AI data-centre capacity — comparable to the peak consumption of New York State
5 427
data centres in the US — ten times more than any other country
362
documented AI incidents over the year
50 pp
the gap between experts and the public in assessing AI's impact
Physical constraints are at least measurable; the last one on the list is not, and it is the most stubborn.
The gap between experts and the public is 50 percentage pointsAI Index 2026
Share expecting a positive effect of AI on how people do their jobs.
One and the same question put to two audiences: US adults and US AI experts. The distance between the answers is 50 percentage points.
The 50-point gap between experts and the public is a fact; the cause of that gap is not established in the data. It may be related to experience of use, exposure to the technology, expectations of the labour market or risk perception, but the AI Index does not let you pick between those versions.
For a corporate reader the correct conclusion is already available: a single message that “AI is useful” is not enough. Different groups may assess the consequences of the same technology very differently. Inside a company that gap has to be measured separately between levels and roles, rather than explained by psychology without data.
Part II
Eight dimensions
The outside view showed the main thing: the scale of capital does not on its own explain the gap in model quality, and capabilities grow unevenly. Next comes what happens inside organisations, broken out along eight axes.
Every dimension is built the same way: how it looked in 2025 → where it stands in 2026 → where the tension is → what to measure. A dimension is an axis you can put numbers on, not a slogan.
Dimension 1
The economics of AI
You have to start with money: until it is clear what an organisation gets back, the other seven dimensions are a conversation about means without an end.
What we measure. How much goes in, what comes back, whether it can be measured at all, and how return is distributed between the leaders and everyone else.
3.1 How it looked in 2025: belief versus reality
The year began with near-unconditional belief and ended with the first honest numbers. The gap between the two columns below is the content of 2025.
Expectation
8 out of 10 CEOs demanded AI scale both savings and growth within 18 months
85% expected positive ROI on scaled savings initiatives by 2027, 77% — on growth initiatives
Through the whole first half of the year KPMG recorded unanimity: productivity as a return metric — 98%, profitability — 97%, quality of work — 94%
Reality
Over three years only 25% of initiatives delivered the expected ROI, only 16% were scaled to enterprise level
Average AI ROI per CAIO data — 14%
60% CEO still in pilot — although a year earlier more than two thirds expected to be out of it by 2025
The share expecting to drive growth with AI fell from 67% to a little over 50%
Only 52% CEO get value beyond cost reduction
In marketing the picture was harsher still: 19% of pilots with the expected ROI, 25% scaled to several business units, 14% enterprise-wide. 58% of CMOs admitted they invest in technology before they understand its value, out of fear of falling behind.
Two answers from the same people in the same survey.
72%
CAIOs say the organisation risks falling behind without measuring AI's impact
68%
of the same CAIOs launch projects whose impact they cannot assess
72% and 68% are the same people answering two adjacent questions. They know that without measurement they will fall behind. And they launch the unmeasurable anyway. That is neither stupidity nor hypocrisy — it is a description of a market where the price of inaction exceeds the price of error.
In 2025 that was how the market had to be read: investing before understanding value was an honest description of strategy, not a mistake. In 2026 the same behaviour began to cost measurably more — because over 2024–2025 the top of the distribution rose from 200% to 250%, while the bottom rose only from −35% to −20%.
Put the two columns side by side and the gap turns out to be not between “optimists” and “pessimists”, but between the horizon of a promise and the horizon of a report. 85% expect ROI “by 2027” — that is about the future, and therefore safe. 25% achieved ROI “over three years” — that is about the past, and therefore checkable. These two figures come from different slices of the management audience, so they are worth comparing as expectation versus the market's actual result, not as a change of mind in the same people.
3.2 Where it stands in 2026: money jumped sharply
That admission had no effect on budgets whatsoever: in the first quarter of 2026 the spending forecast jumped at once from $124m to $207m.
Average spending forecast at US corporations with revenue ≥ $1bn, $m. Six KPMG waves.
The low of the series is $88m in the second quarter of 2025, back to $130m in the third, a jump to $207m in the first quarter of 2026 and holding at $202m in the second.
AI spend as a share of the IT budget: +71% in two years
Hollow fill is forecast, solid is actual.
For comparison: a year earlier IBM forecast growth in the share of about +31% a year — expectations accelerated.
3.3 Measurement got harder, not easier
The money was found quickly — but the ability to explain what it was spent on sagged over the same year.
Barriers to demonstrating ROI: measuring got harder, not easierKPMG Q1’26
Q1 2025 → Q1 2026. Three of the four barriers almost doubled.
The only item going down is privacy and cybersecurity risk. Everything else used to explain AI's return to the board became harder to explain over the year.
78% of leaders agreed as early as Q3 2025 that traditional business metrics are becoming insufficient for measuring AI's impact. By 2026 this produced new metrics: “improved analytics used by senior management in decision-making” rose as an ROI metric from 62% to 83% in a quarter and held for a second quarter running.
3.4 A new object of measurement — the cost of AI itself
Q2 2026 raises the question of operating economics for the first time. The answer is not encouraging.
Visibility of AI operating costs: control exists, visibility does notKPMG Q2’26
Two panels from one survey: what has been put in place to manage costs and what is actually visible. Shares of organisations, Q2 2026.
84%
of technology leaders have not operationalised financial management of AI
85%
do not have full real-time visibility of spend
It is not only invisible how much goes out — it is also invisible for how long: the asset the money goes into lives far less than is usually assumed.
The assets are different now: an AI model's life is about 14 monthsIBM Tech 2026
This casts doubt on business cases that assume the model layer stays unchanged over a three- to five-year horizon.
14
months — average useful life of a model
Main reasons for replacing a model
First: the budget almost doubled over the year — while three quarters of organisations cannot see the cost of the thing it goes into. Between “we have a tool” and “we have a picture” there are forty points, and that, to my mind, is the most underrated figure of the year: the tool was bought, the process was not built.
Second: 14 months is the average model life cycle in IBM's sample. That is not the same as the accounting useful life of the whole AI system, which includes infrastructure, data, integrations, software and training. But a business case that assumes one and the same model layer for three to five years really does become fragile. The main reason for replacing a model is the appearance of a better one (71%). Here an asset more often loses relevance through external progress than through physical wear.
No single accounting treatment follows from this: that depends on the specific asset and the reporting rules. What does follow is a management principle that IBM itself states: run AI as a continuously reviewed portfolio — refresh models, close weak bets and reallocate capital. Indirect benefits are then better counted at the level of the process or the portfolio, so that swapping a particular model does not zero out the whole business case.
3.5 The forecast that did not come true
If the return horizon has compressed to eighteen months, it is worth checking how well these forecasts hit their own deadlines at all.
IBM published a table of its own past forecasts, and it deserves separate attention.
A systematic two-year error on timing alongside rising confidence in the outcomeIBM CEO 2026
Share of CEOs expecting AI to be primarily driving growth.
In 2024 49% CEO expected that by 2026 AI would be primarily driving growth; the 2026 actual is 10%. Confidence in the same outcome by 2030 did not fall over that time but rose — from 55% to 72%.
Here the source published its own miss — a rarity in corporate research. In 2024, 49% of CEOs expected AI to be primarily driving growth by 2026; the actual share in 2026 is 10%. Meanwhile the expectation of the same outcome by 2030 rose from 55% to 72%. This shows the durability of expectations even after a miss on timing, but one such series is not enough to derive a universal law of technology markets.
The second line is more interesting though. Faced with the miss, the market did not lower expectations — it raisedthem: the 2030 forecast rose from 55% to 72%. That is a recognisable structure: the deadline slides, confidence grows. Expectations behave this way in every technology market, and this is exactly how they behave right before expectations correct.
A simpler rule follows: distant CEO forecasts here should be read as expectations, not calendar promises. One 2024 → 2026 miss does not license shifting every forecast two years automatically; it does justify revisiting the horizon regularly as the facts come in.
3.6 Where the tension is
Two statements cannot be reconciled directly: KPMG records that 97–98% of executives name productivity and profitability as the top return metrics for AI, while PwC, on CEO Survey data, records that only 8% of CEOs report a more than negligible revenue gain from AI. That is not a divergence of data but a divergence of questions. Any AI ROI taken from an executive survey should be read as a sentiment indicator, not a financial fact.
If ROI from surveys cannot be trusted on its own, what is left is to measure what can be checked without a survey — the structure and its cost.
expected revenue growth for those who built all three pillars of structural readiness at once; the same group shows +7% operating margin and more than twice as many agents
How much higher ROI is for one group than another, as a percentage of ROI itself. Early architectural replaceability is a 2025 reading.
Benchmark
01
Share of initiatives that achieved the expected ROI25% (2025)
02
Share scaled enterprise-wide16% (2025)
03
AI spend as a share of the IT budget14,5% → 19,0% → 24,9%
04
ROI spread between the top and bottom of the range+250% / −20% (2025)
05
Average model life~14 months
Dimension 2
Scaling
The economics produced a spread: multiples for some, losses for others on the same technology. The first practical difference between those two groups is visible straight away — some reached industrial operation, others stay in pilot for a sixth wave running.
What we measure. Where organisations actually stand on the path from experiment to industrial operation, and how far claimed maturity matches observed maturity.
4.1 How it looked in 2025: the barrier moved from people to architecture
The year began with a mass exit from experimentation and ended in an intermediate state. Microsoft: 24% of leaders reported organisation-wide AI deployment, 12% remained in pilot; 82% called the year a turning point; 81% expected agent integration within 12–18 months. BCG from the employee side: only 13% see agents integrated into end-to-end workflows, while 77% consider agents important over 3–5 years and only 33% understand what they are.
Barriers to agent deployment by wave. The set of barriers changes from wave to wave: in Q2 2025 the barrier was called human, in Q3 technical, in Q2 2026 architectural.
Staff resistance — halvesData readiness — appears straight in first placeOther barriers
Staff resistance fell from 47% to 21% in a quarter and in 2026 is no longer published in the barrier list at all. Its place was taken by data readiness — 58% — and human-oversight skills — 43%. An empty position means the barrier was not published in that wave.
By Q4 2025 complexity held as the top barrier for a second quarter running, while structural problems came to the fore: inconsistent use across business units rose from 19% to 45%, absence of organisational infrastructure — 41% (threefold in two quarters), unclear corporate strategy — from 20% to 32%.
4.2 Where it stands in 2026: overall deployment stalled, chaining agents together doubled
The structure of deployment: Q1 2026 → Q2 2026KPMG Q2’26
The total number that deployed barely changed (55% → 53%), but a qualitative shift happened inside it. In its own wave KPMG Q1 2026 scaling across several functions is 40%; in the series it is the retrospective figure for the same indicator from the Q2 2026 deck.
13→30%
integration of agents into end-to-end workflows, BCG, over the year
×2,3
×15
growth in the number of active agents in the M365 ecosystem year over year
×18 in large enterprises
1 661
agents per enterprise expected by 2027
+38% on today
61%
believe that in 3 years agents will be able to do half their job
frontline employees 52% · managers and leaders 65%
Awareness rose, and so did understanding, with one caveat: have heard of agents — 72% → 84% (+12 pp), understand only in a limited way what they are — 52% (−15 pp against 2025). Among frontline employees, understanding is limited for 61%.
While people work out what an agent is, some decisions are already being taken without them — and that share has been measured.
Decision autonomy: 25% today → 48% by 2030IBM CEO 2026
Share of operational decisions taken by AI without human intervention. This covers pricing, inventory allocation, rerouting shipments, auto-remediation of incidents.
65%
of organisations plan or already run AI autonomy in demand forecasting
61%
of organisations — in inventory optimisation
Stagnation in overall deployment alongside a doubling of chaining agents into a system is the most substantive pair of figures in this dimension. The usual reading is “growth has stopped”. To my mind it is the opposite: growth changed direction from extensive to intensive. Organisations stopped adding new agents and started connecting the ones they have.
The barriers changed their nature entirely over the year: in Q2'26 first place goes to data readiness (58%), which was not on the list at all in 2025. People stopped being the problem in about two quarters. Data managed to become one.
And one figure to treat with care: 61% of employees believe that in three years agents will be able to do half their job. That is not a forecast — it is the self-assessment of people, 52% of whom “understand only in a limited way what an agent is”. The value of this figure is not its precision but the direction of expectations: people are already ready for what organisations are not ready for yet. Only 11% of technology leaders feel fully ready for the scale of deployment.
4.3 Where the tension is: claimed maturity versus observed maturity
Who is asking
01
KPMGWhom US executives · Result53–55% deployed agents
02
BCGWhom Employees at all levels · Result30% see agents in their workflows
03
Stanford HAIWhom Secondary data · Result Deployment in single-digit percentages across almost every function
04
IBMWhom CIO / CTO · Result Only 11% fully ready for scale within 12 months — while 80% get their transformation mandate directly from the CEO
A spread from “single-digit percentages” to 55% is not noise. These are different definitions: “the organisation has deployed at least one agent somewhere” ≠ “agents work in an end-to-end workflow” ≠ “a function is systematically executed by agents”. Your own maturity has to be measured by the strictest of the three.
Unevenness by function: a 47-fold spread between the extreme functionsKPMG Q3’25
Share of organisations that have deployed agents into the workflows of the relevant function, Q3 2025.
Marketing and HR are the functions talked about most in connection with AI; they are also the ones with the least deployed.
HR — 2%. The function responsible for reskilling, for change management, for mixed “human × AI” teams, barely uses AI itself. And it is HR that is expected to solve both the guidance gap and the training gap, which has not moved in two years.
The shape of the curve shows that AI is more often deployed where the result is easier to check (IT 95%, operations 89%), and less often where the result depends more on judgement about people (HR 2%, marketing 18%). This is consistent with the jaggedness Stanford describes at the model level: a 26% gain in software development and a weak or negative effect in tasks requiring judgement. The coincidence is interesting but does not on its own prove that model capabilities determined the order of adoption.
4.4 What to measure
Reduce all of the above to what is worth tracking at home and six lines remain — and each has a benchmark from this section.
Share of business functions where agents are integrated into an end-to-end workflow rather than launched as a standalone application
Ratio of agents “chained into a system” to the total deployed — there is no direct benchmark; the nearest reading uses a different denominator — at the stage of chaining agents together are 18% of organisations (Q2'26)
Number of agents per organisation and its forecast — benchmark 1 661 by 2027
Share of operational decisions without a human — benchmark 25% today → 48% by 2030
Readiness for scale by technology leadership's own assessment — benchmark 11% “fully ready”
The gap between the leaders' assessment and employees' assessment on the same indicator
Dimension 3
Data and architecture
Over the year the barriers to scaling changed their nature entirely: staff resistance halved in a quarter, and first place went to something that was not on the 2025 list at all — data readiness, 58%. So the next layer is the foundation all of this stands on.
What we measure. The ability of the technology foundation to keep rebuilding itself faster than the technology itself changes.
5.1 How it looked in 2025: the bottleneck was called “data”
A year ago the answer to “what is in the way” was the same everywhere, and it fitted into one word: data.
IBM devoted a separate study to the problem. From the foreword of the CDO Study: data is locked inside functions, each function has its own, there is no shared taxonomy, no end-to-end visibility — and every AI initiative turns into a data-cleaning project six to twelve months long.
An architectural shift appeared too: 81% of CDOs bring AI to the data rather than centralising data for AI. The share of the IT budget going to data strategy rose over two years from 4% to 13%.
Solid — by formal criteria. But ask not “is it done” but “will it hold” and the answers fall threefold. Five such questions were put to different audiences in different studies, and all five landed in the same corridor.
The confidence ceiling: five independent readings converged on a quarterIBM CDO 2025
Share of executives confident in the relevant element of the foundation. Five different questions, one and the same range.
For context: 92% CDOs are required to be benchmarked on outcomes, but only 29% fully agree that they have clear measures of the value of that outcome.
A coincidence like that is not chance. It means we have run not into a particular hole but into a general ceiling: roughly a quarter of organisations have actually reached a state where the foundation is ready for the next step.
And here is what is genuinely interesting. Integration of data strategy rose from 52% to 81%. Platforms — from 41% to 75%. Enormous progress. Confidence — a quarter. Which means requirements grew faster than the foundation. That is the content of the rewording of the bottleneck: not “the data is bad” but “we are running and the horizon is receding faster”.
5.2 Where it stands in 2026: the bottleneck is called “adaptability”
When the horizon recedes faster than you run, the question “are we ready” stops working in its old form — and in 2026 it was rewritten.
IBM redefined the very notion of readiness. Readiness used to mean stability — conformity with a known operating model optimised for continuity, cost and predictability. At machine speed that stopped working: when models, platforms and tools change faster than planning cycles, stability turns into a constraint. The new definition of readiness is replaceability: the ability to move workloads, change models and absorb new capabilities without a large-scale rebuild.
What you pay for on the way out of a cloud: barriers to moving workloadsIBM Tech 2026
What looks like inflexibility today is the consequence of yesterday's rational decisions.
88%
of organisations plan or are attempting to move workloads between providers
25%
of their workloads they call easily portable
88% of organisations are attempting or planning to move workloads — and technology leaders call only a quarter of those workloads easily portable. The three rows above are what that difference runs into: it is not laziness, it is what you pay for on the way out of a cloud.
+48%
by this much cloud spend exceeded the original forecasts
80%
of organisations: data transfer costs above expectations
+10%
return on AI investment for those who designed replaceability early
The original reasons for moving to cloud: cost optimisation 60%, enabling innovation 59%. Optimising for cost is what created the dependency.
IBM shows an important trade-off: decisions taken yesterday for cost and convenience can constrain portability today. 60% named cost optimisation as one of the reasons for moving to cloud, and now call only a quarter of workloads easily portable. These two facts are consistent with a story of accumulated vendor dependency, but do not on their own prove that cost optimisation caused it.
The redefinition of readiness changes the engineering priority: portability and the ability to swap components quickly are added to predictability. These are not strictly opposing goals, but between specialising for a particular platform and portability there is often a trade-off in cost and complexity. The higher reported ROI at organisations that invested in replaceability earlier is an association, not the “price” of one particular decision.
5.3 How data looks in 2026 from other angles
So far this is one source's diagnosis — but the other three run into the same point, each from its own side.
KPMG: named data quality as a barrier to AI strategy goals in 2025 — 82%. By Q2 2026 data readiness and availability became the main barrier to deploying agents — 58%, ahead of human-oversight skills (43%), the complexity of agentic systems (38%) and technical skills (37%). BCG: adequate guardrails are seen by 59% of employees at companies redesigning processes, against 41% at those simply deploying tools. Stanford: the physical ceiling became visible — 29.6 GW of AI data-centre capacity and the whole frontier-chip supply chain depending on one Taiwanese contractor.
5.4 Data as a source of differentiation
Everything said so far describes data as a burden. It has a flip side too, and it is stated far more quietly: it is the one thing a competitor cannot copy.
This is the most underrated line in the report, and in 2026 it strengthened. In 2025: 84% CDO say their unique data products have already delivered significant competitive advantage; 78% name the use of proprietary data as a top differentiation task; 72% CEO — that proprietary data is the key to unlocking the value of generative AI; 82% CDO consider data sovereignty a critical element of risk management.
In 2026: 83% of all CEOs — and 97% of CEOs at AI-benchmarked companies — say AI sovereignty is material to their business strategy; 63% CEO agree that competitive advantage in 2030 will come primarily from the sophistication of their AI models.
Model strategy swaps places: general-purpose → hybridIBM CEO 2026
Organisations' primary model strategy, 2026 → expectation for 2030.
Hybrid is not chosen out of caution: for those who have already found their mix, the lead over “one big model” shows up immediately on three measures.
+24%
greater productivity gain for those using the right model mix
+55%
greater improvement in operating margin by 2030
×2
greater reduction in process cycle time
Compared with those relying predominantly on large pre-trained models. A typical organisation was already using 11 generative models in 2025 and planned at least 16 by the end of 2026.
The shift of 39% → 13% on general-purpose pre-trained models and 13% → 50% on hybrid is in substance a forecast that “just take the best model” will stop being a strategy. And organisations expect that of themselves, not of the market.
Two more figures should be kept alongside this. A model's life is 14 months, the main reason for replacement is a better one appearing (71%). The number of models in operation rose from 11 to 16. With that combination an organisation swaps more than one model a month. No “pick the right model” strategy works at that tempo — only a “be able to swap fast” strategy does. That is exactly the same replaceability, only at the model level rather than the infrastructure level.
And separately on sovereignty: 83% of all CEOs against 97% of CEOs at AI-benchmarked companies. The 14-point gap between those who have already rebuilt and everyone else is a leading indicator. AI-benchmarked companies consistently lead the average on every measure by a year to eighteen months.
5.5 Where the tension is
Formally the foundation looks solid. But confidence that this foundation will hold the next step stayed at the level of a quarter. The progress is real, but it has not caught up with the growth in requirements. That is exactly what the rewording of the bottleneck from “data” to “adaptability” means.
5.6 What to measure
Share of workloads genuinely easy to move between providers — benchmark 25%
Deviation of cloud spend from plan — benchmark +48%
Share of the IT budget on data strategy — benchmark 4% (2023) → 13% (2025)
Presence of clear measures of the value of data-driven decisions — benchmark 29% fully agree
Share of models in the portfolio: general-purpose / small specialised / custom — benchmark for 2030 50% hybrid strategy
Number of models in operation — benchmark 11 → 16
Time from setting a task to having the needed data available
Dimension 4
Governance and trust
The foundation turned out not to be “bad data” but an inability to change: 88% of organisations plan to move workloads, while only a quarter of the workloads themselves are easily portable. But replaceability without control gives you not speed but 54 incidents a year. Hence governance.
What we measure. AI governance as the condition for scaling, rather than as compensating control.
6.1 How it looked in 2025: governance as a risk function
Trust was the main theme of the first quarter: leaders named trust in the accuracy and fairness of AI outputs the biggest societal challenge to 2030 (32%), followed by misuse of AI by bad actors (30%). Over the year the priorities then swapped: by Q2 misuse took first place — 38% against 37% for trust in accuracy; it had risen from 30%, and by Q3 it grew to 45%, while trust in accuracy fell to 28%.
Oversight tightened all through 2025 and stabilised on output validationKPMG Q2’26 · six waves
Agent risk-mitigation measures by quarter. A gap means the measure was not published in that wave.
Trusted providersA human on every actionNo access to sensitive dataOutput validation
Oversight was not tightened out of caution: by the end of 2025 almost everything blocking AI strategy came down to one family of threats.
Two different questions from one wave, Q4 2025: what blocks AI strategy goals and which threats are of most concern.
Half of leaders planned to allocate $10–50m to protecting agentic architectures, improving data traceability and tightening model governance. At the same time IBM recorded that only 22% of organisations had set clear rules for using AI in automated decision-making — that is, roughly eight in ten had none.
6.2 Where it stands in 2026: governance as the condition for scale
The change of frame is recorded literally. KPMG Q1 2026 puts it in the executive summary: “governance is the condition for scaling”. 91% of leaders name data security, privacy and risk as the main factor shaping AI strategy over the next six months; in Q2 2026 — 92%.
The scale of the problem is already visibleIBM Tech 2026
Technology leaders' self-assessment: the gap between the pace of adoption and the ability to control it.
Accountability was handed out, the tools for it were not. Two more readings the source gives only in words: more than two thirds of technology leaders say business units bypass IT when adopting AI, and two thirds say they are accountable for systems they do not fully control.
So far all of this is a feeling: “we are not keeping up”. In 2026 the feeling acquired a counter for the first time.
Incidents became measurable: 54 a year for the average organisationIBM Tech 2026
On the left, how long an incident takes to contain; on the right, how it ends for the business. High severity by the source's definition means more than four hours to contain.
Time to contain
Main consequences
Since incidents happen anyway, the whole question is what puts them out: a procedure for each case or an architecture for all of them at once.
The key figure of the dimension: designed-in control versus manual governanceIBM Tech 2026
Control built into the architecture versus control implemented through procedures.
×16
more agents deployed
×4
smaller share of the AI budget spent on it
+18%
operating margin
“Control shifted from approving inputs to continuous oversight of outputs and outcomes — from gates to guardrails”.Dena Almansoori, ADNOC — quoted from IBM Tech Leader 2026
What designed-in control is: platform teams own the shared guardrails — telemetry, model registry, identity, logging, rollback; risk and compliance set policy thresholds; architecture sets reference patterns; business domains own outcomes inside the guardrails; incident response maintains predefined shutdown and recovery procedures.
The shape of the oversight curve changes with scale: “a human on every action” first rises and then disappears; “output validation” appears later and stays. That is consistent with a shift towards more scalable forms of control: checking every action grows with the number of actions, whereas checking outcomes can be aggregated. At 54 incidents a year and 17% high severity that works out to roughly nine high-severity incidents a year; by IBM's definition each of them requires more than four hours to contain. The two frequent consequences — data exposure (37%) and cascading failures (33%) — make response speed material. The gap between mandate and readiness is itself large; one hypothesis is that governance architecture delivers a slower and less visible result than launching a new tool, but the data does not test that cause directly.
At 54 incidents a year and a high-severity share of 17% that works out to roughly nine high-severity incidents a year. By IBM's definition each such incident requires more than four hours to contain; that does not mean each lasts exactly half a working day.
The most awkward thing here is the size of the gap between the transformation mandate and the sense of readiness. One possible reason is that governance architecture delivers a slower and less visible result than launching a new tool; the data itself does not test that cause.
Readiness for scale
11%
of technology leaders feel fully ready for the expected scale of agents — while 80% get their transformation mandate directly from the CEO
6.3 Governance through the employee's eyes — nothing is ready there
Everything above is the view from the office where ×16 agents and +18% margin have already been counted. One floor down, where mixed teams work every day, none of it is visible yet.
No clear guidance on managing mixed “human × AI” teamsBCG 2026
Share of employees who see no rules, by level of hierarchy.
47% put accountability for AI decisions in the top 3 of their concerns for the next 2–3 years — and that is the same at every level: frontline employees 46%, managers 47%, leaders 46%. The only question on which the hierarchy produces no divergence.
On everything else the gap between levels is 10–25 points. Here the gap persists: from 55% among managers to 41% among leaders. What converges on a single point is a different reading — concern about accountability for AI decisions, 46–47% at every level.
This strikes me as an important signal. Usually agreement across levels means the question affects everyone equally — and accountability for an AI decision really does affect everyone equally, because nobody knows the answer. A leader is afraid of being accountable for what they do not control (two thirds say exactly that); a doer is afraid of being accountable for what they did not choose. This is not a perception gap but shared uncertainty about the legal and organisational frame. And it is resolved not by communication but by actual rules, which half of companies do not have.
Outside company walls the count runs the same way, only it is kept by researchers rather than security teams.
6.4 The macro perimeter
Stanford: documented AI incidents over the year — 362. Practically all leading frontier-model developers report on capability tests, but reporting on responsible-AI tests remains selective. And a separate awkward finding: improving one dimension of responsible AI, say safety, can worsen another, say accuracy.
6.5 Where the tension is
Recognised as the condition for scaling — 91–92% of leaders put governance at the head of strategy
Economically justified — ×16 agents at a quarter of the cost and +18% margin
And not built —
6.6 What to measure
Share of agents and models that are registered, observed, with an assigned owner and the ability to stop them — the minimum production standard per IBM
Number of agent incidents and the distribution of time to contain — benchmark 54 incidents/year, 17% longer than 4 hours
Ratio of growth in incidents to growth in the number of agents — the key indicator of governance maturity
Share of controls implemented in code rather than in policy documents
Oversight model: human in the loop on every action / output validation / autonomy inside guardrails — benchmark 52% on output validation
Share of employees who see clear rules for mixed teams — benchmark 50% do not see them
Dimension 5
Operating model
Governance is recognised by almost everyone and built by almost nobody: 92% put it at the head of strategy, 11% are ready for scale. A gap that size is no longer explained by technology — it is explained by how the work itself is arranged.
What we measure. The main watershed: deploy a tool versus redesign the work. Here all six sources agree — and here the evidence base is largest.
7.1 How it looked in 2025
The whole argument about what “adopting AI” means came down in 2025 to one question: are you changing the tool or the work itself — and a ruler for that question appeared for the first time.
BCG introduced a scale that became the de facto standard for everyone: Deploy — rolling out tools and raising productivity; Reshape — end-to-end redesign of processes and functions; Invent — new business models and products. Even then BCG offered its own observation: the companies creating the most value with AI direct 80% of investment into reshape and invent — in a few key processes, rather than spreading it across all of them.
A recommendation is a recommendation, but the mass of organisations moved up that scale over the year — fastest of all where yesterday there was almost nothing.
Share of organisations working at each level. These are not parts of a whole — an organisation can do all three at once, so they will not sum.
An organisation can work at all three levels at once, so they add up to more than a hundred per cent here.
7.2 What redesign gives: the same actions deliver both value and satisfaction
The scale says who is at which level — and here is what it means for the person inside: BCG put side by side the answers of those whose work was redesigned and those who were simply handed a tool.
Reshape and invent versus deploy: the resultBCG 2026
Answers from employees at companies redesigning work versus employees at companies only rolling out tools. 2026.
The same questions put to two groups. The gap holds across all six rows: from 24 points on measurable improvement in business metrics to 11 on confidence in working with AI.
The same two groups, but by practice rather than by result. The biggest gap is employees' participation in process redesign.
12% versus 43%. Where tools were simply handed out, one in eight takes part in process redesign; where processes are being redesigned, almost half do.
A caveat that must not be lost
Employees at companies redesigning processes fear losing their job more — 46% versus 34% at companies that stopped at rolling out tools (against an average of 41%). In groups with deeper redesign both reported benefit and anxiety are higher at once.
7.3 Strategic clarity is more strongly associated with impact than access to tools
One of the six practices pulled away from the rest far enough to be worth testing separately — and BCG tested it.
BCG split respondents in its global survey on two criteria — how clear their company's AI strategy is to them and how available the tools are — and compared the four resulting groups by the share reporting measurable impact.
In this cut the groups differ more by strategic clarity than by access to toolsBCG 2026
Share reporting measurable impact from AI in each of the four groups.
At equal access to tools, high clarity separates the groups by 23–25 points; at equal clarity, broad access separates them by 3–5.
If one figure had to be kept from the whole report, I would keep this one. Not because it proves causality, but because the difference between the groups on strategic clarity is markedly larger than the difference on access to tools.
Clarity versus tools
+25/+5
difference from the base group: +25 pp at high clarity and limited access; +5 pp at broad access and low clarity
And employees measure it themselves: 52% call the strategy clear where processes are being redesigned, and 31% where tools were merely handed out.
Separately on the caveat of 46% versus 34% fear of losing a job. In groups with deeper redesign there is both more reported benefit and more anxiety. That does not prove causality, but it may reflect that the changes to their work are deeper and more visible. The practical conclusion is simple: the more the work changes, the more important clear communication about roles, expectations and consequences becomes.
7.4 IBM: the same finding from the CEO side
BCG asked employees — IBM asked the people who sign those decisions, and got the same picture in different units.
×2
more often meet their business goals — CEOs who actively redesign cross-functional work
×4
more often — those who redesigned five key areas as a single system
87%
of CEOs actively build AI into end-to-end workflows
60–80%
recommended share of productivity gains to be reinvested
IBM's wording: when functions evolve independently, the effect adds up linearly; when they are redesigned as a single system, the effects reinforce each other. The five areas are technology, finance, HR, operations and cross-functional collaboration. And a hard statement about priority: redesigning processes comes before redesigning roles. “Do not fund reskilling until there is a redesigned way of working in which those skills can be applied”.
One of the most methodologically transparent findings in the report: the model, the factors and the “association, not causation” limit are stated explicitly. Microsoft tested 29 factors (10 organisational, 9 individual, 10 demographic) against self-reported impact from AI on a sample of ~20,000. The authors' wording: “The real question isn't whether people have the right skills. It's whether the organization is built to unlock them”.
Where AI users sit on two measures: personal ability to work with AI and the organisation's readiness to support it.
31% of AI users are in a mismatch between themselves and their organisation — that is, outside the “leaders' zone” and “forming” categories.
The mechanics of the paradox: the same forces that accelerate adoption also hold it backMicrosoft WTI 2026
Three answers from the same respondents.
Metrics, incentives and norms keep rewarding the old way of working.
Three figures — 65 / 45 / 13 — show a mismatch between new possibilities and old rules. A set of metrics and incentives like that makes rethinking work less attractive for an employee.
“Hands tied” at 10% matters especially: people rate their own skills high and their organisation's support low. This is a potentially expensive gap between capability already created and the conditions for applying it, but the survey contains no data on cost, so this category cannot be called “the most expensive”. For comparison, “untapped potential” is 5%; this shows the mismatch more often comes from the organisation's side than from personal readiness.
The 13% figure points to incentives as one possible organisational barrier. They are worth checking alongside metrics, manager support and the rules of work. This data does not prove that changing rewards should be the first step or that it is sufficient on its own.
7.6 The organisation as a learning system — a new theme in 2026
In 2025 this line barely existed. In 2026 it appears at two sources at once. Microsoft frames it as “every company is a learning system”: the firms pulling ahead focus on mastering AI, not merely adopting it — turning the result of work into understanding, and understanding into a changed way of working.
What that looks like day to day can be seen in the 16% of the sample that Microsoft calls the leading group: what differs is not their tools but their team habits.
What employees in the leading group (16% of the sample) do differentlyMicrosoft WTI 2026
Practices of team work with AI.
The gap holds across every row, but the lowest pair is documenting workflows at the organisation level: 25% even in the leading group against 14% for the rest.
Even among employees in the leading group, workflows are documented at the organisation level in only 25% of cases. That is a strong signal of weak institutionalisation of the practices found.
Which means that almost nobody is currently accumulating their own intelligence. Knowledge lives in people's heads and in private chats, not in institutional memory. And that, to my mind, is the most underrated risk of the year: the local wins of 2026 will disappear along with the people who found them, unless they are codified. In a technology where a model lives 14 months, the only genuinely accumulating asset is precisely how you apply it.
7.7 Where the tension is
This is one of the most consistent dimensions in the report: different sources point to the importance of organisational redesign. Causality is not established, so it is more useful to measure the gap between declared redesign and its actual depth. An indicator of that gap is right there in the data: 57% of organisations declare reshape, but only 43% of their employees take part in process redesign, and only 30% see agents built into their workflows.
7.8 What to measure
Distribution of investment across the “deploy”, “reshape”, “invent” levels — 2026 benchmark 78% / 57% / 42%; recommendation from BCG: 80% of investment into reshape and invent, in a few key processes
Number of end-to-end business areas redesigned as a single system — benchmark 5 areas → ×4 likelihood of meeting goals
Share of employees who consider the AI strategy clear — benchmark 52% at the leaders versus 31%
Share of employees taking part in process redesign — benchmark 43% versus 12%, the largest gap
A fixed share of productivity gains to be reinvested — IBM's recommendation 60–80%
Share of employees rewarded for rethinking work regardless of the result — benchmark 13%
Share of teams whose agent workflows and quality standards are documented and reproducible
Dimension 6
Leadership
Strategic clarity turned out to be five times stronger than access to tools: +25 points against +5. Clarity cannot be bought — someone has to formulate it and get it across. Who exactly, and whether it reaches the bottom, is the next question.
What we measure. Who is accountable for AI, how the top level of management is being rebuilt, and whether direction reaches the doer.
8.1 How it looked in 2025
The CAIO role was only emerging: of 2,300 organisations surveyed only 26% had a CAIO (in 2023 it was 11%). 57% of CAIOs were appointed from the internal bench, 66% expected most organisations to have a CAIO within two years. In IBM's sample, having the role was associated with a higher reported result: +10% on ROI on AI spend and +24% likelihood of outperforming competitors on innovation.
But a line in the org chart settled nothing on its own: at the same title the result diverged depending on how the AI function is arranged beneath that leader.
The differences by operating model were larger than the fact of having the roleIBM CAIO 2025
Distribution of organisations by maturity stage depending on the AI operating model.
CAIOs running hub-and-spoke or centralised models move twice as many pilots into production and get 36% higher ROI. Note the bottom row: at decentralised ones the direction is reversed — more pilots, less scaled.
61%
of CAIOs controlled the organisation's AI budget
57%
reported directly to the CEO or the board
76%
other CxOs consult them on important AI decisions
32%
of CAIOs named the CHRO as one of the main opponents of AI
One floor up sits the board, and by the end of that same 2025 AI reached it too.
What the board was dealing with by the end of 2025KPMG Q4’25
Topics boards work through in connection with AI, and the share whose board has “substantial” AI expertise.
The board discusses AI constantly, but only four boards in tencall their own expertise “substantial”. The gap between how often it is discussed and how competent the discussants are is a characteristic mark of 2025.
8.2 Where it stands in 2026: functional boundaries declared obsolete
The CAIO role became standard: 26% of organisations in 2025 → 76% in 2026, and 100% of CEOs expect the CAIO's influence to grow by 2030. This is the fastest change in organisational structure of anything measured here. At the same time IBM gives a non-trivial recommendation on the mandate: give the CAIO authority over AI priorities, standards and funding gates — but not ownership of business results. Accountability for the result stays with line leaders. “It is precisely that separation that allows speed without chaos”.
The division of authority is itself a particular case of what CEOs started saying in 2026 about their management design in general.
What CEOs say about the structure of management in 2026IBM CEO 2026
Share of CEOs agreeing with the statement.
IBM's caveat on decentralisation: you first have to explain who decides what, otherwise it turns into paralysis. The specific recommendation: tie at least 30% of senior leadership's compensation to shared results — growth, margin, customer trust — rather than to functional metrics.
If the boundaries are erased and the reward is shared, what remains is to work out who actually takes the decision.
Ownership of the agenda became distributedKPMG Q2’26
Who is accountable for AI-informed or AI-executed business decisions, Q2 2026.
67% agree that the CEO actively owns AI as a strategic priority with clear accountability for results across the organisation. KPMG's wording: “AI leadership is anchored at the top but executed by a broad group of leaders”. · For comparison, in 2025, per 87% of leaders, AI strategy was led by CIOs — ownership was technical.
Confidence in AI at the level of strategic decisions flipped over the yearIBM CEO 2026
62%
2025: CEOs considered generative AI too risky to use in core business functions
64%
2026: CEO are comfortable taking major strategic decisions on the basis of AI outputs
The questions differ: in 2025 the question was about the risk of generative AI in core business functions, in 2026 about comfort in taking strategic decisions on its outputs.
What makes me wary here: the data does not show that the risk itself objectively fell. Documented incidents over the year: 362; only 22% of organisations had clear rules for automated decisions; 77% admit governance is lagging. Meanwhile executives' attitude to using AI in strategic decisions became markedly calmer. IBM's recommendation to separate the mandate from accountability for results is logical as management design: the CAIO sets standards, priorities and funding gates, while the P&L stays with the business. But this data does not test whether such a scheme causally reduces conflicts of interest or improves the result.
8.3 The gap between leader and doer is a constant
While executives rebuild the top floor, something else is visible from below.
One question put to both sides: leaders see it differentlyMicrosoft WTI 2026
Leaders' and employees' answers to identical statements.
Leaders are twice as likely to think rethinking work is rewarded — 21% versus 10%. That is a notable perception gap between levels; it is consistent with the transformation paradox but does not causally explain it on its own.
Three readings from 2026: two taken among frontline employees, the third among all regular AI users.
All three readings were taken in 2026, from different angles and by different researchers — and all three come out at roughly a third or less.
A potentially expensive consequence of the gap — unused capabilityIBM CEO 2026
Two answers from the same CEOs in the same study — but about different things: about themselves and about the workforce.
The skills are acknowledged almost unanimously, and by the same people's estimate a quarter is at work. One thing is paid for, another is used.
“The gap between capability and deployment is an organizational design problem more than a skills problem. If AI is not being used, that should be read as an operational failure, not a qualification problem”.IBM CEO Study 2026, pp. 36–37
The 86% / 25% gap really does look like unused capability: the same CEOs simultaneously consider skills widely available and regular application limited. That is consistent with IBM's framing of an organisational design problem, but gives no cost for that gap. Separately, the perception gap between executives and employees repeats across several cuts and different sources. That makes it less like one-off noise, but does not prove a universal “property of hierarchy”. Strategic clarity is therefore better measured from both ends — among those who formulate it and those who have to work to it; the employee view is critically important here as an outcome metric.
Idle capacity
86/25
per cent of CEOs say employees have the skills to work with AI — and the same CEOs put regular AI use at a quarter of the workforce
A similar self-assessment gap appears at two independent sources and in several cuts. That makes the pattern less like one-off noise, but does not prove a universal “property of hierarchy”. Strategic clarity is better measured from both ends — among those who formulate it and those who have to work to it; the employee view is critically important here as an outcome metric.
8.4 What leadership support is associated with
The gap has a mirror side: at BCG, employees who report explicit leadership support simultaneously report regular use and positive effects of AI markedly more often. This is a comparison of groups, not an experiment.
A 40–47 point gap between frontline employees with support and without itBCG 2025
The same level of hierarchy; the groups differ by reported leadership support, but may also differ on other unobserved factors.
The differences between the groups BCG splits by reported leadership support are +41, +40 and +47 points. These are associations between groups, not a causal effect of support.
Employees in the leading group work in exactly that environment: their manager uses AI openly (85% versus 64%), sets quality standards for AI work (83% versus 57%), creates space for experiments (84% versus 61%), encourages more ambitious redesign of work (87% versus 61%).
8.5 Where the tension is
That is why some benchmarks in the list below are deliberately taken not from the leader but from the frontline employee.
8.6 What to measure
Having a CAIO and separating their mandate from line accountability — benchmark 76% of organisations in 2026
AI operating model: centralised / hybrid / decentralised — benchmark +36% ROI at the first two
Share of senior leadership compensation tied to shared rather than functional results — recommendation ≥30%
The board's AI expertise — benchmark 8% → 40% “substantial”
Clarity of the AI strategy as measured among frontline employees — benchmark 33% BCG, 26%Microsoft
Words matching actions as measured among frontline employees — benchmark 28%
The gap between “employees have the skills” (86%) and “employees use AI regularly” (25%)
Dimension 7
The labour market
Executives explain the “skills exist — they are not used” gap by people's qualifications. Market data explains it differently: qualification on the labour market is doing something of its own, and it looks nothing like a survey.
What we measure. What is happening to people and the labour market — from objective job-posting and financial-reporting data rather than surveys. This dimension rests on PwC and Stanford HAI: the only two sources that do not ask executives but look at the market.
9.1 How it looked in 2025: aggregate optimism
PwC analysed around a billion job postings and thousands of financial reports across six continents and reached four conclusions, each of which refuted a widespread fear.
The common view and what the data showed
01
AI is not affecting productivity yetThe industries most able to apply AI show three times greater growth in revenue per employee
02
AI pushes wages downWages are growing twice as fast in the industries most exposed to AI
03
AI cuts the number of jobsThe number of openings is growing in practically every occupation exposed to AI
04
AI devalues automatable rolesWages are growing in both automatable and augmented roles; automatable ones even raise their skill requirements faster
The first of these four conclusions is worth unfolding: here is what “three times” looks like broken out by quartile of exposure.
Productivity growth by quartile of AI exposure, industriesPwC 2025
A reading from the 2025 edition: the unit of observation is the industry, base 2018. The top quartile outpaces the bottom by more than threefold.
Here the quartiles are by industry; a similar series in the first part of the report is built by company. Wages across the same quartiles, in the same order: 16,7% / 12,6% / 8,2% / 7,9%.
Productivity and wages are only two cuts of that billion job postings; here is what else PwC pulled out of it.
56%
the AI skills premium on average (2025) — every industry analysed pays it
+38%
growth in openings in AI-exposed occupations against +65% in the rest over 5 years
−7…−9 pp
fall in demand for formal degrees in AI-exposed roles
>50%
women in AI-exposed occupations in every country analysed
2025 was the year the data refuted the fears, and PwC played that role honestly. But one line in its own conclusions already contained the story of 2026: there are more jobs everywhere, but in AI-exposed occupations they grow more slowly. A year later that single caveat turned into the main plot.
And here is what I think is underrated in that tile about women. If they are the majority in AI-exposed occupations in every country, then both the benefits and the risks of both tracks are distributed unevenly by gender — and not one source tracks this beyond stating it. One of the most conspicuous holes in the data.
9.2 Where it stands in 2026: the aggregate split into two tracks
PwC 2026 takes the same corpus (now more than a billion postings) and changes the coordinate system itself. Instead of “AI-exposed / not AI-exposed” there are three categories by what exactly AI takes over.
And once the market is recounted across those three categories, it turns out it is not divided evenly at all.
What the job market is made of: three categories by what AI takes overPwC 2026
AI barely touches itWhat AI does Almost nothing · Examples Cooks, builders, mechanics
02
Expertise requirements riseWhat AI does Takes the simple tasks, leaves the expert ones · Examples Radiologists, recruiters, air traffic controllers
03
The entry bar fallsWhat AI does Takes the expert tasks, leaves the simple ones · Examples Software developers, loan officers, financial managers
PwC's historical analogy: in the 1980s spreadsheets lowered the entry bar for bookkeeping clerks (a gradual but steady decline in numbers) and raised expertise requirements for financial analysts (sharp growth continuing into the 2020s). The same mechanism, a different scale. PwC offers four questions for forecasting the fate of any role: how AI changes the level of expertise required; how demand for the role and the supply of workers will change; where a human is needed for oversight or atypical cases; which external forces constrain the use of AI.
Those four questions are needed where a role's future is still open. But there is a point in the market where the answer is already visible, without any forecast — the very beginning of a career.
9.3 Entry to a profession — the sharpest story of the year
Entry-level openings most exposed to AI are seven times more likely demanding skills once expected of an experienced specialist: motivational leadership, team building, managing people and stakeholders, process management, mentoring, data-driven decision-making. The top quartile by AI exposure is the only one where the number of early-career openings has plateaued.
PwC's conclusion: entry to a profession is not disappearing, it is “growing up in its requirements”. Companies and education will have to redesign onboarding, mentoring and early careers so that people demonstrate the skills of an experienced specialist much sooner.
The phrase “entry is not disappearing, it is growing up in its requirements” sounds calmer than the data itself. If an early-career opening demands leadership, mentoring and stakeholder management, a real question arises: where will a person acquire those skills if they are already becoming a condition of entry? PwC's data shows a rise in requirements of the candidate's level, but says nothing about the pay for such openings — so a “mid-level work at junior pay” conclusion cannot be drawn here.
The learning chain used to look more like this: simple tasks → experience → more complex responsibility. One hypothesis is that some of the first tasks are now automated while the requirements of the next level remain. But job postings do not let you prove that AI specifically “took the first link”. It is more accurate to speak of observable structural pressure on early careers and the risk of weakening the traditional path of gaining experience. Stanford's data on developers aged 22–25 and CEOs' expectations strengthen this signal, but do not turn it into a proven causal chain either.
The signal here is already observable, not merely projected: Stanford records a fall in employment among US developers aged 22–25 by 2024, and 49% of CEOs report plans to cut hiring of junior specialists. Together with PwC's rise in the level required of candidates this increases the risk for entry to a profession, but still does not prove that one cause explains all three series.
9.4 Skills: the earthquake is accelerating
The gap in skill-change speed between the extreme quartilesPwC 2026
By what percentage faster requirements change in the occupations most exposed to AI versus the least exposed.
116%
2.2 times faster requirements change in the occupations most exposed to AI in the 2025 edition the same gap was 66%
The axis shows data years, not barometer editions: 25% for 2023, 66% for 2024, 116% for 2025. Each successive year adds more to the gap than the last.
The more AI is applied, the more distinctly human expertise is valuedPwC 2026
Average EPOCH score of new tasks added to roles since 2022. The EPOCH framework (MIT, Loaiza and Rigobon) covers five abilities where the human matters especially.
Empathy and emotional intelligencePresence and connectionOpinion, judgement, ethicsCreativity and imaginationHope, vision, leadership
The more strongly AI affects an occupation, the more often its new tasks rest on abilities where the human matters especially: 0.47 points in the top quartile against 0.19 in the bottom.
9.5 The AI skills premium and demand for specialists
For skills that live a year and rest on what AI cannot do, the market has started paying extra — only it pays very unevenly.
The AI skills premium by sector, 2025 — per the 2026 barometerPwC 2026
How much higher the average advertised wage is in openings that require AI skills.
The average across 16 sectors rose over the year from 56% to 62%. The spread, though, is more than sevenfold — from 16% in the public sector to 118% in consumer markets — so a “62% premium” does not work as a benchmark for any particular company.
The premium is paid for a skill in an ordinary candidate. Separately from that the market chases AI specialists themselves — and here is how fast.
×8
how many times faster than overall hiring AI-specialist hiring grew in 2025
11,4%
of all 2025 openings in TMT were for AI specialists
1.3m
AI-related openings created over two years — roles that did not exist five years ago
70%
of the skills used today will change by 2030 (LinkedIn forecast)
Everything above in this subsection is annual snapshots. Between two adjacent quarters the picture moves no more slowly.
Employers doubled their willingness to pay more in a single quarterKPMG Q1’26
Share of leaders prepared to pay 11–15% more for candidates with strong AI skills.
In Q4 2025 76% of leaders were prepared to pay up to 10% more and only 22% — 11–15% more.
9.6 What is happening to hiring inside companies
Overpaying for a ready-made person is only one way to close the shortfall. The second companies apply inside themselves, and it is larger in scale.
Reskilled for a different role — from 19% to 29%, for the current one — from 41% to 53%IBM CEO 2026
Share of workers reskilled for a different role and for their current one: actual for the past year and CEOs' forecast for 2026–2028.
For comparison, in 2025 CEOs expected that 31% of workers would require reskilling over three years; 54% were hiring for roles that did not exist a year earlier; 65% planned to close skill gaps with automation.
To reskill is to decide what exactly to teach. Over the past year employers' answer stopped being technical.
Top strategies for securing AI talent: reskilling current staff 87%, hiring into new roles 68%, redesigning existing roles 55%, outsourcing 39%, acquiring a team for its people 17%.
While companies retrain their own, agents have already rewritten hiring itself — and not where it was expected.
In a single quarter agents reached the hiring of experienced specialistsKPMG Q1’26
Share of organisations whose hiring approach agents have already changed.
The claim “AI only touches juniors” stopped being true in exactly this quarter — and stopped not because the pressure on juniors eased, but because the queue reached the experienced.
That is the average picture across everyone. In one function that has lived on data longer than the rest, it thickens to the limit.
Inside the data function the problem is sharper than averageIBM CDO 2025
The state of hiring in the data function.
They have to hire for roles that did not exist a year ago — and only 53% of chief data officers say recruiting and retention deliver the skills they need, against 75% a year earlier. Team composition changes faster than recruiting can adjust to it.
9.7 Where the tension is: whom to believe about jobs
Three sources give three different answers to one question. PwC: headcount at the companies where AI applies most broadly grows markedly faster — 52.2% versus 35.7% against a 2018 base, and the gap widens every year since 2022 — “AI may turn out to be not a job killer but a job multiplier”. Microsoft 2025: 33% of leaders are considering headcount reductions in the next 12–18 months — while 45% put expanding capacity through digital labour among their top priorities, and 47% reskilling. Stanford: the effect is concentrated not in total headcount but in the hiring funnel and among the youngest workers in AI-exposed occupations.
These three answers do not contradict each other if read as different levels of aggregation: total headcount is growing, hiring of the young is contracting, an intention to cut is present in a third of executives. All three are true at once.
From this follows the conclusion I consider the main one in this dimension: the aggregate statement “AI creates jobs” is true and simultaneously useless for a twenty-two-year-old specialist. It describes a sum inside which two opposing processes are running. It cannot be used as a basis for a personal or educational decision.
And one caveat PwC makes honestly while readers usually skip it: the company-level analysis suffers from survivorship in the sample. It contains those who lived through 2018–2025. Those who did not are not in the data — and we do not know whether AI exposure was the reason.
9.8 What to measure
Distribution of your own roles: expertise requirements rise / entry bar falls / AI barely touches it — benchmark 22% / 52% / 26%
Speed of change in required skills in your roles — benchmark gap 116% between the extreme quartiles
The AI skills premium inside your own bands — benchmark 62% market, 16–118% by sector
Share of new tasks resting on EPOCH abilities where the human matters especially — benchmark ×2,5 in AI-exposed roles
Entry-level hiring dynamics and the share of entry roles with experienced-specialist requirements — benchmark +35% versus −10%
Share of workers reskilled for a different role and for their current one — benchmark 19% / 41% over the year, forecast 29% / 53%
Share of hiring into roles that did not exist a year ago — benchmark 54% CEO, 82% CDOs in the data function
Dimension 8
Employee experience
The labour market is visible from outside — through openings, wages and requirements. The last dimension looks from inside: what is happening to the person who already sits at a desk and opens AI every day.
What we measure. What actually happens to a person at work — and why individual gain does not turn into organisational gain.
10.1 Use: the “silicon ceiling” has been broken
Last year this dimension had one immobile line — frontline employees stuck at the level of the year before last while the upper floors moved ahead. That is where we start.
Daily or several times a week — four consecutive readings from one source.
Frontline employeesManagers and Leaders
+23 pp over the year among frontline employees — the fastest change in any behavioural indicator over the year. In 2025 this was called the “silicon ceiling”: overall use 72%, but among frontline employees it stalled — 52% a year earlier and 51% in 2025. The source does not explain the dip among leaders in 2025 (88% → 85%).
The vertical gap is not the only one here: the same question asked by country turns the familiar picture upside down.
Geography stays counter-intuitive: the Global South consistently leads the NorthBCG 2026
Regular AI use among frontline employees, 2026. Average — 74%.
In 2025 the picture was the same: India 92%, Middle East 87%, Spain 78%, Brazil 76% against the US 64% and Japan 51%. · By function (frontline employees, 2026): IT 88%, marketing 85%, finance 83%, data and analytics 83%, compliance 76%, HR 75%, procurement 73%, logistics 72%, administrative work 72%, sales 68%, operations and manufacturing 61%.
10.2 Time: savings grow, while converting them into organisational value stays unclear
Two years running the surveys show the same tension, although the measures are not fully comparable between years. In 2025 47% saved more than an hour a day, while guidance on where to put the savings went to only a third. In 2026 52% save at least a working day a week (leaders 60%, managers 52%, frontline employees 42%), while 61% receive limited or no guidance and 45% do not convert the savings into strategic work.
Eight ways of using the time saved, 2025. Respondents could pick more than one.
The most frequent answer is “perform more tasks”. This may be a rise in individual productivity, but the answer on its own does not show whether the time saved turned into margin, revenue or strategic capacity for the organisation.
The answer “perform more tasks” shows where some people direct the time gained. But it does not let you claim that all the time saved automatically went back into the same work, or that this happened without a management decision.
Separately, BCG records a rise in cognitive load and KPMG a rise in load-related resistance. These series are consistent with a risk of intensification, but do not prove a causal chain of “more tasks → load → fatigue”. The practical question is therefore not “where did the time go” but what share of it turned into additional output, margin, revenue or strategic work.
Frontline employees report both smaller time savings (42%) and less clarity about how to use them (66% without adequate guidance). That shows an uneven distribution of the effect by level, but does not establish why it arose.
10.3 Training: the strongest association with regular use
There is, though, one factor with a very strong association with regular use — training. These surveys do not measure its cost or causal effect, so its “cheapness” cannot be compared with other levers.
The five-hour threshold: the main boundary runs between “none” and “any at all”BCG 2025
Share of regular AI users depending on the volume of training received.
Crossing the threshold → +45 ppOver 10 hours → plateau
+12 pp
is added by in-person training — 79% versus 67%
+14 pp
is added by access to a coach — 84% versus 70%
In-person format and a live coach add to the hours, they do not replace them.
Above, training; below, tools: what employees expect and what they get.
The 52-point gap between expectation and provision reproduced twice — on samples of 10,635 and 11,749 people. This is not noise.
No other lever stays unused two years running.
The sharpest threshold in the data
+45 pp
difference in regular use between the “no training” and “up to five hours” groups: 18% versus 63%. This is an association between groups, not an estimate of the causal effect of training.
Shadow AI should be read as a double signal. 54% are prepared to use unauthorised tools, and among Gen Z and millennials — 62%. This may point to friction or the inadequacy of corporate tools, but may also reflect convenience, culture and attitudes to policy. Shadow AI is worth measuring both as a governance risk and as a quality signal about your own stack.
10.4 What AI does to the work itself
How many people and how many hours has been counted; the next question is different: what this work now consists of.
The work changed not in volume but in compositionBCG 2026
What exactly changed, in employees' own view, 2026.
That is what employees say about themselves. The telemetry says the same thing — only in shares of conversations.
The structure of conversations in Microsoft 365 Copilot by telemetry — all work with the assistant, broken down by type of request.
66% of AI users say it let them spend more time on high-value work; 58% — that they produce work they could not have done a year ago. Among employees in the leading group (16% of the sample) — 80%.
And the premium on judgement is rising at the same time. The most important human skills in the view of AI users themselves: quality control of AI output — 50%, critical thinking — 46%. 86% treat AI output as a starting point rather than a final answer, and “stay responsible for the thinking”. Employees in the leading group go further: 43% versus 30% deliberately do part of the work without AI so as not to lose the skill; 53% versus 33% deliberately stop before starting work to decide what AI should do and what the human should.
10.5 The “joy paradox”: work became both better and harder
Work in which you constantly have to decide what to give the machine and what to keep turns out to be both more interesting and more tiring.
Two thirds get more enjoyment. Four in ten — more strainBCG 2026
Both are strongest among leaders.
Rise in job satisfactionRise in cognitive load
Not one lever improves one at the expense of the otherBCG 2026
Top 5 organisational levers: uplift in enjoyment and uplift in measurable impact, percentage points.
Uplift in enjoymentUplift in measurable impact
Value and enjoyment do not look like a simple trade-off here. Stronger organisational practices are associated with both greater reported value and greater enjoyment of work at once; the causal effect of individual practices is not isolated.
10.6 The “honeymoon” and what sustains the effect
Everything so far is a snapshot of one moment. But adoption has an age, and a year on the picture looks different.
Novelty halves. Clarity strengthens. The obstacles do not weaken at allBCG 2026
What helps and what hinders enjoyment of work among those using AI for under six months and for over a year, percentage points.
Novelty as a source of enjoyment halves over the year, clarity of purpose strengthens, and the obstacles — the negative values — do not weaken at all. Each row is one factor in the first six months of working with AI and a year on.
The answer to “what happens a year from now” is unpleasant: everything the enthusiasm of the first six months rests on — novelty, cognitive challenge, the sense of mastering something new — halves over a year. While what is boring to build and impossible to show in a quarterly report — strategic clarity and clear guidance — gains weight instead.
Neither the difficulty of showing your unique value nor a lack of training disappears automatically with experience. So an adoption programme built on enthusiasm alone risks sagging as the technology normalises — especially if the company invests in tools but not in clarity of roles, training and rules of use.
10.7 Resistance came back — in a new form
And here is what that dip looks like once it has already happened: the latest wave caught the turn within a single quarter, live.
Resistance changed its nature: fear went away, fatigue almost doubledKPMG Q2’26
The structure of reasons for resisting agents, Q1 2026 → Q2 2026.
5→20%
express resistance to agents, in total
×4 in a quarter
55→43%
employee agent usage rate — fell over the same quarter
−12 pp
In the structure of reasons for resistance the share of job and skill concerns falls, while the share of answers about load and complexity rises. That is consistent with BCG's signal about rising cognitive load, but two studies do not form a causal chain in time.
A separate detail from the same quarter
41% of leaders would consider introducing a “token race” — incentives for maximum token consumption with internal leaderboards — to encourage AI use.
KPMG itself warns: efforts that put usage metrics above meaningful outcomes risk locking in the wrong behaviour. At the same time 47% agree that employees who use AI effectively and productively outperform the rest.
The pattern of the second quarter looks less like a fear of “I won't cope” and more like fatigue from the load.
On the token race I will say plainly: 41% of leaders considering internal leaderboards for token spend is an alarming signal, because KPMG itself warns about the risk of putting usage metrics above outcomes. Elsewhere in the section you can separately see “more tasks”, rising cognitive load and rising load-driven resistance. These series do not prove a single causal loop, but they give a solid reason not to incentivise AI consumption for its own sake.
The fall in agent use from 55% to 43% in a quarter is a rare reversal among the 2026 series. One wave does not make a trend, especially with KPMG's shifting sample, but it is a point worth checking in the next wave before concluding that adoption has turned.
10.8 The 2025 context: people had no slack left
What remains is to understand why the fatigue arrived so fast — and for that you have to look at the state people were in before any AI.
80%
of the global workforce say they lack the time or energy for their work
53%
of leaders meanwhile say productivity must rise
the capacity gap
275
interruptions a day — every two minutes during working hours
60%
of meetings are unscheduled
Telemetry from the same study: PowerPoint edits spike by 122% in the last 10 minutes before a meeting; messages outside 9–5 — 58 a day, +15% year over year; meetings after 8pm — +16% year over year. 48% of employees and 52% of leaders say their work feels chaotic and fragmented. The gap between leaders and employees on AI was recorded at the same time: familiar or very familiar with agents — 67% of leaders versus 40% of employees; believe AI will accelerate their career — 79% versus 67%; almost a third of leaders save more than an hour a day.
This section explains everything else in the dimension. AI arrived not into an empty system but into one that already had no slack — neither time nor energy — and that was being asked to raise productivity anyway.
In such a system any time saved is absorbed automatically. Not because someone maliciously takes it away, but because the deficit was already there. Interruptions and unscheduled meetings are not consequences of AI, they are the backdrop it landed on. And that is exactly why “perform more tasks” became the first answer: in a system with a negative time balance an hour saved does not create freedom, it pays down debt.
10.9 Where the tension is
Time savings may raise individual productivity, but without rules for reinvestment they do not guarantee growth in margin, revenue or strategic capacity.
Rising cognitive load at low clarity is a warning signal; this data does not allow burnout to be diagnosed causally.
BCG and KPMG show related patterns in different surveys. They are better read as risk signals than as a proven causal chain.
10.10 What to measure
Regular use by level, with a mandatory separate reading for frontline employees — benchmark 74% / 88% / 93%
Time saved and the share actually redirected into strategic work — benchmark 52% save a day a week, 45% do not redirect it
Share receiving clear guidance on the time saved — benchmark 61% do not receive it
Training hours per employee with a threshold at 5 hours — benchmark 18% → 63% adoption when the threshold is crossed
Share satisfied with training — benchmark 36%, unchanged for two years
Cognitive load and its trend — benchmark 41% report a rise
The structure of resistance: fear versus fatigue — benchmark fatigue 28% → 51% in a quarter
Share of unauthorised AI use as an indicator that corporate tools are inadequate — benchmark 54% are prepared to
Part III
Synthesis
The five shifts describe what changed. Then comes the section on what this data does not know.
11. Five shifts, 2025 → 2026
The simplest way to see the year is through the five most comparable indicators that have both a 2025 and a 2026 point. This is not a panel of the same companies and not always literally the same question; the chart shows the direction of change, not a clean causal year-over-year effect.
Five shifts in one picture: what moved over the year
The five most comparable indicators, 2025 → 2026. Sources, samples and denominators differ and are named in the rows.
A post can be created by memo by Monday, while a skills premium is paid by the market and revised over years: the org chart covered three quarters of the scale in a year, the AI skills premium a handful of points, from 56–57% to 62%.
The five segments above are the five most comparable readings found in the sources. Some of the questions and samples differ. The five shifts below are a thematic synthesis, not a causal explanation of each line's movement. The first two shifts are not on the picture: no clean enough 2025 → 2026 pair was found for them.
Shift 1. The bottleneck: data → architecture and the operating model
01
How the problem is framed2025 Data is locked inside functions; every initiative turns into a 6–12-month cleaning project · 2026 Systems are designed for human decision speed and cannot absorb machine speed
02
The anchor figure2025 26% of CDOs are confident data will support new revenue-generating streams · 2026 88% of organisations move workloads, only 25% of those workloads are easily portable
03
The new definition of readiness2025 Stability · 2026 Replaceability — the ability to change provider, model, capability without a rebuild
04
The price of error2025 19% of CDOs with full enterprise-grade architecture · 2026 Cloud spend +48% over forecast, model life 14 months
Architecture you can rebuild on the fly runs immediately into the question of who is accountable for what gets built — and the answer to that changed sign over the year.
Shift 2. Governance: a risk function → the condition for scaling
01
Position on the agenda2025 Control that slows deployment · 2026 The condition without which deployment is impossible
02
The anchor figure2025 Only 22% have clear guardrails for automated decisions · 2026 91–92% put security and risk as the main factor in AI strategy
03
The oversight model2025 Human in the loop on every action: 28% → 61% over the year · 2026 Validation of the result, not of every action: 52–57%, stable for two quarters
04
The economic argument2025 Absent · 2026 ×16 agents, ×4 less budget, +18% operating margin with designed-in control
While companies were rebuilding their control perimeter inside, the market they hire from was rebuilding outside — and there the year went exactly the other way: instead of converging into one frame it split in two.
Shift 3. The labour market: growth for everyone → a split into two tracks
01
The claim2025 AI makes workers more valuable; openings and wages grow almost everywhere · 2026 The market split: in 22% of openings expertise requirements rise, in 52% the entry bar falls
02
The anchor figure2025 AI skills premium 56%; skill change +66% · 2026 Rising expertise requirements: +39% openings and +42% on wage growth against roles with a falling entry bar
03
Entry to a profession2025 Not singled out · 2026 −16% at entry level in AI-exposed fields; +35% for roles with raised requirements against −10% for the rest
04
Check2026 Developers aged 22–25: −20% employment
One hypothesis is that agents take over part of the routine tasks on which beginners used to learn. The charts for openings and for agents move in a compatible direction, but do not on their own establish causality. Let us look at what happened to agents over the same year.
Shift 4. Agents: prototype → operation with the operating model laggingBCG 2026
All three series are employees' assessments BCG: the dominant state in both years is pilots and supervised work, while “integrated” and “not yet deployed” swapped places over the year — 13% → 30% versus 31% → 20%.
01
Integration into workflows2025 13% · 2026 30%
02
Deployment (KPMG)2025 11% → 42% over the year · 2026 53–55%, with chaining agents into a system doubling in a quarter (9% → 18%)
03
Scale2025 24% deployed organisation-wide · 2026 1,661 agents per enterprise by 2027; ×15 active agents in M365 year over year
04
What is lagging2025 Understanding — 33% understand what an agent is · 2026 The operating model: 50% see no rules for “human × AI” teams; 54 incidents a year
05
Independent check2026 Agent deployment in single-digit percentages by function (Stanford)
Someone still has to design the operating model for agents. The last shift shows how ownership of AI is being formalised and how accountability is distributed between business and technology.
Shift 5. Leadership: formalising the AI owner + distributed accountability
01
IBM's research focus (context, not a market metric)2025 Four role-based studies: CEO, CMO, CDO, CAIO · 2026 Two: CEO (business) and Tech Leader (CIO/CTO)
02
Penetration of the CAIO role2025 26% · 202676%
03
Ownership of the agenda2025 CIOs lead AI strategy, in the view of 87% · 2026 Distributed: a named senior executive 34%, CEO or executive committee 32%, business-unit head 14%
04
Attitude to risk2025 62% of CEOs: generative AI is too risky for core business functions · 2026 64% of CEOs are comfortable taking strategic decisions on the basis of AI outputs
05
Functional boundaries2026 77%: business/technology boundaries are obsolete; 85%: all functional leaders must become technology experts
Look at the five shifts together and you see that this is one shift shown from five sides. In each of them the bottleneck moves from the technical plane into the organisational one.
And that, to my mind, explains why the gap between the leaders and everyone else is widening rather than narrowing: access to technology can be obtained comparatively quickly, while the way an organisation is built is rebuilt markedly more slowly.
12. What we don't know
Everything you have read so far rests on six sources — and not one of them opened the management accounts of a single respondent.
This section is obligatory. Without it the report claims more than the data allows.
12.1 Not one source provides independently verified ROI
Every return figure here is either executives' self-assessment (IBM, KPMG), or employees' self-report (BCG, Microsoft), or indirect market indicators (PwC — revenue per employee by industry and company; Stanford — a summary of other people's research). Not one source measured the ROI of AI projects from management accounting data. Hence the “97% versus 8%” pair: two correct answers to two different questions. The only two series that can be considered objective are PwC's market data (job postings, Orbis financial reporting) and Stanford's aggregated series.
That is a complaint about the instrument. The second is heavier and concerns not the instrument but who it was ever held up to.
12.2 Small and mid-sized business is practically absent
Each has its own threshold, and the sources word it differently — here are all five, in the words of their own methodologies.
Whom these samples see at all: the threshold for inclusion in the sample
Revenue of the companies that make it into the sources' samples. Scale in millions of dollars.
The lower threshold for inclusion differs by a factor of twenty, and the typical company at IBM has $14.4bn in revenue. A company on $50m trying the report's benchmarks on itself is comparing itself with something other than itself.
Inclusion threshold
01
KPMGOnly companies with revenue of ≥ $1bn, US only
02
IBM CEO 2026Median revenue $5.8bn, average $12bn, four in five are public
03
IBM Tech 2026Average revenue $14.4bn, average headcount ~33,500
04
BCG23% of the sample are companies of $500m–$1bn, but the distribution skews large
05
PwCA filter for companies with turnover from $50m; an explicit warning about survivorship bias
Practically everything written here describes a company with an average headcount of 33,500. But even about that company less is known than it seems: not one source looked at the same company twice.
12.3 What is not covered at all
Russia and the CIS are absent from every sample without exception.
Public sector is represented fragmentarily: it appears in BCG's and IBM's industry breakdowns, but is nowhere studied separately.
Industry depth: there are cuts, but no full industry studies. At most, three priorities and three challenges per industry.
The economics of running AI — a topic that has only just appeared (KPMG Q2 2026) and so far is measured by one wave of 204 respondents.
The long-term effect on skills. Stanford mentions research raising the question of whether heavy reliance on AI may carry long-term costs for learning and slow skill development, but there is no systematic data. Microsoft records a counter-practice among advanced users — 43% deliberately work without AI so as not to lose the skill — but that is self-report.
Listing what you do not know is easy. Not forgetting it on the next page is harder — so what follows are four rules, short enough to keep in your head.
12.4 How these limits change the reading of the report
1
Read any return figure from an executive survey as sentiment, not as fact.
2
Check any maturity figure against a figure from an employee survey and against an independent source — the divergence is itself informative.
3
Treat “leaders versus the rest” differences as correlation, not as a recipe.
4
Apply all benchmarks to large business; for mid-sized, with an explicit caveat.
I consider this section the most useful part of the report, even though it spoils the impression left by all the previous ones: everything is built around the question of return, and nobody tried to measure it from management accounts.
The second thing worth remembering: every series is a new sample every year. “Leaders do X and get Y” never means “do X and you will get Y”. X and Y may both be consequences of these companies being better built to begin with. The data gives no way to separate that, and every recommendation in this report, including my own comments, carries that caveat.
And third, closest to practice. The company-size limitation is the most frequently ignored of all: IBM's $5.8bn median revenue and KPMG's $1bn threshold are forgotten the very second people start trying the benchmarks on themselves.
Above is a share of executives, below a share of initiatives: one cannot be subtracted from the other.
The top three rows are expectations, the bottom four are actuals. All data is from 2025.
How “expected ROI” is counted at IBM.IBM CEO 2025 The respondent compares an initiative's actual result with what was stated in its business case. That is self-assessment, but self-assessment against an internal criterion: it is harder to lower a business case retrospectively than to answer “yes, there is an effect”.
How that differs from KPMG's “return”.KPMG Q2’25 KPMG asks whether an improvement is observed on a metric (productivity, profitability, quality). There is no threshold and no comparison with plan. Hence 97–98% against 25%.
PwC's superstar effect.PwC 2026 The top 20% within the quartile of companies where AI applies most broadly show productivity growth of 163% against 33.5% for the whole group. Selection is by actual growth in revenue per employee, not by self-assessment. Twenty per cent of companies take 74% of the total gain.
Why the top of the range rises while the bottom barely does.It is not about the volume of spendbut about how capital is reallocated and refreshed. Organisations in the lower part of the distribution spend comparably, but on assets with a 14-month life and no portability.
“Machine speed” is not a metaphor. Behind it are three dimensions, on each of which systems designed for humans hit a ceiling.
Volume. By 2027 enterprises expect an average of 1,661 agentsIBM Tech 2026. At hundreds of decisions per agent per day that is hundreds of thousands of autonomous decisions daily.
Autonomy. Today 25% of operational decisions are taken without a humanIBM CEO 2026; CEOs expect 48% by 2030.
Frequency. A model's life is about 14 monthsIBM Tech 2026, and the main reason for replacement (71%) is a better one appearing. Most organisations' planning cycle is longer than the asset's refresh cycle.
How it looks from the employee's side.Microsoft WTI 2026 Microsoft, independently and by a different method, reaches the same conclusion: organisational factors explain 67% of the effect of AI, individual ones 32%. The strongest single factor is the organisation's AI culture, roughly 2.5 times stronger than the most significant individual one. This is an association, not a proven cause.
KPMG's claim that “governance is the condition for scaling” can be checked by calculation.KPMG Q1’26
How agent incidents end
These are parts of a whole: the structure of consequences of 54 incidents a year at an average organisation. The first two types are exactly the ones where four hours to contain is critical, and 17% of incidents are of that kind.
1 661
agents per enterprise by 2027 — 38% more than today
54
agent incidents a year at an average organisation, 17% longer than 4 hours
The five roles of designed-in control.IBM Tech 2026 Platform teams own the shared guardrails — telemetry, model registry, identity, logging, rollback. Risk and compliance set policy thresholds. Architecture sets reference patterns. Business domains own outcomes inside the guardrails. Incident response maintains predefined shutdown and recovery procedures.
IBM's minimum production standard: every agent is registered, has an owner, is observable and can be stopped. Four properties without which an object should not reach industrial operation.
Why manual gates do not scale. The cost of approvals is linear in the number of agents. The cost of designed-in control is almost fixed. At 1,661 agents the difference between those two curves becomes insurmountable — hence the observed ×16 agents at a quarter of the budget share.
A regional detail. Growth of the agent fleet 2026 → 2027 is 60–87% depending on region, but growth in the number of incidents differs by almost twofold. In organisations with weak governance more agents mean proportionally more incidents; in those with strong governance the rate stays relatively flat even as deployment accelerates.
The term “democratisation” sounds positive, but at PwC it means that prior expertise becomes less of a barrier to entry — and is paid less as a rare skill.
Definition. The category is defined by how the required level of expertise changes in a role's job postings. “Professionalisation” means AI takes the simple tasks and expertise requirements rise. “Democratisation” means AI takes over some of the expert tasks and the entry bar falls.
PwC's historical analogy.PwC 2026 In the 1980s spreadsheets lowered the entry bar for bookkeeping clerks — a gradual but steady decline in numbers. And in the work of financial analysts they raised expertise requirements — sharp growth continuing into the 2020s. The same mechanism, a different scale.
Four questions for forecasting a role's fate. How does AI change the level of expertise required? How will demand for the role and the supply of workers change? Where is a human needed for oversight or atypical cases? What external forces — regulation, process bottlenecks — constrain the use of AI?
Baselines. Skills and openings are measured against 2018, wages against 2021.
Self-reported time savings grew, and in 2026 a marked gap in guidance and in converting time into strategic work persists at the same time. The measures of the two years are not fully comparable, so it cannot be claimed that “conversion has not risen”. Below is where employees direct the time gained and what gaps are visible by level.
The gap runs across one rung, not down the whole ladder.BCG 2026 66% of frontline employees are left without guidance and exactly the same share of managers; among leaders it is 52%. On converting time into strategic work the ladder is monotonic: 58 / 43 / 36, a 22-point gap.
Training has not moved in two years.BCG 2026 88% expect major reskilling within five years, 36% consider themselves adequately trained. Both figures are identical in 2025 and 2026 — on samples of 10,635 and 11,749 people. The 52-point gap reproduced twice; this is not noise.
What BCG says to do about it.BCG 2026 The first of five imperatives for CEOs: “Change the scoreboard: measure value, not reach. The time people save leaks out of the organisation unless it is tracked and deliberately reinvested”. IBM offers a specific commitment: fix the share of gains to be reinvested (60–80%) before the gains appear.
A five- to sixfold spread is not noise and not a contradiction. These are three different questions, and they form a scale of strictness.
The third reading, Stanford, is given only in words — “single-digit percentages across almost every business function”, without figures.
The stricter the question, the smaller the figure.KPMG Q2’26 “Have you deployed agents” (at least one, somewhere) is KPMG. “Are agents built into an end-to-end workflow” is BCG. “What share of functions actually runs on agents” is Stanford. Your own maturity is worth measuring by the strictest definition.
Microsoft's telemetry is the only telemetric reading here of organisational agent use.Microsoft WTI 2026 The number of unique active agents in the M365 ecosystem grew 15× year over year, 18× in large enterprises. It does not ask an opinion, it counts objects — but it says nothing about value either: an agent created and launched twice counts twice.
Unevenness by function.KPMG Q3’25 IT 95%, operations 89%, risk 66%, finance 45%, marketing and sales 18%, HR 2%. A 47-fold spread between the extreme functions — per KPMG Q3 2025.
In 2025 IBM published a separate study for almost every role in the C-suite. In 2026 two are left — and they are divided not by job title but by perimeter of management.
2025 — four, by role
01
CEOstrategy and growth
02
CMOmarketing
03
CDOdata
04
CAIOAI as a separate mandate
2026 — two, by perimeter
01
CEO Studythe business perimeter
02
Tech Leader Studythe technology perimeter, CIO and CTO together
What this means. The structure of publications shows a change in IBM's research focus, but is not itself a market metric. IBM does not comment on whether this relates to converging functions, editorial strategy or a new programme design; so any conclusion about causes here stays a hypothesis.
The substantive 2026 data emphasises the seam between business and technology more strongly: formalisation of the CAIO is rising, accountability is spread wider, and CEOs talk about the blurring of functional boundaries. It is those indicators, rather than the structure of publications, that support the claim of a new perimeter of management.
In parallel at other sources. PwC changed its cut from “AI-exposed / not AI-exposed” to “roles with rising expertise requirements / roles with a falling entry bar”. Microsoft introduced the zones “leaders' zone” / “forming” / “stalled” instead of demographic cuts. BCG moved from levels of hierarchy to the “deploy — reshape — invent” scale. Four independent sources changed their coordinate system in one year — and all four towards “what the organisation does” rather than “who is accountable”.
The pair “97 and 8” is not the only place in the report where two numbers look comparable and are not. Here are four more.
Actually measures
01
97% name profitability a top metric / 8% reported a revenue gainThe choice of indicator versus the result in the accounts
02
53% “deployed agents” / 30% “agents in workflows”The presence of an object versus how embedded it is in a process
03
86% “have the skills” / 25% “use it”Capability versus the deployment of capability
04
21% of leaders / 10% of employees: “rethinking work is rewarded”Intent versus how it is perceived
05
81% “data strategy is integrated” / 26% “data will hold what comes next”The foundation's progress versus confidence that it is sufficient
The common structure of all five pairs.KPMG Q2’25 On the left is a figure named by whoever is accountable for it. On the right is a figure measured by whoever it acts on, or by an independent counter. One cannot be subtracted from the other: the pair has no common denominator.
How to rephrase your own reading. Replace the question “do you observe an effect” with “how large is the effect relative to what was stated in the business case”. The first gives 97% (KPMG, top ROI metrics), the second 25% (IBM, share of initiatives that achieved ROI). The second is checkable.
Three admissions from 2025, made by different audiences in different studies.
Why two thirds of CEOs got the exit-from-pilot date wrong.IBM CEO 2025 In 2024 more than two thirds expected to leave the pilot stage by 2025. In 2025, 60% were still in pilots. The share expecting to drive growth with AI fell over the same period from 67% to a little over 50%.
The marketing cut is the harshest.IBM CMO 2025 19% of pilots with the expected ROI, 25% scaled to several business units, 14% enterprise-wide. And only 22% of organisations had set clear rules for using AI in automated decision-making.
The price of the “just don't fall behind” approach became measurable only in 2026.IBM Tech 2026 In 2025 the lower bound of the ROI distribution was not published — only averages were. In 2026 it appeared: the bottom of the distribution is −20%. Investing “before understanding value” was a description of the market in 2025. In 2026 that behaviour acquired a price tag.
Five independent readings, put to different audiences in different studies, produced answers in the 19–29% range.
A coincidence like that is not chance. It means the constraint is not a particular problem but a general ceiling: roughly a quarter of organisations have actually reached a state where the foundation is ready for the next step.
And this is against objective progress.IBM CDO 2025 Integration of data strategy with the technology roadmap rose from 52% to 81% in two years. Platforms on top of scattered data — from 41% to 75%. The IT-budget share for data strategy — from 4% to 13%. Enormous progress, confidence a quarter.
Which means requirements grew faster than the foundation. That is the content of the rewording of the bottleneck: not “the data is bad” but “we are running and the horizon is receding faster”. That is exactly why IBM in 2026 changes the very definition of readiness from stability to replaceability.
54 incidents a year at an average organisation. The distribution by time to contain looks reassuring until it is recalculated into absolute numbers.
Recalculated into absolute numbers.IBM Tech 2026 17% of 54 is nine incidents a year, each lasting longer than half a working day. The two main types of consequence — data exposure (37%) and cascading system failures (33%) — are exactly the kind where four hours is critical.
The key indicator of governance maturity is not the number of incidents in itself but the ratio of growth in incidents to growth in the number of agents. In organisations with weak governance more agents mean proportionally more incidents. In those with strong governance the rate stays relatively flat even as deployment accelerates.
Macro context. Documented AI incidentsAI Index 2026 per Stanford — 362 over the year.
BCG broke 9,923 regular AI users down along two axes and looked at the share reporting measurable impact from AI.
Two ways to spend a budget — and a fivefold difference in the result.BCG 2026 Buying tools where there is no clarity is +5 points on the share reporting measurable impact. Creating clarity where there are no tools is +25.
Why almost everyone spends the first way. Licences are bought within a quarter and land in the report as a measurable action. Strategic clarity is not procured, not installed and not reported to the board. Meanwhile employees measure it themselves: the strategy is called clear by 31% at “deploy”-level companies against 52% at “reshape/invent”-level ones. The largest gap in that comparison, though, is not clarity but people's participation in process redesign: 12% versus 43%.
What “high clarity” means in the reading.BCG 2026 AI is declared a priority, there is a clear AI strategy, and there is guidance on what to do with the time saved. Not “I have heard about the strategy” but “I understand what is required of me and why”.
Confirmation from the other side. IBM:IBM CEO 2026 CEOs who actively redesign cross-functional work meet their business goals more than twice as often; those who redesigned five key areas as a single system, four times as often.
Each of the seven findings has figures that support it and figures that argue against it. Both are here with a link to the report; the button next to a finding opens its full breakdown.
7. A change of cutSupportthe chief AI officer role 26% → 76% in a year · Argue against this is an observation about what the analysts chose to measure, not about the market itself
The word “ROI” appears in the article next to +250%, next to 25% and next to 33.5% — and those are three different quantities with different denominators. What follows are the 2025 and 2026 series in full, a breakdown of what exactly each source measures, and the reason you can no longer compare yourself with the market average.
Twenty per cent of companies take three quarters of the gain
These are parts of a whole: shares of the total gain from AI. The stratification of return is not “everyone grows at different speeds” but concentration.
The top of the ROI distribution is +250%, the bottom — −20%: 270 percentage points between them. The source's wording: “the ceiling is rising faster than the floor — and not because anyone spends more, but because they allocate and refresh capital differently”.
Productivity at the companies most exposed to AI 33,5% versus 24,0% at the least.
The top 20% within the leaders — 163%, five times the group.
20% of companies take 74% of the gain (PwC 2026, citing PwC's 2026 AI performance study).
Only 8% of CEOs report a more than negligible revenue gain from AI over the year.
Three independent readings give the same distribution: a long tail near zero and a small group with multiple returns. That makes the comparison “our ROI is above/below the market average” meaningless — you have to compare yourself with a quartile, not with the average.
An objection to keep in mind. None of these series is built on management accounting. PwC comes closest to objectivity (Orbis financial reporting), but measures revenue per employee by industry and company with AI exposure, not the return of specific AI projects. Exposure is the opportunity to apply AI, not the fact of applying it; PwC itself notes that actual use may lag exposure.
The bottleneck's move is visible only by comparing two formulations — IBM 2025 on scattered data and IBM 2026 on “machine speed”. What follows are both in full, the figures that support the shift, and the objection worth keeping in mind: the data foundation objectively improved over the year, requirements simply grew faster.
What explains the effect of AI
Microsoft, 20,000 respondents, 29 factors. The values show a statistical association, not a causal effect.
Verbatim from the foreword of the CDO Study: data is locked inside functions — finance has its own, HR its own, marketing, supply chain and legal their own; there is no shared taxonomy, no common standards, no end-to-end visibility. The consequence: every AI initiative turns into a data-cleaning project six to twelve months long; teams spend more time finding and reconciling data than getting conclusions from it.
IBM introduces the term machine speed — a shift in the volume, autonomy and frequency of decisions, and what that shift demands of corporate systems. The diagnosis changes: the problem is not the absence of data but that architecture, governance and funding models are designed for human decision speed.
The key rewording of readiness: readiness used to mean stability — conformity with a known operating model optimised for continuity, cost and predictability. “When models, platforms and tools change faster than planning cycles, stability turns into a constraint”.
Microsoft reaches the same conclusion from the employee rather than the infrastructure: organisational factors explain 67% of the reported effect of AI, individual ones 32%; the strongest single factor is the organisation's AI culture, roughly 2.5 times stronger than the most significant individual one. The authors' wording: “The real question isn't whether people have the right skills. It's whether the organization is built to unlock them”.
By 2025 the data foundation already looks decent, and that has to be acknowledged: integration of data strategy with the technology roadmap 52% → 81%, having a platform on top of scattered data 41% → 75%, the IT-budget share for data strategy 4% → 13%. The bottleneck moved not because data stopped being a problem, but because requirements grew faster than progress.
×16 agents at a quarter of the budget share is a result, not a mechanism. What follows is the arithmetic that stops manual oversight coping, two strategies that both fail to scale, and the five roles across which designed-in control is distributed.
Governance: recognised, justified — and not built
The upper panel is recognition, the lower one implementation. The gap between them is larger than in any other dimension. The denominators differ and are named in the labels: US executives at KPMG, technology leaders at IBM, employees at BCG.
The calculation that makes manual oversight impossibleIBM Tech 2026
By 2027 enterprises expect an average of 1,661 agents, which is 38% more than today's level. Each agent takes hundreds or thousands of decisions a day. The result is hundreds of thousands of autonomous decisions daily per organisation. IBM's wording verbatim: “Manual governance does not cope with this arithmetic”.
IBM describes the fork technology leaders land in:
Speed first. Business units push ahead, governance catches up. Local speed rises, visibility and containment degrade. The consequence: 54 agent incidents a year on average, of which 17% are high severity (more than four hours to contain).
Safety first. Deployment slows under the weight of checks and approvals. Immediate exposure falls, but learning stops, the competitive position weakens, and manual oversight becomes unmanageable anyway.
The source's conclusion: “One path trades safety for speed, the other speed for safety. Neither scales, and both accumulate strategic debt”.
Growth of the agent fleet 2026 → 2027 by region is 60–87% (the leader is the Middle East and Africa, +87%), by industry 63–83%. But growth in the number of incidents differs by almost twofold: in organisations with weak governance more agents mean proportionally more incidents; in those with strong governance the rate stays relatively flat even as deployment accelerates.
The source's practical conclusion: governance investment should match operational risk. In regulated sectors, sovereignty-sensitive markets and customer operations, underinvestment turns AI growth into operational exposure.
Dena Almansoori, Group Chief Technology and Innovation Officer, ADNOC: “Control shifted from approving inputs to continuous oversight of outputs and outcomes — from gates to guardrails”.
“Professionalisation” and “democratisation” are PwC's terms, not judgements. They describe what exactly AI takes out of a role: the simple tasks or the expert ones. From this you can see how PwC changed its coordinate system, what the three diverging series are made of, and which roles break the scheme — from nursing assistants to childcare managers.
In 2025 the barometer divided occupations by AI exposure (the presence of tasks where AI applies) and by type of impact — augmentable (AI complements judgement and expertise) versus automatable (AI performs tasks autonomously). Methods: the Felten et al. exposure index, the IMF's distinction between augmentability and automatability.
In 2026 the divider is different: what exactly AI takes out of the role (PwC 2026, drawing on the work of Teeselink and Carey, 2026).
Share of openings
01
AI barely touches itMechanics AI barely applies · Examples cooks, builders, mechanics
02
Expertise requirements riseMechanics AI takes the simple tasks, the expert ones remain · Examples radiologists, recruiters, air traffic controllers
03
The entry bar fallsMechanics AI takes the expert tasks, the simple ones remain · Examples software developers, loan officers, financial managers
The source's illustrations: for a recruiter, AI automated CV screening and contract negotiation remained (professionalisation). For a stock clerk, AI took over inventory management and moving goods remained (democratisation).
The spreadsheets of the 1980s. For bookkeeping clerks they performed the hardest part of the work — the entry bar for the role fell and numbers went into a gradual but steady decline. For financial analysts they provided a tool of unprecedented analytical power — expertise requirements in the role rose, numbers went into steep growth that continues into the 2020s, with rising wages.
Four questions for forecasting a role's fatePwC 2026
PwC suggests not guessing by category but taking a specific role apart along four axes:
Expertise. How does AI change the level of human expertise required?
Supply and demand. How will demand for the role and the supply of workers change as it is rebuilt?
AI's limits. Where is a human needed — oversight, quality checks, atypical cases?
External forces. What constrains the use of AI: regulation, process bottlenecks, availability of people?
Nursing assistants — a role with rising expertise requirements, but the effect may be limited by regulation, difficulty of integration into workflows and a shortage of people willing to enter a demanding profession.
Childcare managers — a role with a falling entry bar, but the number of openings more than doubled since 2019 (+111%) because of enormous latent demand; wages meanwhile rose only 8%. That is exactly the mechanism David Autor describes: lowering the required expertise widens the pool of suitable candidates, and the occupation grows in numbers while wages stagnate.
The surveys show a rise in reported time savings, and in 2026 a large gap between the savings and clear rules for reinvesting them. The 2025 and 2026 metrics are not fully identical, so this is not a clean “conversion” series. What follows is where employees direct the time and what BCG and IBM propose.
Shares of respondents naming a way of using the savings (BCG 2025; the total exceeds 100%, multiple choice):
The most frequent answer is intensification, not reallocation. That is the mechanics of the leak: without explicit guidance, time saved goes by default into increasing the volume of the same work.
Training is the second channel of the same leak, and it has not moved over the year: 88% expect to need major reskilling within five years, 36% believe they have been trained adequately; neither figure changed from 2025.
The first of five imperatives for CEOs: “Change the scoreboard: measure value, not reach. Adoption tells you that people are using AI, not that it pays off. The time individuals save leaks out of the organisation unless it is tracked and deliberately reinvested. Look at business outcomes, not at usage”.
IBM's parallel recommendation with a specific number: fix the share of productivity gains to be reinvested — typically 60–80% — before the gains appear and before the next budget cycle closes.
The five- to sixfold spread between KPMG, BCG and Stanford is not a dispute about facts but three different questions put to different people. What follows are all three questions verbatim, KPMG's quarterly curve with the break in it, and a separate telemetric reading of organisational use.
agents per enterprise expected by 2027 — 38% more than today's level
25%
of operational decisions today are taken by AI without human intervention
65%
plan or are already introducing AI autonomy in demand forecasting; in inventory optimisation — 61%
How the question is worded
01
BCG“are agents integrated into broader workflows” — an employee's assessment
02
IBM“do you feel fully ready for the expected scale” — a technology leader
01
Stanford HAIHow the question is worded secondary data on agent deployment across business functions · Resultsingle-digit percentages almost everywhere
02
KPMGHow the question is worded “what stage of working with AI agents is your organisation at” — an executive's self-assessment · Result53–55% deployed
The spread is not noise: these are three different questions. “The organisation has deployed at least one agent somewhere” ≠ “agents work in an end-to-end workflow” ≠ “a function is systematically executed by agents”.
Why KPMG's series cannot be read as an exact time series
The full quarterly table — all three stages across five waves — is in the deep dive “The agents' path from experiment to work”. What matters here is one feature of it: the Q4 2025 deck presents 26% with the wording “more than double the 11% in Q1”, ignoring the already published 33% and 42%. The likely reason is a change of question wording or base. The 2026 charts are internally consistent; the Q4 2025 wave cannot be included in the chart. The breakdown is in the appendix “Divergences between sources”.
Microsoft: the number of unique active agents in the Microsoft 365 Copilot and SharePoint ecosystem grew 15× year over year, 18× in large enterprises (rolling 28-day window).
The same source gives an industry pattern: in software and technology adoption is broad (almost one in five companies using agents), in manufacturing fewer companies but a markedly larger scale within each. Individual behaviour is the same across industries — the difference is in where agents are embedded and how deeply.
1,661 agents per enterprise by 2027, +38% on the current level.
25% of operational decisions today are taken by AI without human intervention; the expectation for 2030 is 48%. The decision types going first: price updates, inventory allocation, rerouting shipments, automatic incident remediation.
Most executives already plan or are implementing AI autonomy in demand forecasting (65%) and inventory optimisation (61%).
Areas where agents should not decide, per IBM: regulatory reporting, material disclosures, sensitive legal judgements.
A change of cut is easy to dismiss as report packaging. But IBM surveys the same audiences for years and changes its division only when the old one stops producing differences — and PwC, Microsoft, BCG and Stanford changed their cut after it, each in their own way.
Functional boundaries declared obsoleteIBM CEO 2026
All three rows are shares of CEOs surveyed by IBM in 2026. The share of organisations with a chief AI officer itself — 11% in 2023, 26% in 2025, 76% in 2026 — is a different kind of quantity and is shown alongside in the article as a separate line.
What was asked2025 how your function applies AI · 2026 who decides and who is able to execute
The change is confirmed from inside the studies themselves: 77% of CEOs say the boundaries between business and technology are obsolete, and that talent and technology roles are converging; 85% — that all functional leaders must become technology experts in their area.
How the cut changed at the others
Publisher
01
PwC2025 AI exposure: most / least · 2026 mechanics: rising expertise requirements / falling entry bar · What the change means from “how much AI” to “what exactly it takes over”
02
Microsoft2025 three trends about the new shape of the firm · 2026 three levels: employee / leader / organisation · What the change means from describing the phenomenon to levels of intervention
03
BCG2025 “Momentum builds, but gaps remain” · 2026 “Strategy matters more than tools” · What the change means from measuring adoption to measuring the conditions
04
Stanford HAI2025 8 chapters, “Science and medicine” together · 2026 9 chapters, science and medicine separate · What the change means domain applications grew into subjects in their own right
11% (2023) → 26% (2025) → 76% (2026) of organisations have a chief AI officer. 100% of CEOs expect the CAIO's influence to grow by 2030.
This is the fastest change in organisational structure of everything measured here. As recently as 2025, 66% of CAIOs merely expectedthe role to appear at most organisations within two years — the forecast came true in one.
Why this is data, not packaging
IBM's change of cut is an interesting indirect signal, but not proof that the market itself changed. It can be read as a move from a functional view to a split between the business and technology perimeters. But from the structure of publications alone you cannot conclude that the differences between a CMO and a CDO have narrowed: that stays a hypothesis, not a measured result.
Everything on the outside view in one place: where AI in science and medicine already works and where it hits a wall, how the AI Index editions themselves are built across two years, and what changed between them.
80 150
AI publications in the natural sciences in 2025 — 26% more than in 2024
5,8–8,8%
of scientific output is taken by AI depending on the field — against less than 1% in 2010
33%
accuracy on Earth-observation questions, with agents' code failing to run in 58% of cases
83%
by this much doctors cut the time spent on clinical notes with automatic generation
Where AI in science hits a wall
Both scales are the share of correct answers, higher is better. For comparison: on ChemBench frontier models outperform the average human chemist, while agents' code fails to run in 58% of cases. Capabilities have no single scale of difficulty.
The 2025 edition has 8 chapters, 457 pages. The 2026 edition has 9 chapters, 425 pages; the “Science and medicine” chapter was split in two. Below are also the chapters that did not make it into the main text of the report.
AI publications in the natural sciences reached about 80 150 in 2025, +26% on 2024. AI takes 5,8–8,8% of scientific output depending on the field — against less than 1% in 2010.
Frontier models outperform the average level of human chemists on ChemBench (more than 2,700 questions), but fail at reproducing published research: under 20% on ReplicationBench in astrophysics, 33% accuracy on Earth-observation questions, with agents' code failing to run in 58% of cases.
Smaller models outperform larger ones. MSAPairformer (111m parameters) beat previous methods on ProteinGym; GPN-Star (200m parameters) outperformed a 40bn-parameter model.
In 2025 astronomy got its first general-purpose model (AION-1, trained on more than 200m celestial objects from 5 major surveys), its first visualisation benchmark and a 100 TB training set.
AI ran a full weather-forecast cycle for the first time: Aardvark Weather replaced the traditional numerical pipeline with a single machine-learning system; FourCastNet 3 builds a 60-day global forecast in under 4 minutes, 8–60 times faster than previous approaches.
On end-to-end research tasks the best agents deliver roughly half the result of a PhD expert: 38.8% against 83.5% on PaperArena; around 17% on BixBench (real bioinformatic analysis).
Tools for automatic clinical note generation spread widely in 2025: doctors report up to 83% less time on notes and a significant reduction in burnout; one hospital system reported a 112% return on investment.
The FDA authorised 258 AI medical devices in 2025, predominantly via device-modification pathways not requiring new clinical trials; only 2.4% of devices with clinical studies relied on randomised trial data.
A multi-agent system (Microsoft's AI Diagnostic Orchestrator paired with OpenAI's o3) scored 85,5% on complex published clinical cases against 20% for doctors working without their usual tools. Multi-agent setups in general give a gain in diagnostic accuracy of 7% to more than 60% over single-agent baselines.
AI summaries appear at the top of 84–92% of medical search queries on Google; for symptoms and general health questions, in 92% of cases.
The evidence base, though, is thin: a review of more than 500 clinical AI studies showed almost half relied on examination questions rather than real patient data; only 5% used real clinical data.
2025: two thirds of countries offer or plan to offer computer science in school — twice as many as in 2019, with the greatest progress in Africa and Latin America, though access in African countries is limited by the absence of electricity in schools. 81% of US computer science teachers agree AI should be part of the core curriculum, but fewer than half feel ready to teach it. The number of AI master's graduates in the US almost doubled between 2022 and 2023.
2026: enrolment in computer science at US four-year institutions fell 11% between 2024 and 2025, while master's output in AI-related fields grew 17% over 2023–2024. Four in five US school and university students use AI for study, but school policies lag: only half of middle and high schools have them, and only 6% of teachers consider them clear. More than 90% countries offer computer science; China and the UAE introduced compulsory AI education from the 2025/26 academic year. The number of new AI PhDs in the US and Canada grew 22% over 2022–2024, and the entire increase went into academia rather than industry — a reversal of a decade-long trend.
2025: US states lead on legislation amid slow federal progress — from 1 law in 2016 to 49 in 2023 and 131 in 2024 alone. Mentions of AI in the legislative proceedings of 75 countries grew 21,3% (1,889 against 1,557), and more than ninefold since 2016. State investment in infrastructure: Canada $2.4bn, China $47.5bn on semiconductors, France $117bn (elsewhere in the report €109bn), India $1.25bn, Saudi Project Transcendence $100bn. National AI safety institutes were launched in 2024.
2026: national AI strategies appear fastest in countries that had no formal policy five years ago — more than half of the new 2024 strategies came from developing economies. AI sovereignty is becoming a central principle of national policy, but infrastructure is unevenly distributed: Europe and Central Asia grew state AI supercomputing clusters from 3 to 44 between 2018 and 2025.
2025: the share seeing more benefit than harm in AI products is 55% across 26 countries. Regional differences are persistent: China 83%, Indonesia 80%, Thailand 77% against Canada 40%, the US 39%, the Netherlands 36%. Two thirds of people expect a significant effect of AI on everyday life within 3–5 years.
2026: the gap between experts and the public is 50 percentage points: 73% of experts expect a positive effect of AI on how people do their jobs, against 23% of the public. Trust in one's own government to regulate AI is lowest in the US — 31%; globally the EU enjoys more trust as a regulator than the US or China.
What is covered in this section
In the article money is shown in one cut — the US against China. Here is the same capital at full height: a ten-year series, cloud providers' capital expenditure, the economics of the AI companies themselves and the caveat about China's state funds, without which a two-country comparison misleads.
$bn
01
Corporate AI investment, worldwidemore than doubled
02
Generative AIgrowth over 200%, almost half of all private funding
03
Funding events of $1bn+almost doubled
The decade's trajectory. Cumulative investment grew roughly fortyfold since 2013 — to $581.69bn in 2025, up 129.9% over the year.
Capital expenditure. Large cloud providers accelerated capital expenditure; Google reported more than $150bn in annual capital expenditure for 2025.
The obligatory caveat on China. The private investment figures probably understate China's total spend: state guidance funds placed an estimated $184bn in AI companies over 2000–2023. Any US-versus-China chart without that footnote misleads.
What changed qualitatively. Capital did not merely grow — the regime changed: private investment added 127,5% over the year, and generative AI stopped being a segment and became half of all private funding.
The economics of the AI companies themselves. Revenue at frontier companies is growing at historically fast rates, but spending on compute and infrastructure is hitting records at the same time.
On the 88% ceiling and the gap between “we use AI” and “AI does the work” — now from both sides of the gap: who added the most, what AI delivered in specific functions and how small that effect is for most. And what to measure adoption by, when the fact of use no longer tells anyone apart from anyone.
The key contradiction inside a single source. AI adoption in organisations reached 88% — practically saturation. Agent deployment meanwhile measures in single-digit percentages by function. The gap between “we use AI” and “AI does the work” is the central fact of the year, and it is visible inside one, the most objective, source.
53% in three years — faster than the personal computer or the internet. Adoption correlates strongly with GDP per capita, but there are outliers both ways: UAE 64%, Singapore 61% — higher than income predicts; the US at 24th place with 28.3%, despite the country leading on both investment and model development.
What this means for comparing companies
Adoption in organisations has exhausted itself as an indicator: at 88% it no longer tells companies apart. What does tell them apart are next-level indicators — the share of functions with agents in an end-to-end workflow, the share of decisions without a human, the share of employees using AI regularly.
The surplus appears in the article as a single figure — but it has a point from last year, and the whole conversation about return rests on the distance between them. What follows is what this implies for corporate calculations, including the “what others are getting” benchmark.
Estimated consumer surplus from generative AI in the US: $112bn (early 2025) → $172bn a year (early 2026), growth of +54%. Median value per user tripled over the same period. Most of these tools remain free or nearly free.
Consumer surplus grew by half in a year
The value users receive above what they pay. Almost none of it passes through companies' revenue.
Three implications for corporate calculationsPwC 2026
A significant part of the value AI creates does not pass through companies' P&L at all. It goes to the consumer directly, bypassing corporate revenue. That partly explains why, with investment doubling, only 8% of CEOs see a more than negligible revenue gain.
Internal ROI and societal return diverge. A technology can be economically successful at the level of society and still not pay off at the level of an individual company — a classic situation for infrastructure technologies in an early phase.
The “what others are getting” benchmark misleadsif counted by consumer value. Corporate return is measured in a different coordinate system.
Olympiad gold and analogue clocks are only two points on the jagged frontier. Here it is in full: where models have caught up with or overtaken the human, where they fail — from robots to arithmetic, how far the gap with China has closed, and by how much a query to a model became cheaper over the same time.
Gains on the 2023 benchmarks in a single year
The gain is in percentage points, not per cent: this is a difference of shares and cannot be added to growth percentages.
SWE-bench Verified: performance rose from 60% to almost 100% of the human baseline in a single year. For context: in 2023 systems solved 4.4% of tasks of this class, in 2024 — 71.7%.
Leading models reach or exceed the human baseline on PhD-level natural-science questions, multimodal reasoning and olympiad mathematics.
Gemini Deep Think took gold at the International Mathematical Olympiad.
Gains on the 2023 benchmarks in a single year: +18.8 pp on MMMU, +48.9 pp on GPQA, +67.3 pp on SWE-bench.
The source's own term is the jagged frontier: capabilities are distributed unevenly, and the failures are where you do not expect them.
The same model that takes gold at the IMO reads an analogue clock correctly 50.1% of the time.
Agents on OSWorld (real computer tasks across operating systems) rose from 12% to ~66% success, meaning they still fail roughly every third attempt.
Robots succeed in only 12% of household tasks, against 89.4% success at manipulation in the RLBench simulation.
Complex reasoning stays a problem: systems do not reliably solve tasks that have a provably correct solution — arithmetic and planning, especially on examples larger than those in training.
The lead changed hands several times from the start of 2025: in February 2025 DeepSeek-R1 briefly matched the best US model, by March 2026 the best US model leads by 2.7%, and all year the gap stayed within single-digit percentages. The US still releases more top-tier models and more cited patents, China leads on volume of publications, citations, patents and industrial-robot installations.
Industry produced more than 90% of the notable frontier models of 2025, and the most capable systems became the least transparent: training code, parameter counts, dataset sizes and training duration are no longer disclosed for several of the most compute-intensive systems, including models from OpenAI, Anthropic and Google.
For comparison, the 2025 edition noted an improvement in transparency: the average Foundation Model Transparency Index score then reached 58%. Over the year the trend reversed.
The unevenness of productivity gains is the most practical seriesAI Index 2026
Gain
01
Customer support14–15%
02
Tasks requiring deep reasoningweaker or negative
The source's wording: the gain is largest in structured, measurable work where the result is easy to track. That explains why corporate ROI is smeared: the effect concentrates where it is easiest to count, which means the sample of successful use cases is systematically biased.
A separate warning from the source: evidence is emerging that heavy reliance on AI may carry long-term costs for learning and slow skill development.
The cost of a query to a GPT-3.5-class model (64.8 on MMLU) fell from $20.00 to $0.07 per million tokens between November 2022 and October 2024 — more than 280-fold in about 18 months. Depending on the task, inference prices fall 9- to 900-fold a year. Hardware performance grows 43% a year (doubling every 1.9 years), cost falls 30% a year, energy efficiency grows 40% a year.
Electricity, one Taiwanese contractor and distrust — the article names them but does not take them apart. Here is each constraint from its own side: how much states have announced they will invest, what training models costs in tonnes and in water, how incidents are growing and why “responsible AI” is not one scale but a set of conflicting requirements.
State investment in AI: announced programmes
All values in dollars. These are announced programmes, not funds actually spent.
29.6 GW
comparable to the peak consumption of New York State
5 427
data centres in the US — more than ten times as many as in any other country
72 816
tonnes of CO₂ equivalent — estimated emissions from training Grok 4
50 pp
gap between experts and the public in assessing AI's effect on work: 73% versus 23%
AI data-centre capacity grew to 29.6 GW — comparable to the peak consumption of New York State.
The US hosts 5 427 data centres, more than ten times any other country, and consumes more energy than any other country.
Almost every frontier AI chip is produced by one Taiwanese company — TSMC, which makes the global AI hardware supply chain dependent on a single contractor; TSMC's US expansion began operating in 2025.
Models' compute grows roughly 3.3-fold a year since 2022.
Estimated emissions from training Grok 4 — 72,816 tonnes of CO₂ equivalent.
Annual water consumption for inference at GPT-4o alone may exceed the drinking water needs of 1.2m people.
The training series over the years: AlexNet (2012) — 0.01 t; GPT-3 (2020) — 588 t; GPT-4 (2023) — 5,184 t; Llama 3.1 405B (2024) — 8,930 t. For comparison, the average American emits 18 tonnes a year.
In 2024 this was already driving the energy agenda: Microsoft announced a $1.6bn deal to restart the Three Mile Island reactor for AI; Google and Amazon also concluded nuclear power agreements.
The growth rate holds at around 55% a year; the absolute number keeps rising.
Practically all leading frontier-model developers report on capability tests, but reporting on responsible-AI tests remains selective.
The awkward finding of 2026: improving one dimension of responsible AI, say safety, can worsen another, say accuracy. This means “responsible AI” is not a single scale but a set of conflicting requirements.
A separate line from 2025: the contraction of publicly available data. The share of restricted tokens in actively maintained domains of the C4 dataset jumped from 5–7% to 20–33% over the year because of new anti-scraping protocols. The consequences are for data diversity, model alignment and scalability.
Models trained to be explicitly unbiased retain implicit biases — in tests GPT-4 and Claude 3 Sonnet more often associate negative terms with Black people, women with humanities rather than STEM fields, and men with leadership.
The number of AI researchers and developers relocating to the US fell 89% since 2017, 80% of that in the last year. Gender gaps in AI talent have not narrowed in any country since 2010.
Everything on the section is gathered here, plus a table of all the ROI metrics that appear in the reports.
Every ROI metric and what each actually measures
+The full list13
01
Share of initiatives with the expected ROIWhat it measures hitting their own business case · Who answers the question CEO
02
The same in marketingWhat it measures the same · Who answers the question CMO
03
Share scaled enterprise-wideWhat it measures reach, not return · Who answers the question CEO
04
The same in marketingWhat it measures the same · Who answers the question CMO
05
Average AI ROIWhat it measures ratio of effect to cost · Who answers the question CAIO
06
ROI premium for having a CAIOWhat it measures the difference between groups · Who answers the question CAIO
07
ROI premium for the operating modelWhat it measures the difference between groups · Who answers the question CAIO
08
ROI premium for early replaceabilityWhat it measures the difference between groups · Who answers the question CIO/CTO
09
Expectation of measurable ROI within 12 monthsWhat it measures expectation, not fact · Who answers the question US senior management
10
Achieved or expect ROI within 12 monthsWhat it measures mixed wording · Who answers the question US senior management
11
Revenue gain from AI over the yearWhat it measures self-assessment, narrow wording · Who answers the question CEO
12
Expected revenue growth with the three pillarsWhat it measures expectation · Who answers the question CIO/CTO
13
Revenue gain at CEOs of AI-benchmarked companiesWhat it measures actual growth, self-assessment · Who answers the question CEO
01
Top and bottom of the ROI rangeValue +250% / −20% · What it measures the distribution across the sample · Who answers the question CIO/CTO
02
Top ROI metricsValue productivity 98% · What it measures the share naming a metric · Who answers the question US senior management
03
Growth in revenue per employeeValue 33,5% / 163% · What it measures an external proxy from financial reporting
04
Financial effect by functionValue savings <10%, revenue <5% · What it measures self-assessment of the size
The practical conclusion: there is not one ROI metric built on management accounting. The closest to objectivity are PwC's series on Orbis financial reporting and HAI 2025's soberest series on the size of the effect (savings under 10%, revenue gain under 5% as the most common values).
What is covered in this section
Both columns — expectation and reality — had to be cut in the article to the most telling items. In full they are longer: what was claimed all year and what came of it, plus the marketing cut, the harshest, separately.
Top ROI metrics rose in a single quarter
Near-unanimity on all three metrics — while 25% of initiatives delivered the expected ROI over three years. This is a reading of sentiment, not a financial fact.
85%
expected positive ROI on scaled savings initiatives by 2027
77%
expected the same on growth and expansion initiatives
65%
prioritised AI use cases by ROI; 68% claimed to have clear metrics for measuring it
68%
claimed to have clear metrics for measuring innovation ROI
Eight out of ten CEOs demanded that both AI-driven savings and AI-driven growth be scaled within 18 months.
85% expected positive ROI on scaled savings initiatives by 2027, 77% — on growth and expansion initiatives.
65% prioritised AI use cases by ROI, 68% claimed to have clear metrics for measuring innovation ROI.
Through the whole first half of the year KPMG recorded near-unanimity on return metrics: productivity 98% (94% a quarter earlier), profitability 97% (was 95%), improvement in quality of work 94% (was 87%).
82% of leaders agreed that their industry's competitive landscape would look different in 24 months.
Expected investment growth doubled over the year: in generative AI for 2025–2027 31% against the 15% stated a year earlier for 2024–2026; in traditional AI 31% against 13%.
Expected investment growth: what was stated a year earlier for 2024–2026 and what is stated now for 2025–2027. Both lines arrive at the same point — 31%.
Over three years only 25% of initiatives delivered the expected ROI; only 16% were scaled to enterprise level.
Average AI ROI per CAIO data — 14%.
60% of CEOs are still in the pilot stage, although a year earlier more than two thirds expected to be out of it by 2025.
The share expecting to drive growth with AI over the next five years fell from 67% to a little over 50%.
Only 52% of CEOs said they get value beyond cost reduction.
60% of organisations invested predominantly in pilots; since 2023 only 25% of AI initiatives delivered the expected ROI.
The marketing cut is the harshest
Over three years: 19% of AI pilots delivered the expected ROI, 25% were scaled to several business units, 14% were rolled out enterprise-wide.
The source's explanation is “pilot purgatory”: stacks designed for a different era, a Frankenstein portfolio of disconnected technologies, data fragmentation, too many tools. The average portfolio grew to nine tools, adding two in two years; 68% of CMOs agree that simplifying infrastructure will raise operational efficiency.
investing out of fear of falling behind, and its priceIBM CMO 2025
58% CMO admit that the risk of falling behind makes them invest in technology before they understand its value.
64% CEO say the same in their own words: the risk of falling behind pushes them to invest in some technologies before clearly understanding the value they bring.
And yet only 37% of CEOs agree that in technology adoption it is “better fast and wrong than right and slow”. So they invest out of fear but do not consider themselves reckless.
72% CAIO say the organisation risks falling behind without measuring AI's impact. 68% of the same CAIOs admit they launch AI projects even when they cannot assess their impact — because the most promising opportunities are often the hardest to measure.
This is the most honest pair of figures from 2025.
Risk aversion and reluctance to disrupt what exists
Lack of expertise and knowledge
Absence of a clear innovation strategy
Limited budget
Inadequate technology
Insufficient or poorly integrated data
The top two places are organisational, not technological. This is an early version of the conclusion everyone will reach in 2026.
The jump from $124m to $207m is one point in a series KPMG has run for six waves. What follows is the whole series with the source's comment on each wave, the categories the money goes into, and a check on whether it agrees with IBM's series on the IT-budget share.
KPMG's series: AI spending forecast twelve months outKPMG Q1’26
67%
said AI would remain a top investment priority even in a recession over the next 12 months
59%
will keep investing in AI regardless of their ability to measure tangible ROI, Q4 2025
31%
annual growth in the AI spending share was forecast by IBM a year earlier — the actual turned out higher
73%
agree that tariffs will sharpen the focus on AI efficiency and optimisation
The source's comment
01
Q1 2025the series' base point
02
Q2 2025the year's low
03
Q3 2025“the peak of the year”
04
Q4 2025“close to the Q3 peak”; AI called recession-proof
05
Q1 2026“almost double the same period last year”
06
Q2 2026“holding steady from last quarter”
The structure of investment by category (share of organisations planning to invest $10–49m):
Top categories
01
Q1 2025R&D 34%, new technology solutions 28%, data and analytics 26%
02
Q2 2025cybersecurity and data protection 67%, risk and compliance 52%, operations 48%
03
Q3 2025data and analytics, R&D, procurement of generative AI technology
04
Q1 2026cybersecurity 63%, data and analytics 61%
05
Q2 2026data and analytics 48%, R&D 45%
Recession resistance. In Q4 2025 67% said AI would remain a top investment priority even in a recession over the next 12 months; in Q1 2026 — 79%. In Q4 2025 59% also indicated they would keep investing regardless of their ability to measure tangible ROI.
Growth +71% in two years. The source's wording: this means hundreds of millions of dollars flowing into an investment category that behaves nothing like traditional IT.
For comparison, a year earlier IBM recorded growth in the AI spending share of 62% over 2022–2025 with a forecast of annual growth of 31% two years ahead. The actual pace turned out higher than the forecast.
Do the two series agree
Yes. KPMG measures absolute sums at large US companies, IBM the IT-budget share globally. Both show acceleration in 2026: KPMG almost doubling year over year, IBM +71% in two years. There are no divergences.
The effect of tariffs on the AI agenda, Q2 2025: 73% agree tariffs will sharpen the focus on AI efficiency and optimisation; 66% — that tariffs will raise the cost of using AI (energy, data); 57% — that tariffs will increase the total planned volume of AI investment; 44% — that tariffs will stimulate AI adoption. In Q3 2025 the tariff-driven focus on efficiency held at 76%.
Four barriers are what is visible at once. Behind them is a bigger question: what to measure with at all, once traditional metrics stopped coping, and how much it costs to maintain what already works.
“Achieved or expect ROI”: the series by quarter
The wording is mixed: “achieved or expect” adds fact and intent together, so the series reads as sentiment rather than result.
78%
agreed that traditional business metrics are becoming insufficient for measuring AI's impact
66%
of organisations have set up AI cost monitoring dashboards
Three of the four barriers almost doubled. Measuring got harder, not easier — as organisations left pilots the question shifted from “does this work” to “how do you count an effect smeared across a process”.
The intermediate point of Q4 2025: risks (privacy and cybersecurity) were named the biggest challenge to demonstrating ROI (74%) and a key factor in revising generative AI investment strategy (71%).
Q2 2025productivity 98%, profitability 97%, quality of work 94%
02
Q3 2025productivity 97%, profitability 94%, quality of work 91%
03
Q4 2025improved analytics for senior management decisions rose from 62% to 83%; productivity, profitability and generated revenue remain top metrics
04
Q1 2026improved analytics from senior management stays the top metric for a second quarter running (83%); customer experience 64%
The key admission: as early as Q3 2025 78% agreed that traditional business metrics are becoming insufficient for measuring AI's impact; 78% named pressure to demonstrate value to investors or the board a critical factor in their generative AI strategy for the next six months. In Q2 2026 pressure to demonstrate value to investors or the board was named the top factor 76%.
Expectation of measurable ROI within 12 months: 57% (Q3'25) → 59% (Q4'25) → 62% (Q1'26, worded as “achieved or expect”).
The new theme of 2026: the cost of running AIKPMG Q2’26
Q2 2026 raises the question of operating economics for the first time, not just of investment.
Only 26% say the operating costs of their AI systems are fully visible today.
What is already in place to manage costs: cost monitoring dashboards 66%, cost review at AI approval 61%, architecture and prompt-design standards 47%, usage or token budgets 36%.
Among the main barriers to deploying agents a new skill has appeared: economic literacy in AI — understanding usage-, token- and inference-based costs — 35%.
One of the factors shaping strategy: 66% agree their organisation is able to make its AI strategy future-proof — adapt to new technology and regulation, scale responsibly, avoid lock-in to particular tools and vendors.
68% name balancing AI scaling against the carbon footprint as a strategy factor.
Confirmation from IBM's side: 84% of technology leaders have not operationalised financial management of AI, 85% do not have full real-time visibility of AI spend. From an interview: “We are starting to see how quickly costs can get out of control. Building agents is powerful, but the key question is: how much does each agent actually cost?” (Miguel Santos, CTO of MasOrange).
Separating operational and strategic AIIBM Tech 2026
IBM's central recommendation on portfolio discipline:
01
What it isOperational AI proven use cases at scale, infrastructure for agents · Strategic AI experiments, bets on new capabilities, decisions on refreshing models
02
The management logicOperational AI TCO: optimisation, monitoring, vendor discipline · Strategic AI portfolio: explicit owners, success criteria, exit points, reallocation of capital as data arrives
03
What is acceptedOperational AI minimising cost · Strategic AI higher unit cost in exchange for speed, replaceability and asymmetric upside
The source's wording: assessing all investment through a single cost-optimisation lens systematically undervalues investments with asymmetric upside.
The practical minimum for portfolio accounting: if for each use case you cannot see spend, owner, model, business goal and expected return — portfolio management is not happening.
The average useful life of an AI model is roughly 14 months.
71% of technology leaders name as the main reason for retiring or replacing a model the appearance of better models, 60% — changing business needs and use cases.
The source's wording: most models are retired not because they broke but because the organisation worked out what works.
This breaks business cases built on three-to-five-year asset life cycles and requires mechanisms for refreshing and reallocating capital in real time, not on an annual schedule.
Where financial discipline gives an advantageIBM Tech 2026
Organisations where technology leaders built strong financial discipline:
+17% likelihood that cost accounting was in place from the very start of cloud migrations
+16% likelihood of joint IT and finance decision-making with shared accountability
×1,5 likelihood that cloud initiatives exceeded expectations
×2,4 more AI agents deployed without increasing the AI or IT budget
×3 likelihood of claiming full readiness for an AI surge
Multiples against other organisations where technology leaders built strong financial discipline: how many times more often (readiness for an AI surge, exceeding expectations on cloud initiatives) or more (number of agents deployed without budget growth). The vertical line is the comparison base, 1×. The first two items on the list, +17% and +16%, do not go on this scale: they are percentages, not multiples, and stay in the text.
Two lines from IBM's table made it into the article — the ones that argue with each other. In full it has four rows: you can see where what did not become growth went, and that the same miss had already happened a year earlier.
The main purpose of AI investment, forecasts for 2026 and the actual:
Forecasts for 2030:
55%
the forecast share of “growth and expansion” for 2030 in the 2025 edition — in the 2026 edition it rose to 72%
60%
of CEOs reported in 2025 that they were still in pilots — although a year earlier more than two thirds expected to be past them
A caveat on extraction. The correspondence of rows and columns was recovered from the coordinates of the slide's text blocks (IBM CEO 2026, p. 15): the values are laid out separately from the labels, and the text extraction order does not match the visual one.
The systematic error on timing is of the order of two years. What was expected by 2026 did not arrive; the analogous expectation was moved to 2030.
Confidence in the outcome meanwhile rises rather than falls. The forecast share of “growth” for 2030 rose from 55% to 72% — after the 2026 forecast missed fivefold.
A change of definition softens the comparison. In 2024 “advanced AI” meant generative, in 2026 agentic. Formally these are two different forecasts; in substance it is the same transfer of expectations onto the next technology wave.
The practical implication for planning: forecasts from industry research about when growth arrives should be shifted by about two years, and you should assume that the dominant mode of return within the planning horizon will remain efficiency rather than growth.
The report's own check: the same failure happened a year earlier — more than two thirds of CEOs expected to be past pilots by 2025; in 2025, 60% reported they were still in them. So this is not a one-off miss but a persistent pattern.
The divergence of questions is clearest verbatim: what exactly KPMG, IBM and PwC asked before each got its own figure. And after that, the requirements for wording under which your own reading becomes checkable rather than merely optimistic.
the most common size of AI savings is less than this: the only series that asks about the size of the effect
20%
the bottom of the ROI range against a top of +250%: which is why you count the distribution, not the average
$112bn
consumer surplus from generative AI in the US in early 2025; a year later — $172bn
Question
01
KPMG Q3’25do you expect to achieve measurable ROI within the next 12 months
02
IBM CEO 2025what share of AI initiatives delivered the expected ROI over the last three years
03
IBM CAIO 2025do you launch AI projects even when you cannot measure their impact
04
PwC 2026did AI deliver a more than negligible revenue gain over the past year
01
KPMG Q2’25Question which categories do you consider top metrics for measuring and demonstrating ROI · Answer productivity 98%, profitability 97%, quality 94%
02
AI Index 2025Question what is the size of the savings and revenue gain in the function · Answer most often <10% and <5%
Why 97% and 8% do not contradict each other
Different questions, different objects, different scales:
KPMG asks about the choice of metric, not about the size of the effect. The answer “productivity” does not mean productivity rose.
PwC asks about the size of the effect on a specific reporting line — revenue — over a specific period.
HAI gives the only series that asks about size: and there the most common values are savings under 10% and revenue gains under 5%.
Adding these figures together or using them as an argument against each other is a methodological error on the analyst's part. The conclusion “executives exaggerate” does not follow from this data.
How to rephrase your own reading so that it is checkableIBM Tech 2026
The problem is not bad faith but that the question is almost always broader and softer than checking requires. The minimum set of requirements for the wording:
Name the line. Not “do we see an effect” but “on which P&L line is it visible and how large is it”.
Name the period and the base. “Over what period against what base” — otherwise the figure is uncheckable.
Separate expectation from fact. “We expect ROI within 12 months” and “we got ROI” are different questions; in Q1 2026 KPMG merged them in the wording “achieved or expect”, and that figure (62%) is therefore not comparable with the 57% and 59% of earlier waves.
Separate the use case from the portfolio. The return of a single use case and the return of the AI portfolio as a whole are different quantities; they are constantly conflated.
Fix the metric before launch. IBM's recommendation: if for a use case you cannot see spend, owner, model, business goal and expected return — no measurement is happening.
Count the distribution, not the average. With a spread from −20% to +250% the average describes nothing.
A significant part of the value AI creates goes to the consumer directly and does not pass through corporate revenue: estimated consumer surplus in the US grew from $112bn to $172bn a year. That is a structural reason why corporate return may lag societal return, and it is worth keeping in mind when interpreting any low ROI.
Everything on the section is gathered here, plus a summary of definitions.
The structure of deployment: Q1 2026 → Q2 2026
The total barely changed, while chaining agents into a system doubled. Growth changed direction from extensive to intensive.
Every source's definition of a “deployed agent”IBM Tech 2026
01
KPMGDefinition the stage of the organisation's work with agents by an executive's self-assessment: experiment / pilot / deployment (within which: scaling, chaining agents together, multi-agent systems) · Consequence the broadest reading; “deployed” = launched at least part of it
02
BCGDefinition “agents are integrated into broader workflows” in an employee's assessment; the alternatives are “used experimentally, in pilots or under human supervision” and “not yet deployed” · Consequence stricter: requires integration into a process
03
IBM Tech 2026Definition “an AI system able to initiate, coordinate or execute multi-step actions with limited human intervention” · Consequence a functional definition, not a stage-based one
04
Microsoft WTI 2026Definition unique active agents on the Microsoft 365 Copilot Agents and SharePoint platform over a rolling 28-day window · Consequence telemetric, the only non-survey one
05
AI Index 2026Definition agent deployment by business function (secondary data) · Consequence the strictest in fact
The practical rule: measure your own maturity by BCG's definition or stricter — the share of functions where agents are built into an end-to-end workflow.
What is covered in this section
Between these two figures stand the barriers — and over three quarters they travelled from people through technology to architecture and strategy. Here is the whole quarterly series, the Q4 2025 wave that is not in the article, and what people began demanding of platforms by the end of the year.
The quarterly series of barriers to agent deploymentKPMG Q2’25
+Barrier11
01
Complexity of agentic systemsQ2 2025 39% · Q3 202571% · Q4 202565% — the top barrier two quarters running · Q2 2026 38%
Q2 2025 — a human barrier. Staff resistance in second place (47%), technical skills first (59%). The problem framed as: people cannot and will not.
Q3 2025 — a technical barrier. The complexity of agentic systems almost doubled (39% → 71%), staff resistance more than halved (47% → 21%). The source's wording: “the complexity of agentic systems almost doubled, while staff resistance fell by more than half”.
Q4 2025 — a structural barrier. Three new items come to the fore: inconsistent use across business units rose from 19% to 45% (the high of three quarters), absence of organisational infrastructure 41% (threefold in two quarters), unclear corporate strategy from 20% to 32%. KPMG calls these three traps: the complexity wall, fragmentation risk, the foundation gap.
This is the key story of 2025: the barrier travelled from people through technology to architecture and strategy in three quarters.
A notable counterpoint: lack of trust and leadership support and budget constraints were consistently named the least significant barriers quarter after quarter — 8% and 14% in Q3 2025. Money and a mandate from the top were never the problem.
The Q4 2025 wave — it is not in the article's chart, which has Q2 2025, Q3 2025 and Q2 2026. In small type is the previous reading where the source prints it. The complexity of agentic systems stayed the quarter's top barrier — 65%; it did not make the chart.
By Q4 2025 executives converge on platform standards that consistently manage identity and permissions, data access, tool catalogues, policy enforcement and observability — “so that every new agent strengthens the system rather than adding fragility”.
Platform requirements: speed of adoption 83%, security, compliance and auditability 75%, reliability 69%, speed of adoption 65%.
The list opens with embedding into existing systems, followed by control: security, compliance, reliability. Speed of adoption closes the four.
The headline figure did not move between quarters, but this is the last point in a series KPMG has run since the start of 2025 — and in the series you can see pilot giving way to deployment. The same path from below, through employees' eyes: what they consider built into the work, what they understand about agents and how understanding changes attitudes.
of employees have limited understanding of what agents are (2026) — it was 67% a year earlier
71%
of those who understand agents well see them as a valuable tool — against 25% among those who have merely heard of them
01
Q4 2025Deployment26% ⚠
⚠ The definition break. The Q4 2025 deck presents 26% with the wording “more than double the 11% in Q1”, ignoring the published 33% and 42%. Use the series as an indicator of direction, not as an exact time series. See the appendix “Divergences between sources”.
Worth noting too is the Q2 2025 wording: “most organisations (90%) have moved past the experimentation stage with AI agents, of which 33% have reached deployment of at least some agents after two consecutive quarters at 11%”. The phrase “after two consecutive quarters at 11%” itself sits poorly with the series.
The structure of deployment: Q1 versus Q2 2026KPMG Q1’26
The quarter's main conclusion: overall deployment stalled, but chaining agents together doubled. This is a move from “a set of separate agents” to “a system” — a qualitative shift amid quantitative stagnation.
How agents change the coordination of workKPMG Q1’26
The ways agents enable cross-functional collaboration:
Method
01
Automating workflows that span several functionsQ1 202673% · Q2 2026 48%
02
Routing information and decisions between teamsQ1 2026 53%
03
Shared knowledge bases and common monitoring dashboardsQ1 2026 51%
04
Aligning shared goals, targets and success metrics across functionsQ2 202664%
05
Supporting joint decision-makingQ2 2026 49%
The shift is marked: in Q1 the main thing was automating end-to-end processes, in Q2 — aligning goals and metrics across functions. KPMG's wording: “a move from experiments to operational integration, where agents help unify execution and ensure consistency”.
Three mutually exclusive states on one base — employees' assessment, BCG. Awareness and understanding are shown as a separate pair: they have a different base.
Awareness and understanding — a different base
One base — all employees surveyed by BCG, three different questions. In the bottom row, movement to the left means improvement: it measures a lack of understanding.
01
Have limited understanding of what agents are2025 67%* · 202652% · Δ −15 pp
02
Believe agents will be able to do at least half their job within 3 years202661%
Breakdown by level, 2026: have heard of agents — leaders 94%, managers 87%, frontline employees 72%. Limited understanding — frontline employees 61%, managers 58%, leaders 39%. Consider agents important within 2–3 years — leaders 86%, managers 75%, frontline employees 57%. Believe agents will be able to do half the job — managers and leaders 65%, all 61%, frontline employees 52%.
The same figures as the paragraph, 2026: awareness and expectation rise with level, limited understanding falls — 61% among frontline employees against 39% among leaders. The paragraph's fourth question, about half the job, is not shown: in it managers and leaders are merged into one group.
The geography of agent integration, 2025: Germany 18%, Brazil 17%, India 16%, Spain 15%, the US 13%, the UK 13%, South Africa 13%, the Middle East 11%, Italy 11%, France 11%, Japan 7%; average 13%.
Understanding changes attitudes. Employees who understand agents well and can explain how they work see them as a valuable tool in 71% of cases against 25% among those who have merely heard of them; consider them a threat to jobs 2% versus 16%. The gap is +46 pp and −14 pp.
The share of decisions without a human has been measured — and with it, how many agents will be taking those decisions. Where autonomy has already been handed over, where it will not be, and by what rules IBM proposes extending it.
By 2027 enterprises expect to deploy an average of 1,661 AI agents — which is 38% more than today's level.
At hundreds or thousands of decisions per agent per day that means hundreds of thousands of autonomous decisions daily per organisation.
Regional breakdown of expected agent-fleet growth 2026 → 2027:
Growth
01
Range across all regions60–87%
Industry breakdown:
Growth
01
Range across all industries63–83%
The source's key observation: the pace of deployment does not determine success — governance maturity does. With the agent fleet growing 60–87%, growth in the number of incidents differs by almost twofold depending on governance maturity. Industries with complex regulated operations — pharma, energy, telecom, retail banking — show the most marked reduction in incidents where governance is mature.
The number of unique active agents on the Microsoft 365 Copilot Agents and SharePoint platform grew 15-fold year over year, and at large enterprises — 18-fold (rolling 28-day window).
The industry pattern (telemetry for March 2025 – March 2026, Microsoft WTI 2026): in software and technology adoption is broad — almost one in five companies among those using agents. In manufacturing there are fewer user companies, but the scale within each is markedly higher. Individual behaviour is the same across industries: people use agents in roughly the same way. The difference is where agents are embedded and how deeply the organisation has integrated them.
Prompts per user in the leading industries: 66.0 / 60.2 / 58.5 / 57.8.
Share of operational decisions taken by AI without human intervention202625% · 2030 (CEOs' expectation)48%
Where this is already happening. Areas where consistency and guardrails can be codified: price updates, inventory allocation, rerouting shipments, automatic incident remediation.
Where it goes next. Decisions requiring speed, scale and continuous optimisation, where a human physically cannot keep up: real-time demand sensing and inventory optimisation across global networks, dynamic workforce scheduling, automatic rerouting during disruptions. Most executives already plan or are implementing AI autonomy in demand forecasting (65%) and inventory optimisation (61%).
Where agents should not decide, on the source's explicit instruction: regulatory reporting, material disclosures, sensitive legal judgements.
The change in the human's role. IBM's wording: the human's role will shift from taking every decision to designing decision logic, setting guardrails and intervening only where exceptions carry material, ethical or strategic consequences.
Focus on decision points, not use cases. Find a small set of repeatable, high-frequency decisions where consistency matters more than judgement, and deploy agents there first. Move people from approving every decision to handling exceptions. Measure success by decision quality and speed.
Govern in learning mode. Review AI decisions by what they revealed, not by the flawlessness of the outcome. Set exit criteria in advance: time to value, decision quality, conditions for closing. Normalise early termination of low-value automation.
Fix reinvestment in advance. Before the next budget cycle closes, agree a fixed share of productivity gains to be reinvested — typically 60–80%.
Extend autonomy deliberately. Increase AI's authority only where guardrails are explicit, auditability is proven and accountability is assigned to a named leader. AI executes where consistency can be codified; people govern and verify; confidence grows on evidence.
The three readings diverge not because they argue with each other but because they answer three different questions. How each is stricter than the last, how the same figure falls apart a second time across functions inside a company, and what to fix so that your own figure has something to be compared with.
The spread between the extreme functions is 47-fold. Any aggregate figure for agent penetration across an organisation hides this distribution.
For context, back in Q1 2025 executives named as the functions benefiting most from agentic AI: technology 76%, operations 74%, risk 56%, finance 39%, marketing and sales 35%, HR 8%. The order held, the values rose — except in HR, where both expectation and fact stayed at the bottom.
AI adoption among frontline employees by function in 2026 shows a similar but not identical picture: IT 88%, marketing 85%, finance 83%, data and analytics 83%, compliance 76%, HR 75%, procurement 73%, logistics 72%, administrative work 72%, sales 68%, operations and manufacturing 61%. In HR employees themselves use AI at a level of 75%, while agents are deployed at only 2% of organisations — the gap between how people use AI themselves and where the organisation puts agents.
The minimum requirements for a figure to be comparable with anything at all:
The denominator. A share of what: of all functions, of all processes, of all employees?
The integration threshold. “Launched” / “in a supervised pilot” / “built into an end-to-end workflow” / “systematically executes a function” — four different answers.
Who answers. An executive systematically gives a higher figure than a doer. The reading is worth taking from both sides, with the gap between them used as a separate indicator.
Chaining agents — separately. The share of agents working in chains connected into a system is a fundamentally different maturity indicator from the total number of agents. Benchmark: 18% of organisations are at the stage of chaining agents into a system.
Do not confuse it with AI adoption. AI adoption in organisations reached 88% and no longer tells companies apart. Agent deployment is the next level of measurement.
Everything on the section is gathered here, plus a glossary of data architecture.
The price of decisions taken yesterday
The upper panel is intent, the lower one capability, and they are counted in different units: organisations versus workloads. It is not laziness: optimising for cost is what created the dependency.
Data architecture — the overall framework and blueprint for managing data: from collection and storage to integration and use. It includes the policies, data models, processes and technologies that ensure availability, security, quality and alignment with business goals.
Data models — conceptual and technical representations of data structure: how elements relate, by what rules they are stored and used. Blueprints for designing databases and applications.
Data products — packaged, reusable information assets that solve a specific business task, by analogy with applications. They join raw data, analytics and conclusions into a ready tool or dashboard that a business team can use without technical training.
Hybrid cloud — a combination of own infrastructure and public cloud services in a single environment where data and applications move between the private and public perimeters. For AI it allows compute needs to be met while respecting data residency requirements.
Data fabric — a data management architecture giving a single intelligent layer over heterogeneous sources and environments. It creates a virtual centralised view of data without physically moving it.
Data mesh — an organisational and architectural approach in which data is treated as a product owned by domain teams rather than by central IT. Decentralised ownership under federated governance; self-service in data discovery, faster innovation cycles.
What is covered in this section
The formal solidity — 81% and 75% — did not appear on its own: two years ago there were different numbers in their place. From here the whole path is visible: from what mark the growth started, what it demanded of people, and what data still trips over when it is called on to feed AI.
The data foundation: 2023 → 2025
The progress over two years is enormous. Confidence that this foundation will hold the next step stayed at the level of a quarter — requirements grew faster than the foundation.
74%
have built mechanisms for continuous collection and refresh of data for availability in real or near-real time
83%
say their teams constantly ask which additional data points are missing
77%
are comfortable relying on the results of agents' work
The CDO's mandate became clearer. The source's wording: in 2023 CDOs knew they had to move from managing data to delivering business value, but that value was defined broadly. In 2025 AI made the mandate specific — use corporate data to feed AI.
The priority shifted. CDOs now name using data for competitive advantage as priority number one — above governance and security as core responsibilities.
CDOs' top priorities and top challenges almost coincide — four of the five main challenges are simultaneously main priorities:
Top challenges
01
Using data for competitive advantageBuilding a data culture
02
Generating actionable conclusions through advanced analyticsIntroducing effective governance and data management practices
03
Introducing effective governance and data management practicesAttracting and retaining talent with advanced data skills
04
Building a data cultureEnsuring data security and privacy
05
Ensuring data security and privacyGenerating actionable conclusions through advanced analytics
The architectural shift.81% CDO say they bring AI to the data rather than centralising data for AI. This avoids the cost and risk of moving data. 81% also say data is prepared for AI predominantly at the function or project level — against reliance on top-down decision structures in 2023.
Continuity.74% have built mechanisms for continuous collection and refresh of data for availability in real or near-real time. 83% say their teams constantly ask which additional data points the organisation is missing.
Trust in agents grew.83% say the potential benefits of deploying agents outweigh the risks; 77% are comfortable relying on the results of agents' work. The source explicitly contrasts this with 2023, when only 44% said leadership trusted the data being collected.
Availability — slow response times to authorised users' queries, low satisfaction
Completeness — share of empty fields, gaps, poor adherence to mandatory fields
Integrity — limited tracking of data lineage, inconsistent entry across systems
Accuracy — high share of errors, incorrect data, validation failures
Consistency — inconsistent formats, codes and nomenclature
Barriers to innovation in data: a focus on short-term results · insufficient or poorly integrated data · resistance to change · lack of expertise · limited budget. The first and third are cultural, not technical.
An example: a single source of truth at IBM CIOIBM CDO 2025
Data was scattered across disconnected sources, including seven different data lakes on different technologies. The consequences: the team could not reach a significant share of the data; governance, access control and processes were inconsistent across repositories; pervasive duplication and data pumping undermined quality, raised operating costs and stifled innovation.
In the third quarter of 2024 the organisation built a single data hub, replacing the scattered stores with centralised control. Users previously extracted data laboriously from different environments; now a single query pulls data from all systems at once.
The result: $5.3m in savings by the first quarter of 2025 through retiring legacy systems and reducing redundancy, plus improved governance from a single source of truth.
Manual matching of invoices, delivery confirmations and purchase orders. They deployed an AI solution for automatic document extraction and matching that copes with complex formats, stamps, poor scans and non-standard layouts.
In four weeks they automated document matching in one unit. The results: invoice processing time fell from 20 minutes to 8 seconds, accuracy above 99%, staff freed for higher value-added work.
The answers fall threefold in five readings at once, and each asked about something different — too tight for coincidence. Look at what exactly was asked: the whole quarter rests on the difference between “agree” and “strongly agree”.
Five independent 2025 readings converged in the narrow 19–29% range:
The formal foundation versus confidence in it
The break between the second and third rows is the confidence ceiling. Five independent readings converged in the 19–29% range: this is not about a particular problem but about a general limit.
68%
consider an integrated enterprise-wide data architecture critical
72%
say using proprietary data is the key to unlocking the value of generative AI
Reading
01
of CDOs are confident their capabilities will support new AI-based revenue streams26%
02
of CDOs are confident the organisation can extract business value from unstructured data26%
03
of executives fully agree the IT infrastructure can support scaling AI across the enterprise25%
04
of COOs say enterprise-wide data architecture is fully built out and integration is scaled across all functions19%
05
of CDOs fully agree they have clear measures of the value of data-driven decisions29%
The distinction matters: in all these readings the question is about strong agreement (strongly agree) or about full build-out, not about agreement in general. The full scale on the key CDO statements:
+The full list8
01
I should be benchmarked on business outcomes to succeed in the roleAgree + strongly agree 92% · Strongly agree45%
02
When data leaders cannot clearly explain the value of data, it puts the organisation's success at riskAgree + strongly agree 86% · Strongly agree35%
03
I can clearly explain how data priorities deliver key business outcomesAgree + strongly agree 85% · Strongly agree34%
04
I actively engage with business leaders to define data needs and set prioritiesAgree + strongly agree 84% · Strongly agree41%
05
I have the resources and authority to execute a strategic data transformationAgree + strongly agree 78% · Strongly agree29%
06
Data leaders have struggled to articulate the business value of their role clearlyAgree + strongly agree 77% · Strongly agree29%
07
There is no template defining the CDO mandateAgree + strongly agree 74% · Strongly agree26%
08
I have clear measures of the value of data-driven decisionsAgree + strongly agree71% · Strongly agree24%
The gap between “agree” and “strongly agree” is 43 to 51 percentage points on every statement. That is itself a diagnosis: the topic is acknowledged, but there is no confidence in one's own position.
The wide bar slides from 92% down to 71%, strong agreement holds in a 24–45% corridor: the distance between the points barely changes.
Half of CEOs admit the pace of recent investment left them with scattered, patchwork technology across the organisation.
68% consider an integrated enterprise-wide data architecture critical for cross-functional collaboration and innovation.
72% say using proprietary data is the key to unlocking the value of generative AI.
The gap between progress and requirements
But confidence that this foundation will hold the next step stayed at the level of a quarter.
The progress is real, but it has not caught up with the growth in requirements. That is exactly what the 2026 rewording of the bottleneck from “data” to “adaptability” means.
87% of executives say effective data protection is essential to getting the most from AI investment. But only half believe they can protect sensitive data and prevent leaks in most AI use cases.
A separate fact from the same series: more than a quarter of AI initiatives were cancelled, postponed or failed to scale because of security problems.
Since readiness now means replaceability, it is worth reading to the end: what exactly IBM redefined and what is paid for it on the way out of a cloud. Next come the barriers in order, the “tomorrow morning” recommendations, and quantum, where readiness must come before deployment.
It was: readiness = stability. Conformity with a known operating model optimised for continuity, cost and predictability. That made sense while capabilities evolved slowly enough to plan around.
It became: readiness = replaceability — the ability to move workloads, change models and absorb new capabilities without launching a large-scale rebuild.
The source's wording: “When models, platforms and tools change faster than planning cycles, stability turns into a constraint”.
Replaceability does not reduce to infrastructure. It includes: integrating new capabilities without replatforming, swapping models without rewriting workflows, preventing a situation where local optimisation creates enterprise-wide vendor lock-in.
of technology leaders investing in quantum are motivated by future-proofing their compute strategy
82%
of CEOs at AI-benchmarked companies already work with partners in quantum ecosystems — against 50% of all CEOs
01
Cloud spend exceeded the original forecastsby 48% on average
02
Of the workloads they plan to move, technology leaders assess as easily portableonly 25%
The main drivers of rising cloud costs: growth in usage and workloads · unoptimised resources · data transfer and egress fees · limited visibility and tracking of costs.
Barriers to portability:
Why the original decisions were rationalIBM Tech 2026
Technology leaders name as the main drivers of the original cloud migration cost optimisation (60%) and enabling innovation (59%).
The source's wording: “What looks like inflexibility now is often rooted in rational decisions taken earlier. Those decisions created value and introduced dependency at the same time”.
This matters for tone: the point is not past mistakes but a change in the optimisation criterion.
The economic effect of early replaceabilityIBM Tech 2026
Organisations that designed replaceability early — keeping workloads portable, limiting hard dependencies and enabling faster model refresh — showed 10% higher return on AI investment in 2025.
The wording: “Readiness is no longer defined by stability alone. It is about the absence of structural barriers to change”.
What to do: the source's recommendationsIBM Tech 2026
Stop immediately:
treating all workloads as equally important to move;
approving new provider lock-in by default;
calling overruns on egress and migration fees “unexpected”.
Tomorrow morning:
Set a rule for new dependencies. New AI deployments should not create hard-to-reverse provider lock-in by default. Require justification when using provider-specific services.
Fund flexibility deliberately. Portability, resilience and model rotation are strategic capabilities, not a side effect of cost optimisation.
Decide where replaceability matters most. Not every workload needs portability. Focus on the platforms, workloads and data domains where provider lock-in will constrain future business decisions.
This month:
Map lock-in exposure across infrastructure, data and AI models, then pick the 2–3 most valuable steps to raise portability.
Establish cost transparency on egress and data transfer fees so portability decisions rest on real economics rather than assumptions.
Introduce a lightweight review mechanism for new AI deployments that lean heavily on proprietary platform services.
Quantum computing poses a fundamentally different readiness problem. Unlike plug-and-play AI applications, quantum requires algorithm and application development plus testing. Meaningful participation requires a long-horizon view — skills, partnerships and operating models long before business value becomes visible. Investments once started are cumulative and hard to reverse.
Here replaceability comes not from technical mobility but from partnerships with a broad set of competencies. 89% of technology leaders investing in quantum are motivated by future-proofing their compute strategy.
The conclusion is worded as: move fast, but recognise where readiness must precede deployment, and that postponing structural decisions will constrain future strategic paths.
From the CEO side: only 46% have a team identifying specific quantum use cases and the business value attached to them; 82% of CEOs at AI-benchmarked companies already work actively with partners in quantum ecosystems against 50% of all CEOs. Where CEOs expect the greatest return from quantum: advancing AI and machine learning 48%, accelerating complex simulations 45%, optimising operations and logistics 39%, exploring new business models 37%, cybersecurity and encryption 32%, shortening R&D timelines 30%, capturing patents and intellectual property 13%, “too early to judge” 12%.
Where CEOs expect the greatest return from quantum: the top is advancing AI itself, and “too early to judge” closes the list.
What sounds quietest is what a competitor cannot copy — and there are several lines here at once: whose models, whose data, whose perimeter. What follows is how the model mix is changing, what sovereignty costs and why access to data inside a company also turned out to be about differentiation.
Own development and training of AI2026 2% · 2030 (expected) 1% · Δ −1 pp
05
Exploring2026 8% · 2030 (expected) 4% · Δ −4 pp
A hybrid strategy — combining custom, general-purpose and small specialised models depending on specific business requirements.
The economic effect of the right combinationIBM CEO 2026
The organisations scaling AI across workflows most successfully use smaller specialised models or a combination of their own and general-purpose models. Compared with those relying predominantly on large pre-trained models, they expect by 2030 (IBM CEO 2026, citing the IBM IBV study “The enterprise in 2030”):
+24% greater productivity gain
+55% greater improvement in operating margin
twice the reduction in process cycle time
Separately: CEOs who systematically build proprietary data and intellectual property into custom models and agents expect that a 13% larger share of 2030 revenue will come from products and services not offered today.
A typical organisation was already using in 2025 11 generative models in 2025 and planned at least 16 by the end of 2026 — as models move beyond text and language to visual, geospatial and other data types.
A complicating factor: a typical executive has been offered by each vendor more than 30 different generative AI use cases. Some organisations already have tens of thousands of AI assets, which creates a large-scale integration and interoperability problem. And half of organisations say scattered technology limits their ability to use their data.
Managing the model portfolio has become a discipline in its own right, and it is one of the main reasons the CAIO role appeared: to help other members of the C-suite find savings, prevent lock-in to a model vendor and optimise the AI portfolio for the specific needs of the business.
Back in 2025 IBM stated this as a separate CEO recommendation: avoid lock-in to a model vendor, promote a model-agnostic approach, encourage experiments with alternative models, compare their performance and pick the most effective for each task rather than defaulting to large and complex ones; adopt open-source technology for interoperability and freedom to adapt.
82% CDO say the organisation treats data sovereignty as a critical aspect of its overall risk management strategy.
Measures being taken to address sovereignty issues:
The survey's third category — “piloting, evaluating or not considering” — is not labelled by the source on the slide.
The source's wording: as legal boundaries take effect defining where data may be stored and processed, organisations will have to adapt practices and potentially duplicate or relocate AI infrastructure. Some CDOs meanwhile see in complying with sovereignty laws a source of competitive advantage.
Proprietary data as a source of differentiationIBM CDO 2025
84% CDO say their unique data products have already delivered significant competitive advantage — customer 360 views, real-time operational dashboards, datasets for financial forecasting.
78% CDO name using proprietary data as a top strategic differentiation task.
72% CEO say proprietary data is the key to unlocking the value of generative AI.
Only 26% of CDOs are confident the organisation can extract business value from unstructured data. The source explains this by the difficulty: it needs advanced analytics, high data quality and a strong governance framework.
CDOs at organisations with higher ROI on data and AI are 25% more likely to say AI changed how they measure the ROI of data.
83% CDO say strategic partnerships strengthen their capabilities and drive innovation; 82% — that partnerships accelerate their AI and data initiatives.
The full scale:
The value of ecosystem data per the source: it gives a full picture of the market that no single organisation can create alone, and early signals of market shifts — a materials shortage or rising costs, say — months before traditional forecasting.
82% CDO say the organisation is wasting its dataif it does not give people access to it for better decisions.
80% say democratising data helps the organisation move faster.
82% say the data team should make interacting with business data simpler and more intuitive.
74% actively promote a culture of responsible data handling among employees; more than two thirds say their role is focused more on enabling use than on preventing misuse — the source calls this a large cultural shift towards openness.
CDOs at organisations with higher ROI are 20% more likely to say the risk of restricting employees' access to corporate data is greater than the risk of granting broad access; 15% more likely to give access through a single interface while preserving security and governance.
In a separate survey of operations executives 90% said that by 2027 agents will let employees go deeper into analytics for real-time analysis and optimisation.
Everything on the section is gathered here, plus IBM's checklists for all three pillars of structural readiness.
The oversight model by 2026
The one model that scales is highlighted: checking every action costs linearly in their number, checking the result does not. At 1,661 agents the difference becomes insurmountable.
Stop immediately: treating all workloads as equally important to move · approving new provider lock-in by default · calling overruns on egress and migration fees “unexpected”.
Tomorrow morning: set a rule for new dependencies · fund flexibility deliberately · decide where replaceability matters most.
This month: map lock-in exposure · establish cost transparency on egress fees · introduce a lightweight review mechanism for deployments on proprietary platform services.
Pillar 2. Governance built into the architecture
Stop immediately: treating policy documents as working controls · approving autonomous systems without rollback, logging and clear ownership · assuming governance scales through new checks and committees.
Tomorrow morning:
Set minimum production standards. If an agent or model is not registered, has no owner, is not observable and cannot be stopped — it is not deployed. Make that non-negotiable.
Redesign governance end-to-end for one high-risk domain. Pick claims settlement, customer service or code generation — and build a model with clear roles, automatic monitoring and defined escalation paths.
Move one manual control into the platform. Access boundaries, drift detection, escalation thresholds or an emergency shutdown — out of committee oversight and into executable code.
This month:
Build a single registry of all production agents and models: what is deployed, who owns it, what it has access to, how performance is tracked.
Automate one governance control across the whole AI life cycle: model evaluation, deployment approval or compliance checking — built into the platform with real-time monitoring and alerts.
Define and enforce minimum standards for autonomous systems: requirements for observability, auditability, rollback and ownership.
Run an incident simulationto test the governance design: simulate an agent failure and check detection speed, containment protocols and recovery processes.
Pillar 3. Portfolio discipline
Stop immediately: assessing exploratory AI bets only by infrastructure payback logic · funding pilots without accountable owners and explicit success criteria · considering AI economics only within budget cycles.
Tomorrow morning: split AI spend into two categories · require visibility at the use-case level · set a shared rhythm for IT and finance.
This month: build an initial portfolio view of AI spend by use case, owner, model and expected value · agree with finance the criteria for splitting operational and strategic AI · start a regular portfolio review.
What is covered in this section
In 2025 governance was reactive: a guardrail for every new alarm, and the next survey wave caught organisations already on a different one. Quarter by quarter — which oversight measures grew, which vanished without trace and on what grounds an AI decision gets overridden at all.
Not ready for autonomous agents, require a human in the loop on every actionQ1'25 28% · Q2'25 45% · Q3'2561%
03
Do not give agents access to sensitive data without human oversightQ1'25 52% · Q2'25 45% · Q3'2563% · Q4'25 60%
04
A human validates the results but does not control every actionQ2'25 25% · Q3'25 42% · Q1'2657% · Q2'2652%
05
High-risk use cases are defined where autonomous decisions are forbiddenQ1'26 44%
06
Build controls into agents with monitoring and evaluation proceduresQ1'26 43%
How the curve reads. Through 2025 organisations tightened oversight as agents went into production: the requirement for a human in the loop grew from 28% to 61%, the ban on access to sensitive data without oversight to 63%. By 2026 things stabilised on the model “a human validates the result but does not control every action” (57% → 52%, the top approach two quarters running). It is the only model that scales at thousands of agents.
Trusted providers travelled from 23% (Q4'24) to a peak of 74% (Q3'25) and back to 44% (Q1'26) — as organisations moved from buying ready-made solutions to a hybrid strategy.
Misuse of AI by bad actors, including cybersecurity and disinformationQ1'25 30% · Q2'2538% · Q3'2545% · Q1'2644% (worded as: cyber and misuse by employees)
02
Trust in the accuracy and fairness of AI outputsQ1'2532% · Q2'25 37% · Q3'25 28% · Q1'26 the question was not asked
03
Personal trust in generative AIQ1'25 top 3 in the view of more than a third
KPMG's waves: in Q1 2025 distrust of accuracy led (32% against 30%), then the leader changed. In Q4 2025 the question was not published — the break does not join up; the hollow Q1 2026 marker means a different wording of the first item. The third item on the list — personal trust in generative AI — the source gives without a figure.
Over the year the priorities swapped: the theme moved from the quality of outputs to malicious use.
By Q4 2025 cybersecurity was named the single main barrier to AI strategy goals. Half of executives planned to allocate $10–50m in the coming year to protecting agentic architectures, improving data traceability and tightening model governance.
The specific threatsof greatest concern: AI-assisted malware 80%, data poisoning 78%, AI phishing 77%.
Measures:72% prioritise deploying agents from trusted technology providers; 60% do not give agents access to sensitive data without human oversight.
Q1 2026: data security, privacy and risk are the top factor shaping AI strategy over the next six months, for 91% of executives.
Q2 2026: the same factor at 92%; followed by pressure to demonstrate value to investors or the board 76%, constraints on hiring and training initiatives 68%, access to cheaper and better models 68%, balancing AI scaling against the carbon footprint 68%.
Factors shaping AI strategy over the next six months: KPMG, the Q2 2026 wave. The shares do not sum to 100% — each factor has its own.
What causes an AI decision to be overridden, Q2 2026:
A third of organisations override AI outputs with no formal criteria at all — a separate, rarely noticed gap in governance.
Accountability was handed out, the tools for it were not — here is what the missing tool is proposed to be made of. Control that does not approve inputs but holds boundaries: who owns which part of it, why this is not centralisation, and how it looked at the oldest bank in Latin America.
The source's wording: for years control in corporate technology was exercised through policies, approvals, committees and sign-off cycles. That approach worked while systems moved at human speed. At machine speed it does not hold.
When agents move from pilots into production they take decisions continuously and in volumes that no escalation path can realistically control. “In this environment control stops being a permissions problem. It becomes a design problem”.
Platform teams own the shared guardrails: telemetry, model registry, identity, logging, rollback and runtime control.
Risk and compliance define policy thresholds, evidence requirements, review triggers and escalation rules.
Architecture teams set reference patterns, control points and interoperability standards.
Business and domain teams own use-case outcomes, integration into processes, exception handling and accountable operation inside the guardrails.
Incident response maintains predefined shutdown, containment and recovery procedures.
This is not centralised command. The source's wording: “Designed-in control does not mean centralised command. It is a federated operating model where platform, risk, architecture and domains each own their part of the control system — so that autonomy can scale inside shared boundaries”.
The source's caveat: designed-in control requires upfront standardisation and engineering effort, but reduces downstream risk and friction as autonomy scales.
Dena Almansoori, Group Chief Technology and Innovation Officer, ADNOC: “Control shifted from approving inputs to continuous oversight of outputs and outcomes — from gates to guardrails”.
Charles Newhouse, Vice President and CTO, Leidos UK: “Traditional governance functions design processes so the function finds it convenient to do its job, forcing the business to adapt. AI inverts that model — making governance, IT and security think like service providers whose job is to enable the business, not slow it down”.
Farid Boutaghane, CTO, Philip Morris International: “You have to let people move at speed, but inside clearly defined lanes and guardrails”.
The oldest bank in Latin America, a public financial institution with regulatory duties and a public mission. The task: to scale AI responsibly, transparently and confidently in an environment where trust and accountability are non-negotiable, while existing governance approaches had stopped coping as AI spread into critical functions.
What they did:
Built frameworks defining clear roles and responsibilities from model design to deployment in critical environments.
Introduced rigorous evaluation processes for general-purpose models, including benchmark comparisons and vulnerability assessment.
Implemented real-time monitoring, proactive alerts and customised metrics, consolidating a single system of control and trust.
They are testing an integrated approach on generative AI use cases for compliance, performance and risk management requirements.
The key results: automated governance across the whole AI life cycle, reducing manual oversight and speeding deployment · real-time monitoring and explainability for proactive risk and compliance management · traceability and versioning of AI assets.
The feeling of “we are not keeping up” acquired a counter — and it does not only count, it shows where the count comes from. What exactly breaks, how long the repair takes and why, at the same growth in the agent fleet, incidents grow with it for some and stay flat for others.
Speed first. Business units push ahead, governance catches up. Local speed rises, visibility and containment degrade. More than two thirds of CIOs and CTOs say business units bypass IT when adopting AI.
Safety first. Deployment slows under the weight of checks and approvals. Immediate exposure falls, but learning stops, the competitive position weakens, and manual oversight becomes unmanageable — especially when navigating dozens of overlapping AI regulatory frameworks.
The conclusion: “One path trades safety for speed, the other speed for safety. Neither scales, and both accumulate strategic debt”.
The link between incident growth and governance maturityIBM Tech 2026
With the agent fleet growing 2026 → 2027 by 60–87% across regions and 63–83% across industries growth in the number of incidents differs by almost twofold.
In organisations with weak governance more agents mean proportionally more incidents.
In organisations with strong governance the incident rate stays relatively flat even as deployment accelerates.
The source's conclusion: governance investment should match operational risk. In high-risk environments — regulated sectors, sovereignty-sensitive markets, customer operations — underinvesting in governance turns AI growth into operational exposure.
Industries with complex regulated operations — pharma, energy, telecom, retail banking — show the most marked reduction in incidents with mature governance as agents scale.
From the source's checklist: the minimum production standard — an agent or model must be registered, have an owner, be observable and be stoppable. Otherwise it is not deployed.
This report's key derived indicator: the ratio of the growth rate of incidents to the growth rate of agents. Under mature governance it tends to zero; under immature governance, to one.
One floor down, where mixed teams work every day, none of this construction is visible yet. Governance from the desk: who sees no rules, what everyone fears equally at every level, and what changes when a person starts understanding how an agent works.
Rules and accountability: one cut, two questions
Above is the share of employees saying their company has no clear guidance on managing “human × AI” teams; below, the share putting accountability for AI decisions in their top 3 concerns for the next 2–3 years:
The spread on the second row is one percentage point. This is the only question where the hierarchy produces no divergence: accountability worries those who set tasks and those who execute them equally.
22%
of organisations have set clear rules and guardrails for using AI in automated decisions
+18 pp
the gap in having guardrails between companies redesigning processes and those only deploying tools
By contrast, on every other question the gap is 10–40 percentage points.
Only 22% of organisations had set clear rules and guardrails for using AI in automated decision-making. The source's wording: “roughly eight in ten organisations have work ahead to bring their people through a major shift in ways of working”.
The source's comment: “The data shows organisations are investing in AI capabilities but most are underinvesting in the frameworks that protect the humanity of the brand”.
The gap in having guardrails between companies redesigning processes and those only deploying tools:
01
Employees see adequate guardrailsDeploy only 41% · Reshape / invent59% · Δ +18 pp
Governance and redesign go together: those who redesign processes build the rules along the way.
Outside company walls the count is kept by researchers and legislators rather than security teams — and what they count is not only incidents. How many AI laws appeared, where in the world there is more optimism and less, and why “responsible AI” does not reduce to a single number.
Optimism differs twofold by country
Share believing AI products and services do more good than harm. The world average rose from 52% (2022) to 55% (2024).
64%
keep the risk of AI inaccuracy in focus; regulatory compliance — 63%, cybersecurity — 60%
131
AI laws were passed by US states in 2024 alone — against one law in 2016
59
AI regulations — twice as many as in 2023, and by twice as many agencies
+21,3%
growth in mentions of AI in the legislative proceedings of 75 countries over 2024: 1,889 against 1,557; ninefold since 2016
Practically all leading frontier-model developers report on capabilitybenchmarks, but reporting on responsible AI benchmarks remains selective.
A year earlier the AI Index recorded the same problem: standardised responsible-AI evaluations remain rare among large industrial developers, although new benchmarks (HELM Safety, AIR-Bench, FACTS) are starting to fill the gap.
Trade-offs between dimensions of responsible AIAI Index 2026
The key finding of 2026: improving one dimension of responsible AI, say safety, can worsen another, say accuracy.
This means “responsible AI” is not a single scale you can move up but a set of conflicting requirements you have to choose between. The practical implication: any metric of “responsible AI maturity” as a single number misleads.
The gap between acknowledging risk and actingAI Index 2025
A McKinsey survey cited by AI Index 2025: organisations recognise responsible-AI risks, but mitigation lags. The risks of inaccuracy, regulatory compliance and cybersecurity were in focus for only 64%, 63% and 60% of respondents respectively.
Models designed with measures against explicit bias — including GPT-4 and Claude 3 Sonnet — continue to display implicitones: they disproportionately associate negative terms with Black people, more often link women with humanities rather than STEM fields, and prefer men for leadership roles. Bias metrics on standard benchmarks improved, but the problem remains widespread.
The share of restricted tokens in actively maintained domains of the C4 dataset jumped from 5–7% to 20–33% over the year: sites adopted anti-scraping protocols for AI training en masse. The consequences are for data diversity, model alignment and scalability; new approaches to training under constraints may emerge.
US states: from 1 AI law in 2016 to 49 in 2023 and 131 in 2024 alone. US federal agencies introduced 59 AI regulations — twice as many as in 2023, and by twice as many agencies.
75 countries: mentions of AI in legislative proceedings +21,3% (1,889 against 1,557); more than ninefold growth since 2016.
State investment: Canada $2.4bn · China $47.5bn on semiconductors · France €109bn · India $1.25bn · Saudi Project Transcendence $100bn.
National AI safety institutes were launched in 2024: the first in the US and the UK in November 2023, then at the Seoul summit in May 2024 — Japan, France, Germany, Italy, Singapore, South Korea, Australia, Canada, the EU.
AI laws passed by US states per year. The horizontal step is uneven: seven years between 2016 and 2023, one between 2023 and 2024. The unit is a count of laws, not per cent.
2026 (2025 data):
National AI strategies appear fastest in countries that had no formal policy five years ago: more than half of the new 2024 strategies come from developing economies; strategies are in active development in sub-Saharan Africa, Central Asia and the Middle East.
AI sovereignty — the drive for greater self-reliance in national AI capability — is becoming a central principle of national policy, but infrastructure is unevenly distributed: Europe and Central Asia grew state AI supercomputing clusters from 3 to 44 between 2018 and 2025.
Model production meanwhile stays concentrated in the US and China. Open-source development is starting to redistribute participation: the rest of the world's contribution on GitHub already exceeds Europe's and is approaching the US.
Share seeing more benefit than harm in AI products: 52% (2022) → 55% (2024) across 26 countries; it rose in 18 of the 26.
The regional poles: China 83%, Indonesia 80%, Thailand 77% against Canada 40%, the US 39%, the Netherlands 36%.
Growth in optimism since 2022 in previously sceptical countries: Germany +10%, France +10%, Canada +8%, the UK +8%, the US +4%.
Two thirds of people expect a significant effect of AI on everyday life within 3–5 years (+6 pp on 2022); the largest rise in Canada (+17%) and Germany (+15%).
Scepticism about the ethics of AI companies is growing, and trust in AI's fairness is falling.
2026:
The expert–public gap is 50 percentage points: 73% of experts expect a positive effect of AI on how people do their jobs, against 23% of the public. Similar gaps on the effect on the economy and on medicine.
Trust in one's own government as an AI regulator is lowest in the US — 31%.
Globally the EU enjoys more trust as an effective AI regulator than the US or China.
For the corporate agenda this means: internal communication about AI operates in an environment where two thirds of society is sceptical and trust in the institutions meant to regulate the technology is low.
Everything on the section is gathered here, plus BCG's five imperatives for CEOs and Microsoft's four roles in a coordinated rethink of work.
Declared reshape versus its depth
The upper panel is organisations' self-assessment, the lower one the answers of those who work inside them; the denominators differ, which is why the panels are separate. The gap between the panels shows the depth of the redesign.
Change the scoreboard: measure value, not reach. Adoption tells you that people are using AI, not that it pays off. The time individuals save leaks out of the organisation unless it is tracked and deliberately reinvested. Look at business outcomes, not at usage.
Make strategic clarity the top priority and own it personally. Strategic clarity is not a communications task but a leader's stance. Declare AI an explicit top priority, say clearly where the company is going, and make sure it reaches everyone, including frontline employees. CEOs who personally own the transformation outperform the rest on every dimension: value captured, employee enjoyment, trust.
Invest in end-to-end redesign of the work, not in new tools. Most companies still treat AI as a tool for individual productivity, but the more important change is collective: AI changes how teams work together and how tasks flow across the organisation. Capturing that value requires end-to-end redesign of a few core processes.
Put people at the heart of that redesign. Redesign works only if people take part in it. Look ahead at how roles will change over the coming years, train the skills that matter most, and involve people in shaping the change rather than presenting them with the finished thing. What keeps people engaged is understanding how AI helps them grow, not how much faster they work.
Manage this as a moving target, not a one-off programme. The technology moves faster than any company. Treat AI as something you keep steering, not as a programme with a finish line. Put in place a lightweight standing governance loop that re-checks what works, re-measures value and adjusts as models and agents evolve.
Microsoft's four roles in a coordinated rethink of workMicrosoft WTI 2026
Building evaluation infrastructure requires a coordinated rethink of work across four roles:
Employees rebuild their work around setting the task and checking the result.
Leaders redesign processes around outcomes and agents' autonomy.
IT builds the infrastructure for agentic operations at scale — treats agents as managed entities with identities, permissions, policy enforcement and life-cycle management. IT becomes the control plane for agentic operations, extending to agents the same rigour already applied to people and applications, so that scale is not achieved at the cost of visibility.
Security ensures trust is woven into the system itself — accounts for agents' new risks: data exfiltration, unintended system actions, unauthorised access. It requires monitoring, policy enforcement and auditability to be built directly into the platformso that trust works as a structural property of the system.
When these four roles work in concert, the organisation becomes a learning system: work continuously produces understanding, and understanding continuously changes how the work is done.
What is covered in this section
Before moving up the scale it is worth looking at what it is made of: each of the three levels has the source's definition and an example. Then the industries that spread out along that scale, and the 2025 reading that already showed: redesign works and it also frightens.
DeployThe source's definition supporting adoption of generative AI tools and raising productivity · Example rolling out ChatGPT, Microsoft Copilot or Mistral LeChat
02
ReshapeThe source's definition end-to-end redesign of workflows to reimagine functions · Example rebuilding the HR function by embedding AI into critical processes
03
InventThe source's definition building new business models and products for growth · Example creating new revenue streams through AI services
Per BCG's own research, the companies creating the most value with AI direct 80% of investment into reshape and invent — in a few key processes. Not in every process, but in a few.
The gap between the leader and the laggard is 23 percentage points.
The effect of redesign as measured in 2025BCG 2025
Employees at companies redesigning workflows (reshape or invent, n=5,350) against employees at companies only rolling out tools (deploy only, n=1,830):
The source's wording: “Companies redesigning their workflows invest more in transforming people — and it pays off”.
Employees at companies redesigning processes fear losing their job more:
Believe their job will certainly or probably disappear within ten years
The source's comment: this “reinforces the need for clear communication and proper reskilling”.
In groups with deeper redesign both reported benefit and anxiety are higher at once. Ignoring this pair turns a successful transformation into a retention crisis.
By country: the Middle East 63%, Spain 61%, India 48%, Japan 40%, Germany 36%, Italy 36%, the UK 34%, France 34%, the US 33%, South Africa 28%, Brazil 27%.
Share of a country's respondents who believe their job will certainly or probably disappear within ten years. The vertical line is all respondents, 41%, not the average of these eleven countries.
A counter-intuitive finding: the countries with the highest share of regular users — above all India at 92% against an average of 72% — are also among the countries with the greatest fear of losing a job. The source's wording: “Countries with high usage also have the greatest fear of job loss”.
BCG put the same questions to two groups twice — about the result and about what the company does differently; both halves of the reading come with the source's own sample sizes and wording. And two observations that are easy to miss if you look only at the size of the gaps.
Employees at companies on the reshape or invent path (n=8,071) against companies that stopped at deploy (n=2,919); the total base of 10,990 excludes those who do not know how their organisation is adopting AI:
12%
of employees at companies that stopped at rolling out tools take part in process redesign — against 43% where processes are being redesigned
30%
of employees even at leading companies see agents built into their workflows
The source's headline: “Using AI for Reshape or Invent initiatives pays off by delivering more value and providing a better employee experience”.
What exactly these companies do differently (the same base):
01
Sets the direction: employees say the AI strategy is clearDeploy 31% · Reshape / invent52% · Δ +21 pp
02
Invests in skills: employees go through a major reskilling programmeDeploy 47% · Reshape / invent65% · Δ +18 pp
03
Brings people in: employees take part in process redesignDeploy12% · Reshape / invent43% · Δ+31 pp
04
Defines the rules: employees see adequate guardrailsDeploy 41% · Reshape / invent59% · Δ +18 pp
05
Puts agents to work: employees see agents integrated into workflowsDeploy 13% · Reshape / invent30% · Δ +17 pp
06
Measures what matters: employees see that AI value is properly trackedDeploy 27% · Reshape / invent52% · Δ +25 pp
The source's headline: “What makes Reshape or Invent initiatives more successful: a clearer roadmap and deeper investment in people”.
Two observations that are easy to miss
The largest gap is in bringing people into process redesign (+31 pp), not in tools, budget or talent. And the absolute level is striking: at companies that stopped at rolling out tools, only 12% of employees take part in the redesign.
Agents are built into processes even at the leaders in only 30% of cases. So even among those redesigning processes, seven in ten employees do not see agents in their work. It is the same “we see versus we measure” gap, but taken from below.
What exactly BCG called high clarity and broad access — the whole figure rests on this, and that is where the breakdown starts. And it ends with how the same finding arrives from other directions: four independent readings at three sources.
Strong strategic clarity — AI is made a priority, there is a clear AI strategy and guidance on using the time saved.
Broad access to tools — there is access to AI tools or to testing and experimenting with them.
Base: 9,923 regular AI users.
Result
Share of respondents reporting measurable impact (an improvement in key business metrics attributed to AI):
What this means
Giving people tools without direction adds 5 percentage points. Giving direction without tools adds 25. A fivefold difference.
This is the strongest argument against the tool-first approach; it is derived from the answers of 9,923 regular AI users.
In addition: moving from “high clarity + limited access” to “high clarity + broad access” adds only 3 pp. So where clarity exists, tools almost stop being the constraint.
IBM: 83% of CEOs say AI's success depends more on people's adoption than on the technology itself. In 2025 the same thing was said by 71% of CMOs.
Microsoft: in Microsoft's model organisational factors account for 67% of the explained difference in self-reported impact, individual ones for 32%.
BCG, a separate reading: clarity of strategy is the largest gap between companies at the “reshape/invent” level and the “deploy” level (+31 pp).
Four independent readings at three sources point the same way.
The check in full: what the composite impact measure is built from, by what method the factors were ranked and how the explained share was split between the organisation, the person and demographics. And the source's caveat, without which any chart on this data claims more than the data allows.
29 factors (10 organisational, 9 individual, 10 demographic) tested against a 10-item composite measure of AI impact (question Q27).
The AI impact composite includes self-reports that AI: makes the person more creative · enables new kinds of work · gives better first drafts · improves the ability to work · improves collaboration · gives a sense of control · improves career prospects · enables high-value work · raises the likelihood of staying with the company because of AI.
The method: permutation feature importance, random forest, normalised so that the strongest factor = 100%. The ranking is stable across three model families: test R² 0,680 (elastic net), 0,689 (random forest), 0,690 (XGBoost).
The sample: 20,000 surveyed, analytical sample 19 854 after excluding incomplete observations across the 29 factors. 10 markets: the US, Brazil, Australia, India, Japan, France, Germany, Italy, the Netherlands, the UK. Fielded 18 February – 20 April 2026, field partner Edelman Data x Intelligence.
after excluding incomplete observations across the 29 factors
100%
· talent practices (43%) · manager support (43%) ·
Share of the effect explained
01
Demographicsremainder
The composition of the groups
The organisational environment (in order of strength): the organisation's AI culture (100%, the strongest factor) · talent practices (43%) · manager support (43%) · organisational concerns about AI · team AI practices · AI in performance reviews · leadership alignment · governance maturity · organisational adaptability · named barriers.
Individual attitude and behaviour: AI attitude (42%, the strongest individual one) · intrinsic motivation · the work's readiness for AI · sophistication of AI use · external pressure · shift in strategic focus · proactive AI behaviour · progress in skills · concerns about AI.
Demographics: familiarity with AI · market · decision-maker status · job level · generation · function · tenure at the company · company size · industry.
Three conclusions
The three strongest factors are all organisational. The strongest single factor, the organisation's AI culture, is roughly 2.5 times stronger than the strongest individual one (42%).
Demographics explain almost nothing. Generation, industry, company size, job level are at the tail of the list. This refutes the common explanations via “the young pick it up faster” or “it is easier at technology companies”.
The source's conclusion verbatim: “The real question isn't whether people have the right skills. It's whether the organization is built to unlock them”.
The obligatory caveat
The source states directly: the values show a statistical association, not a causal effect. The factors are self-reported assessments of the work, the workplace and one's own use of AI. Any chart on this data without that footnote claims more than the data allows.
Microsoft People Science Agentic Teaming & Trust Survey (July 2025), 1,800 workers globally: 819 leaders (senior management, VPs, directors), 520 managers, 461 individual contributors — all outside Microsoft.
When managers actively model the use of AI, employees report:
+17 points of perceived AI value
+22 points of critical thinking about their own use of AI
+30 points of trust in agentic AI
When managers create psychological safety for experiments, employees report up to +20 points of readiness and value and are 1.4 times more likely to become frequent users of agentic AI.
A map whose two axes both have content: what individual capability is made of, what organisational readiness is made of, and how people fell across the five zones. And alongside it, what the map does not show — leaders and employees answer the same questions differently, and in the leading group the manager behaves differently.
Individual capability reflects how broadly the respondent uses AI and how confidently they steer it, judge its outputs and learn from them. It includes how actively they experiment, share knowledge and create new value with AI — from improving the quality of work and processes to doing work that was previously impossible.
Organisational readiness reflects the environment around them: the culture and management practices supporting AI use; clear rules and guidance on people and AI working together; whether AI use is encouraged and recognised.
Both axes are self-reported composite scores, normalised 0–1 within each market before pooling. Individual readiness is built from questions Q1 (self-efficacy), Q4 (sophistication of use), Q6 (proactive behaviour), Q22 (value creation). Organisational readiness from Q11 (governance maturity), Q12 (manager support), Q14 (AI in performance reviews), Q19 (organisational AI culture).
Analytical sample: 16,971 of 20,000 with complete data on both axes. Neither telemetry nor behavioural observation is used.
The five zones
31%
of AI users are in a mismatch
26%
of AI users say leadership is clearly and consistently aligned on AI
01
Leaders' zoneHigh individual capability and high organisational readiness reinforce each other
02
FormingBoth individual practice and organisational conditions are still forming
03
StalledBoth are low
04
Hands tiedIndividual AI practice is high, organisational conditions are low
05
Untapped potentialOrganisational conditions are high, individual practice is low
31% of AI users are in a mismatch between themselves and the organisation (stalled, hands tied and untapped potential; in the source's wording, 31% mismatched, with the other 50% in the forming zone).
Three figures that explain why people do not move:
01
Fear falling behind if they do not adapt quickly with AI65%
02
Believe it is safer to focus on current goals than to redesign work with AI45%
03
Say they are rewarded for rethinking work with AI even if the results are not achieved13%
The source's wording: “Employees are ready to reinvent how they work, but the system around them — metrics, incentives and norms — keeps reinforcing the old way. The same forces that accelerate AI adoption also hold it back”.
“The transformation paradox is at its core a systems problem. And systems do not fix themselves: they have to be redesigned”.
In 2026 two sources carry this line, and both versions are here in full: at Microsoft, mastery instead of adoption and the signals agents already produce while organisations do not collect them; at IBM, the same conclusion in words about the seams between functions. And three questions a pioneering company must answer.
What CEOs say about their people and processes
The upper panel is about readiness, and next to it stands the same CEOs' estimate: a quarter of the workforce uses AI regularly.
The source's wording: the firms pulling ahead focus on mastering AI, not merely adopting it — they redesign how the work is done and turn the result of work into understanding. When that understanding is captured, spread and built into how the organisation works, a self-sustaining learning system.
emerges. The key observation: many executives focus on hiring the right people and assume results will follow. The data shows otherwise: it is about.
the conditions leaders create for that talent to be unlockedMicrosoft WTI 2026
The signals agents generate As agents take on more, theygenerate valuable signals : what worked, what failed, where outcomes drifted. In most organisations surveyed those signals stay local or spread slowly. Pioneering companies treat them differently:they capture and codify them into shared procedures
, improving future work while preserving accountability and control.Microsoft WTI 2026
01
What employees in the leading group doThe rest 32% · The leading group63% · Δ +31 pp
02
Teams jointly work through and refine business processes to find opportunities for AIThe rest 36% · The leading group61% · Δ +25 pp
03
Share AI prompts, new agents, findings and mistakesThe rest 29% · The leading group54% · Δ +25 pp
04
Discuss quality standards for AI work Agent workflows, handoffs to humans and quality standards are documented and reproducible at the level of theThe rest 19% · The leading group26% · Δ +7 pp
05
team The same at the level of theThe rest 17% · The leading group29% · Δ +12 pp
06
team functionThe rest 14% · The leading group25% · Δ +11 pp
organisation Note the absolute levels of the last three rows: even among the most advanced users, documented and reproducible procedures for working with agents exist for onlya quarter
. This is the least mature practice of all those measured.Microsoft WTI 2026
Evaluation infrastructure The source's wording: the more agents execute, the higher the stakes around human evaluation. Approving one bad output is manageable; when bad outputs pass at scale, risk accumulates. The key is to.
build evaluation infrastructure that keeps up with the agents
Three questions every pioneering company must answer:
Who checks the agents' work?
Who has the authority to update the workflows the agents execute?
How is a local win captured and scaled across the whole organisation? Organisations able to answer these questions build proprietary intelligence
— institutional know-how that accumulates over time, is unique to the firm and hard to reproduce.IBM CEO 2026
IBM's view of the same thing:
Move #4, “Align intelligence — human and artificial” Today AI augments people; by 2030 people will augment AI.
The main shift is not structural but cultural. CEOs who actively redesign how cross-functional teams work together are more than twice as likely
to meet their business goals. Organisations that redesigned five key areas — technology, finance, HR, operations, cross-functional collaboration — are 4 times more likely
to meet their business goals.
The wording: “When functions evolve independently, performance improves incrementally. When they are redesigned as an integrated system, the improvements reinforce each other”. The CEO should focus on the seams
87% CEO between functions — handoffs, decision rights, collaboration flows.
actively build AI into end-to-end processes. The recommendation:redesign workflows before redesigning roles
. “Do not fund reskilling until there is a redesigned way of working in which those skills can be applied”. Reskill for connecting into a system, not for replacement:
prioritise skills in systems thinking, interpreting AI outputs, challenging recommendations and managing exceptions. Redefine performance reviews and manager training to reward those who fuse human and machine intelligence into better results, not merely faster ones. Turn AI use into a key operational metric:
measure where AI is available, where it is actually used and where work reverts to manual defaults. Move #5, “Expect unpredictable futures”
— the same logic on a longer horizon: becoming AI-benchmarked is not the final goal but the foundation for what comes next. The bet is on the ability to adapt, not on a particular technology.IBM CEO 2026
Employees' readiness for the shift
IBM data that is rarely quoted: In every generational group at least twice as many
61% employees would accept rather than resist wider use of AI by their employer.
48% say AI makes their work less routine and more strategic.
And yet say they would be comfortable being managed by an AI agent.CEOs estimate that only 25% of workers use AI regularly in their work 86% CEO — while
say employees have the skills to work with AI. The source's conclusion verbatim:“The gap between capability and deployment is an organizational design problem more than a skills problem”
. And further: “If AI is not being used, treat that as an operational failure, not a qualification problem”.
Everything on the section is gathered here, plus the role-by-role actions from the CAIO Study Action Guide.
Leadership is aligned on AI
The upper rows are measured among executives, the lower ones among doers. There is no point measuring clarity of strategy among the people who formulate it. IBM CAIO 2025)
What the CEO should do (Action Guide, Give the CAIO authority to lead.
Grant a clear mandate, reporting line and visibility in the C-suite. Make the CAIO accountable for ethical and safe AI in collaboration with security and ethics leaders. Require measurable business outcomes.
Build AI dashboards to track business impact, risks and ethical consequences. Enable the CAIO to explore and assess new, less quantifiable opportunities — and creative ways of measuring their effect. Give the CAIO resources.
Allocate a dedicated AI budget and grant ownership tied to ROI. Assemble the CAIO's team with the right mix of technical and business expertise. Choose strategic partners that extend capabilities and align with core AI goals. Build governance, ethics and data-sharing standards into every partnership as non-negotiable conditions.
Inspire employees with a vision for AI. Communicate clearly how the AI strategy will help meet business goals. Encourage questions and be transparent in the answers. Make it so that people are ready to talk about problems and opportunities even without being AI experts.
Encourage a growth mindset. Give employees the chance to experiment with AI through internal competitions and hands-on programmes. Celebrate progress, not only success.
What the CAIO should do (Action Guide, IBM CAIO 2025)
Secure clarity of the role. Actively agree the mandate, responsibilities and objectives with the CEO. Delineate the role from other C-suite members, avoiding unnecessary overlap.
Create and measure clearly defined indicators. Build an AI dashboard to track business impact, risks and ethical consequences. Use analytics to identify the levers that move key metrics. Track different ways AI value is realised, not only quantitative ROI.
Engage with C-suite peers, even when they disagree. Strengthen the network in the C-suite. Understand whom to approach, when, why and how. Engage those who may be sceptical, especially on topics requiring input from several teams — AI ethics and governance. Learn each person's agenda and speak their language.
Scale the team's impact. Mix business, industry and technical skills. Connect the team with key partners in IT, corporate strategy, operations and compliance.
Lead the shaping of the AI operating model. Steer towards a centralised operating model. Set the rules, put frameworks and processes in place. Act as the conductor of the AI ecosystem.
Develop a roadmap for AI-driven digital transformation. Identify the areas where AI creates value, assess readiness, draw up an adoption plan. Work with HR on cultural transformation: promoting data literacy, encouraging experiments, building a culture of innovation.
What is covered in this section
A line in the org chart is not yet a role: it has an origin, a budget, a voice and an operating model beneath it. From here you can see where CAIOs come from, what they control and why results diverge at the same title.
In 2025 IBM surveyed more than 2,300 organisations to find 600+ CAIOs. 66% of those CAIOs expected the role to appear at most organisations within two years — the forecast came true in one.
66%
of those CAIOs expected the role to appear at most organisations within two years
100%
of CEOs expect the CAIO's influence to grow by 2030
57%
of CAIOs were appointed from the organisation's internal bench
36%
higher ROI
100% of CEOs expect the CAIO's influence to grow by 2030.
57% of CAIOs were appointed from the organisation's internal bench.
Professional background — the areas CAIOs focused on over their careers:
The source's wording: CAIOs come with a data-first skill set, but most also have a strong background in business strategy, innovation, enterprise technology and operations. As AI becomes more transformative it also becomes more targeted — use cases focus on industry-specific end-to-end processes, so the CAIO must have deep industry expertise or access to it through the team.
on ROI for AI spend at organisations with a chief AI officer
+24%
on the likelihood of outperforming competitors on innovation
Different quantities and different denominators: a premium on return and an increase in likelihood. They cannot go on one axis.
The operating model matters more than the role itselfIBM CAIO 2025
CAIOs running hub-and-spoke or centralised operating models, against those running decentralised ones:
move at least pilots into production
get 36% higher ROI on AI investment
The source's wording: “With the right operating model organisations can build on the base 10% ROI premium from having a CAIO”.
The distribution of models across maturity stages:
Operating model
As they scale, organisations move from decentralised models to centralised and hub-and-spoke ones. The wording: hub-and-spoke lets the CAIO direct resources to priority initiatives more effectively, and the resulting focus and flexibility give better results.
How CAIOs' priorities change with scaleIBM CAIO 2025
The ranking of the CAIO's most important responsibilities changes as they move from pilot to scale:
01
1At the pilot stage Define the organisation's AI strategy · At the scale stage Raise AI acceptance in the C-suite
02
2At the pilot stage Manage partnerships in the AI ecosystem · At the scale stage Steer AI adoption
03
3At the pilot stage Oversee upskilling of existing AI talent · At the scale stage Manage the organisation's AI budget
04
4At the pilot stage Build the business case for AI · At the scale stage Develop a change management strategy
05
…At the pilot stage … · At the scale stage …
06
lastAt the pilot stage Manage the organisation's AI budget · At the scale stage Manage partnerships in the AI ecosystem
The source's wording: “As AI scales, CAIOs focus less on strategy and more on adoption”.
The wording: this “shows where stronger collaboration and a clear delineation of roles will be needed”.
The three areas successful CAIOs focus onIBM CAIO 2025
The study identifies three areas where CAIOs with greater measurable business impact concentrate their attention:
Measurement. Success cannot be defined in the abstract. Senior leadership must specify how impact will be measured and what business outcomes it expects. The indicators must go beyond project ROI and include metrics of more transformational impact: revenue, profit, customer satisfaction, employee productivity. A dashboard with the right indicators, visible to all relevant decision-makers, is the central tool.
Teamwork. The CAIO must not be an army of one. The average size of a CAIO's team is five people, regardless of AI maturity, and smaller teams are less successful. Teams that prioritise AI specialists, machine-learning engineers and business strategists deliver greater measurable impact. It matters to build a team that complements rather than duplicates the existing technology staff: “If you create a shadow IT department focused only on AI, integration becomes even harder”.
Authority. The CAIO needs a clear mandate grounded in visible support from other senior leaders. But not every form of involvement is equal.
Only 25% of executives fully agreethat their organisation's IT infrastructure can support scaling AI across the enterprise. No CAIO mandate compensates for that.
The role has no hands of its own: everything the chief AI officer does is done by other functions — technology, operations, people. From here you can see whom they negotiate with, whose contribution to the result turned out decisive, and why the person named the main opponent is the one responsible for reskilling.
The top three rows are the role's weight. The bottom one is whom it conflicts with: it is the chief HR officer who is responsible for reskilling, which has not moved in two years.
01
CEO — the supporting sponsorThe function in the AI agenda Defines the CAIO's mandate, gives the authority to break obstacles. Co-creates and loudly backs the AI strategy · The key detail 57% of CAIOs report to the CEO or the board. The CAIO should also work closely with the CFO and CSO on ways of measuring value
02
COO / CSCO — scaling the transformationThe function in the AI agenda The COO must be both an advocate and an active partner in integrating AI into strategy and the operating model · The key detail Without the COO's support the CAIO cannot deliver productivity gains and process innovation
03
CDO — the data engineThe function in the AI agenda Joint work on data strategy, quality, governance and analytics · The key detail Their collaboration is critical for converting corporate data into value through AI
04
CIO / CTO — the technology integratorsThe function in the AI agenda The CTO and CAIO together build the adoption roadmap on the basis of technical feasibility assessment; the CIO ensures enterprise IT readiness · The key detail Aligning the AI, IT and technology strategies is critical
05
CISO — managing security and riskThe function in the AI agenda Partners in building a culture of embedded security, where protection is designed into AI from the start · The key detailMore than a quarter of AI initiatives were cancelled, postponed or failed to scale because of security problems
06
CINO / CDIO — the innovation catalystsThe function in the AI agenda Designing AI-centric solutions, evolving products and services, improving customer experience · The key detail The CDIO and CAIO together execute the AI-driven digital transformation
07
CHRO — the change agentThe function in the AI agenda Aligning talent strategies, defining the skills required, training programmes · The key detail32% of CAIOs name the CHRO as one of the main opponents of AI
The source's wording: “Support from the CEO is critical, but the CTO's involvement and advocacy are the key factors in AI success”.
Three observations:
The CTO is present in all three dimensions — the only role with that position. It is their teams that design, build and deploy AI solutions.
The CHRO is first on advocacy. That same CHRO whom 32% of CAIOs name an opponent. The source comments: if these leaders are on board, they can become the CAIO's most effective advocates, showing employees how AI is good for their careers. This underlines the need for more substantive engagement.
The CEO is only third on support and absent from the other two dimensions. Sponsorship from the top is necessary but not sufficient; execution is delivered by technology and operations roles.
Through active engagement the CAIO aligns the AI strategy with the business, technology, innovation, security and people strategies, focusing the enterprise's efforts on a shared set of AI-driven outcomes. Moreover, they can take on the difficult questions of AI ethics and governance — even when these are not formally considered their direct responsibility.
If the boundaries between functions are obsolete, what decides next is not a job title but who takes the decision. CEOs answer that question themselves — all eight statements about their own design against four in the article — and alongside it stands the boundary of the mandate IBM proposes drawing around the chief AI officer.
Functional boundaries declared obsoleteIBM CEO 2026
+Share of CEOs9
01
The boundaries between business and technology are obsolete77%
02
Talent and technology roles are converging77%
03
All functional leaders must become technology experts in their area85%
04
The CHRO's influence will grow in the coming years59%
05
Are decentralising decision-making79%
06
AI is already changing aspects of the business they consider core69%
07
AI's success depends more on people's adoption than on the technology itself83%
08
Employees have the skills to work with AI86%
09
Estimate of the share of workers using AI regularly25%
What CEOs say about their own design
Four statements from this list are drawn in the article, four more are in D-12.5; here are all eight at once, on one denominator — the share of CEOs: six of the eight are squeezed into the 77–86% band, and markedly lower than all of them is growth in the CHRO's influence (59%). The list's ninth number, 25%, does not go on this axis — it is CEOs' estimate of a share of workers, not a share of CEOs themselves; it is examined in D-9.4 alongside the highlighted 86%.
The expected change in roles' influenceIBM CEO 2026
CEOs expect the influence of every C-suite member to grow between today and 2030. At the same time, in the functions where AI transformation started earlier and has already touched many teams and tools — marketing and HR— the expected increase is lowerthan in more technology-benchmarked areas where AI integration is still unfolding.
The CAIO's mandate: where the boundary runsIBM CEO 2026
IBM's recommendation, easy to miss:
Give AI leadership authority — within limits. If you have a CAIO, clarify their mandate. If you do not — appoint one now. Give them authority over AI priorities, standards and funding gates — but not ownership of business results. Their job is to accelerate decisions, scale what works and stop what does not. Line leaders remain accountable for the results. It is precisely that separation that allows speed without chaos.
The other recommendations on rebuilding the C-suiteIBM CEO 2026
Redesign decision rights before touching the org chart. Find the handful of corporate decisions that hold everything else up — AI deployment, pricing moves, capital reallocation, choice of partners — and assign each a single owner, explicit authority and clear escalation rules. 79% of CEOs are decentralising decision-making, but first you have to help leaders understand who decides what, otherwise there will be no acceleration.
Make every executive accountable for corporate results. Tie a meaningful share of senior leadership's compensation — at least 30% — to shared results: growth, margin, customer trust, rather than to functional metrics. Break down the walls between units, especially between HR and IT. Incentivise joint ownership of workforce redesign, reskilling and AI deployment, so that people and technology decisions are made, measured and delivered together.
Reward leaders who enable safe AI connection across the ecosystem. Ask executives to reassess partners by how easily third-party AI agents can discover, integrate with and transact through your offerings.
What it says about the C-suite of 2030IBM CEO 2026
the C-suite of 2030 must become more AI-native, more technology-centric, more operationally integrated and more inclined to work through ecosystems. “The real difference will not be in job titles on the org chart, but in how these leaders work together — and how ready they are to challenge each other. Speed is the result of productive friction, not of its absence”.
“In a rewired C-suite every leader is obliged to own outcomes, not merely manage tasks. They are accountable for spotting opportunities, making bets and driving change — regardless of whether it falls neatly within their functional domain. The CEO's role is to make that accountability explicit: who owns which outcome, who can decide without consensus and which risks each leader is obliged to take”.
Ownership of the AI agenda: KPMG's seriesKPMG Q2’26
Who leads the AI agenda
01
Q2 2025CIOs continue to lead enterprise-level AI strategies, in the view of 87%
02
Q2 2026distributed: a named senior executive (CIO, CAIO, COO and others) 34% · the CEO or the executive committee 32% · the business-unit head 14% · a centralised AI governance and risk committee 10%
67% agree or strongly agree that their organisation's CEO actively owns AI as a strategic business priority with clear accountability for AI outcomes across the organisation.
KPMG's wording: “AI leadership is anchored at the top but executed by a broader group of leaders, reflecting AI's cross-enterprise nature and its effect on how decisions are made and value created”.
Topics boards work through quarterly in connection with AI (Q4 2025):
For context, in Q2 2025, when demonstrating ROI to investors, profitability and having established accountability and governance policies in place were named the most important factors — 55% each. So governance as an argument before an investor was already equal to profitability by mid-2025.
2025: of CEOs considered generative AI too risky for core business functions
64%
2026: of CEOs are comfortable taking major strategic decisions on the basis of AI outputs
Two different questions, not one series: these percentages are not comparable with each other. What changed is sentiment, not a measured quantity.
01
2025generative AI is too risky to use in core business functions
02
2026are comfortable taking major strategic decisions on the basis of AI outputs
Over a year the position on one and the same subject reversed. The source's caveat: this does not mean abandoning boundaries — regulatory reporting, material disclosures and sensitive legal judgements stay outside the zone of AI decisions.
The same question put to both sides appears several times in the article. There are thirteen such paired readings in all — and in exactly one of them leaders and employees answer identically.
Both panels share a 0–100 scale, so “a point further right” means the same thing from row to row.
86%
say employees have the skills to work with AI
25%
use AI regularly in their work — by the same CEOs' estimate
Where employees' figure is higher
The thirteenth reading is not put on the canvas: accountability for AI decisions is placed in the top 3 concerns by 46% of leaders and 46% of employees — the only question with no gap at all.
+Δ13
01
Rethinking work with AI is rewarded regardless of the result×2
02
It is safe to propose new ways of working with AI+14 pp
03
Managers create space for experiments+19 pp
04
No clear guidance on managing “human × AI” teams+13 pp
05
Limited or no guidance on the time saved+14 pp
06
Do not convert time into strategic work+22 pp
07
Familiar or very familiar with agents (2025)+27 pp
08
Believe AI will accelerate their career (2025)+12 pp
09
Limited understanding of what agents are (2026)+22 pp
10
Save at least a day a week+18 pp
11
Rise in job satisfaction+20 pp
12
Rise in cognitive load+11 pp
13
Accountability for AI decisions in the top 3 concerns0 pp
Question
01
Leadership is clearly and consistently aligned on AIEmployees26%
02
Leadership communicates clearly about AI (frontline employees' assessment)Employees33%
03
Leaders' words match the organisation's actions (frontline employees' assessment)Employees28%
04
Sufficient leadership support (frontline employees' assessment, 2025)Employees25%
A gap is present on every question but one. The only topic on which the hierarchy does not diverge is concern about accountability for AI decisions.
86% CEO say employees have the skills to work with AI. The same CEOs estimate that only 25% of workers use AI regularly in their work.
say employees have the skills to work with AI. The source's conclusion verbatim:. And the recommendation: “Measure where AI is available, where it is actually used and where work reverts to manual defaults. When CEOs estimate that only 25% of the workforce uses AI regularly, they should require their leaders to explain the gaps, remove friction from workflows and share accountability for closing them. If AI is not being used, treat that as an operational failure, not a qualification problem”.
Microsoft's 2025 explanation: “We expect this is because leaders are the first to feel the pressure to have an AI strategy — and the first to be held accountable for making it work. They see what is coming and know they cannot afford to wait. Managing agents also plays to their strengths: delegating, directing and stepping in when needed”.
A Microsoft researcher quoted in the same place: “Working with agents is like onboarding a new team member — you do not micromanage, but you need informed trust”.
An important detail of the trend change. In 2024 the AI wave was led by employees. In 2025, in Microsoft's wording, the picture inverted: on all seven indicators of “agent-manager thinking” leaders are ahead of employees.
The practical implication
Clarity of strategy has to be measured from below, not from above. Any reading taken at the senior management level systematically overstates the result by 15–40 percentage points.
Support that reached a specific person is only one of BCG's three keys to adoption; the other two, training and access to tools, did not fit in the article. And Microsoft's counter-reading: what changes when a manager uses AI in plain sight.
1. Proper training. Only 36% feel properly trained. Effectiveness depends on volume, format and the presence of a coach:
Share of regular users
These three components also significantly raise employees' confidence in AI and improve the quality of results from working with AI.
2. Access to the right tools. Almost four in ten employees (37%) say the company does not provide them with the tools they need. When corporate solutions fall short, 54% say they would use unauthorised AI tools — with the corresponding security risks. Among Gen Z and millennials (n=7,039) — 62%, among the rest (n=3,596) — 43%, a gap of +19 pp.
BCG 2026 adds what was not there in 2025: “CEOs who personally own the transformation outperform the rest on every dimension: value captured, employee enjoyment, trust”. Strategic clarity is called “not a communications task but a leader's stance”.
Everything on the section is gathered here, plus PwC's five next steps for executives.
Roles that did not exist five years ago
Share of employers considering such a role. The role names are left in Latin script — that is how they appear in job postings. Over two years 1.3m AI-related openings were created.
Use AI for growth, not only for efficiency. The companies getting the most value do not apply AI only to cut costs or headcount. They open new revenue, enter new markets, create new forms of value. Build the AI agenda around growth opportunities, especially through partnerships across sector lines. But aiming is not enough on its own: PwC's AI performance study shows that the right foundations — a targeted strategy for scaling the most valuable initiatives and modernised data and technology platforms — strongly raise the ability to deliver on AI's promise.
Shape workforce strategy through the lens of whether expertise requirements are rising or falling. Roles where expertise requirements rise may need deeper investment in advanced AI literacy, in skills where human qualities matter especially, in specialised talent and in retention strategies as demand and wage pressure grow. Roles where the entry bar falls need proactive work redesign, mobility paths and refreshed career propositions.
Invest in AI agents as the main complement to human expertise. PwC's AI performance study: the companies extracting the most value from AI investment are twice as likely to deploy agentsas those with lower ROI.
Reinvent early-career paths. Map the junior roles increasingly demanding experienced-specialist skills. Redesign onboarding, mentoring and training programmes for accelerated development of advanced skills — leadership, stakeholder management, strategic decision-making.
Invest in skills where human qualities matter especially, alongside AI skills. Empathy, judgement, creativity and leadership become more competitively valuable as AI takes on routine and technical tasks. Building a workforce with these abilities is strategically as important as developing AI competencies.
Use AI to transform the whole enterprise, not for isolated use cases.
Treat AI as a growth strategy, not only an efficiency one.
Make AI agents a priority as a tool that multiplies employees' capabilities.
Give the workforce skills: map the skills you have, the skills required and the ways to close the gap — “Buy, Build, or Bot”.
Show people how AI benefits them and unlock the transformational potential by building trust — responsible deployment, clear governance rules, societal and organisational trust.
Comparing the two lists shows the shift: in 2025 the emphasis was on transformation and speed, in 2026 on differentiated work with the two workforce tracks and on entry to a profession.
What is covered in this section
The four conclusions with which PwC closed four widespread fears are the short version of its table: in the full one, what the data actually showed is written against each fear. Along with it are the definitions everything rests on, the quartile series, and the caveats PwC places next to its own conclusions.
AI-exposed occupations — those containing many tasks where AI can be applied. Examples: financial analysts, data entry workers. The term “AI-powered” is used as an equivalent.
Augmentable occupations — AI-exposed ones where AI strengthens or supports human judgement and expertise in many tasks. Examples: surgeons, judges.
Automatable occupations — AI-exposed ones where AI can perform many tasks. Examples: programmers, customer support workers.
AI-exposed industries — those containing many roles where AI applies. Software publishing is AI-exposed, logging is not.
Methods: AI exposure per Felten et al. (AI Occupational Exposure); augmentability and automatability per the IMF's method (“Gen-AI: Artificial Intelligence and the Future of Work”, 2024).
The source's key caveat: “We analyse AI exposure (the ability to use AI) as a proxy for AI adoption, although actual levels of use may lag exposure. As use grows, the effects we find may strengthen”.
ProductivityThe common view AI has not yet had a significant effect on productivity · What the data shows The industries most able to use AI achieve three times greater productivity growth, measured as growth in revenue per employee
02
WagesThe common view AI may affect wages and bargaining power negatively · What the data shows Wages are growing twice as fast in the industries most exposed to AI
03
Number of jobsThe common view AI may lead to a fall in the number of jobs · What the data shows The number of openings is growing across all AI-exposed occupations, though more slowly than in less AI-exposed ones
04
InequalityThe common view AI may worsen inequality in opportunity and wages · What the data shows Wages and employment are growing in both augmentable and automatable occupations. Demand for formal degrees is falling faster in AI-exposed roles, opening opportunities for millions
05
SkillsThe common view AI may “deskill” automatable occupations · What the data shows AI may enrich automatable occupations, demanding more complex skills and decision-making
06
AutomationThe common view AI devalues automatable occupations · What the data shows Wages are growing in both. AI may raise the skill level of automatable roles even faster than of augmented ones
Productivity (growth in revenue per employee, 2018–2024, by quartile of the industry's AI exposure):
The trajectory within the quartile AI affects most: 7,3% (2022) → 15,0% (2023) → 27,0% (2024). For the least AI-exposed over the same period: 9,9% → 9,0% → 8,5%.
Cumulative growth against the 2018 base, not annual growth. In 2022 the top quartile was still behind: 7.3% against the bottom quartile's 9.9% — the lines cross between 2022 and 2023.
The source's wording: “In two years the industries most able to use AI went from productivity laggards to leaders”. The caveat: “We cannot prove causality with confidence, but we know that revenue growth in AI-exposed industries accelerated sharply in 2022 — the year the launch of ChatGPT 3.5 opened AI's possibilities to the world”.
Wages (growth per employee, 2018–2024): 7,9% / 8,2% / 12,6% / 16,7% by quartile.
The AI skills premium: 56% on average; every industry analysed pays a premium. The source's caveat: higher wages may also reflect a shortage of such people, “although a shortage does not on its own translate into high wages — if it did, experts in rare specialisms like medieval calligraphy would be very well paid”.
Number of jobs: growing in practically every type of AI-exposed occupation, with two exceptions globally — keyboard clerks and and ICT specialists. But the growth is slower: +38% over five years in the more AI-exposed against +65% in the less AI-exposed.
Degrees: demand for formal degrees is falling for all occupations, but especially fast for AI-exposed ones. The source's possible explanations: AI helps people acquire and apply expert knowledge quickly (“democratisation of expertise”), which makes formal qualifications less relevant · fast skill turnover makes degrees obsolete faster · strong demand for people with AI skills forces employers to look beyond the narrow circle of the formally trained.
Speed of skill change: +66% in the most AI-exposed against the least — more than twice as fastthan a year earlier (25%).
A clarification by role type: automatable occupations experience the greatest skill disruption — an average net skill change of 3.9 against 3.0 for augmented ones. Since automatable occupations simultaneously show growth in wages and job numbers, the source concludes: they are being rebuilt to create more value, probably shifting towards more complex or creative tasks.
Gender: in every country analysed there are more women than men in AI-exposed occupations. So women have both more opportunity and more risk. At the same time PwC's Workforce Radar 2024 study shows that women's AI usage rate in the US lags significantly behind men's.
Demographics as context. More than a quarter of the world's population lives in countries with a shrinking working-age population; by the 2050s it will be more than half. The source poses the question: might slower job growth in AI-exposed occupations turn out to be useful for countries with ageing populations and help avoid “a looming economic crisis as the older population grows and the pool of workers shrinks”.
Two characters through whom PwC explains the mechanics:
Amina, an information analyst. AI augments her analytical abilities: she delegates research and report drafts to agents and spends more time herself on interpretation, refinement, client service and developing new business. She has had to learn new skills — “she has not learnt so much so fast since school” — but she enjoys the work more, and she brings the company more revenue.
John, a customer support specialist. AI automates many parts of his role, including simple queries. John feared AI would replace him, but in fact AI expanded what he can do: he now helps with more complex queries — working through non-trivial product problems, de-escalating tense situations with empathy. His value grew from “handler of questions” to “solver of complex problems”.
The source's conclusion from the second example: the distinction between augmentation and automation is useful, but automation can serve to augment what a worker is able to do.
A large technology company deployed an omnichannel contact centre on agents with predictive intent modelling, adaptive dialogue and real-time analytics: time on the phone fell by almost 25%, call transfers by up to 60%, customer satisfaction rose by about 10%.
A large hospitality company automated updates, approvals and brand-standard compliance tracking through agentic workflows: review time fell by up to 94%.
A global medical company deployed agentic workflows in oncology: automating the extraction, standardisation and querying of unstructured documents delivered a 50% improvement in access to actionable clinical conclusions and almost a 30% reduction in administrative burden on staff.
Southwest Airlines with PwC modernised the system for tracking crew attendance and leave. Generative AI capabilities were used to extract requirements directly from the source code, which halved project planning time. A vice president's quote: “Rather than replacing employees, these improvements freed up valuable time, letting teams think critically, solve complex problems and drive innovation”.
“Thinking small” — using AI narrowly and retrospectively, doing tasks or creating products the same way as before. The analogy: if electricity had been used only to replace candles with “electric candles” rather than to create computing, telecommunications, satellites, aviation. Erik Brynjolfsson calls this “the Turing trap”: a focus on building AI that imitates a human.
“Thinking big” — approaching AI as a transformative tool that unlocks new capabilities, products and whole industries. For context: two thirds of US jobs today did not exist in 1940, and many of them appeared thanks to technological progress (citing David Autor's work).
The source's conclusion: the data shows companies are so far thinking big — using AI not only to cut headcount but to help workers even in the most automatable roles create more value.
Splitting the market into two tracks is not enough — you also have to be able to say where a specific role lands and what that means for it. PwC has a method for this, three series that diverge between the tracks, and two cases where a role went somewhere other than its category prescribed.
The method of Teeselink and Carey (2026). The expertise data is provided at the level of the SOC-2018 classification. Since the SOC classification is unavailable in Lightcast data outside the US, to obtain global indicators the expertise scores are mapped from SOC-2018 to ISCO-08 — for the purposes of this analysis only (PwC 2026, notes).
The definition of expertise: “requiring specialised knowledge or ability”.
Three categories
Share of openings
8%
wage growth for childcare managers — while the number of openings doubled
111%
growth in childcare manager openings since 2019: a role with a falling entry bar and enormous latent demand
01
AI barely touches itMechanics AI contains few tasks where it applies; limited effect · The source's examples cooks, construction workers, mechanics
02
Expertise requirements riseMechanics AI takes the relatively simple tasks, the complex and expert ones remain for people · The source's examples radiologists, recruiters, air traffic controllers
03
The entry bar fallsMechanics AI takes the relatively expert tasks, the less demanding ones remain for people · The source's examples software developers, loan officers, financial managers
Illustrations of the mechanics:
A recruiter (professionalisation): AI screens CVs automatically, the recruiter is left with the more demanding tasks — contract negotiation.
A lawyer (professionalisation): AI helps with basic tasks like summarising documents, the human is left with building a case in court.
A stock clerk (democratisation): AI performs the complex tasks like inventory management, the human is left with the less expert ones — moving goods around the warehouse.
Medical secretaries and systems administrators — examples of roles where AI automates a significant share of expert tasks previously performed by people.
Growth in the number of skills required (relative to 2018)×2
02
Growth in average advertised wage (relative to 2021)42%
03
Growth in the number of openings (relative to 2018)×2
A methodological caveat: because of data-robustness requirements only six countries are included, those for which Lightcast data is available from 2012.
For bookkeepers and accounting clerks spreadsheets performed many of the hardest parts of the role — in effect democratising it and sending numbers into a gradual but steady decline.
For financial analysts spreadsheets gave a powerful new tool enabling analysis of unprecedented complexity and flexibility — in effect professionalising the role. The number of financial analysts began a steep climb that continues into the 2020s as new areas of financial analysis appeared, many with rising wages (US data).
Four questions for forecasting a role's fatePwC 2026
Expertise: how does AI change the level of human expertise required?
Supply and demand: how may demand for this occupation and the supply of workers for it change as the role is rebuilt?
AI's limits: where is human involvement needed to supervise or assist AI — for example, to check output quality or handle atypical cases?
Environmental forces: how do external forces — from business-process bottlenecks to regulation — constrain or shape the use of AI?
The source's wording: working through these four questions, starting with the change in expertise, helps reveal how AI is redefining the partnership between human and technology in a specific job.
Nursing assistants — a role with rising expertise requirements. AI takes the basic tasks such as tracking vitals and medication schedules, leaving the more expert ones — building relationships with patients, responding to unexpected situations with empathy and tact. You might expect the trajectory of 1980s financial analysts. But AI's effect here may be limited by regulation, the difficulty of integration into workflows or a limited supply of people willing to enter a demanding profession.
Childcare managers — a role with a falling entry bar. AI helps with the more expert parts: managing budgets, meeting government requirements. But instead of stagnating, the number of openings more than doubled since 2019 (+111%) while wages grew only 8% — because of enormous latent demand for childcare combined with the democratisation of the role, letting far more people do it.
The mechanism of the second case is described by David Autor in “Expertise” (2026): when technology lowers the expertise a job requires, the pool of suitable candidates widens, which lets an occupation grow in numbers while wages may stagnate. The source's analogy: taxi drivers — as satellite navigation reduced the need for specialised knowledge of roads, far more people could take the position, driver numbers soared and wages stagnated.
The practical conclusion
The category sets a likely trajectory but does not predetermine it. A forecast for a specific role requires all four questions.
What makes this sharp is that the contraction and the growth happen inside one and the same entry level — and both are visible on a single axis that crosses zero. What follows is how PwC defines a “grown-up” opening, three independent sources that converge on one point, and what employers say from their side.
Entry to a profession: two halves with opposite signs
The axis crosses zero: downward means a real contraction, not slower growth. The only story where three independent sources converge on one point.
−16%
of entry-level jobs in AI-exposed fields
12%
of CEOs expect hiring of experienced specialists to fall — against 49% for juniors
−20%
employment of US software developers aged 22–25 by 2024, while older ones grew
73%
the US share of the job-posting sample — PwC's caveat: the results are mainly driven by US data
Stanford University's analysis: −16% of entry-level jobs in AI-exposed fields. Much deeper declines are recorded in industries AI affects strongly — finance and technology (PwC 2026, citing Brynjolfsson et al., “Canaries in the Coal Mine?”, 2025, and Jacquet, “The Crisis of Entry level Labor in the Age of AI”, 2025).
PwC Global CEO Survey: 49% of CEOs expect AI adoption to reduce hiring of junior employees over the next three years, against 12% for experienced specialists.
Stanford: employment of US software developers aged 22–25 fell by almost 20% by 2024, while the numbers of more senior developers keep growing.
Entry-level occupations most exposed to AI — a junior data analyst, say — are evolving fast, demanding skills traditionally expected of senior employees.
The entry-level openings most exposed to AI are seven times more likely to demand skills traditionally expected of an experienced specialistthan those barely exposed to AI.
What those skills are (the source's verbatim list):
Motivational leadership
Team building
Managing people
Stakeholder management
Process management
Mentoring
Data-driven decision-making
The source's definition: a skill counts as traditionally seniorif in 2019 it was mentioned more than 50 times in postings for experienced specialists that AI affects strongly and no more than 5 times in entry-level postings that AI affects strongly.
The question the source frames this with: “Did your first job require these skills?”
The definition of a role with raised requirements: the posting contains at least 10 mentions of a skill that is both new and traditionally senior. A skill counts as new for an occupation if in 2025 it is mentioned more than 10 times in entry-level postings and no more than 5 times in 2019 for the same occupation.
The quartile AI affects most — the only one where the number of early-career openings has plateaued. The other three quartiles keep growing.
The source's caveats: (1) careful interpretation is required — this is not a claim that AI causes these effects; other shocks and structural features of top-quartile occupations may contribute. (2) The results are mainly driven by US data, which provides around 73% of all postings in the sample of countries with data from 2012.
“Not all AI-exposed entry roles are contracting. AI-exposed entry to a profession is in fact growing up in its requirements, with growing opportunities for workers whose roles AI has rebuilt to be even more complex and demanding. Companies (and education) will have to rethink how they train, mentor and support early career paths to help junior specialists develop and demonstrate the skills of an experienced specialist much sooner”.
The emerging roles most expected in connection with agents:
Q4 2025
01
AI Prompt Engineer — designs and optimises prompts71%
02
AI Performance Analyst — assesses the return from agents59%
03
AI Trainer / Data Curator — manages training sets, labels data, fine-tunes models58%
Microsoft/LinkedIn 2025 adds a list of roles under consideration: AI Trainer (32%), AI Data Specialist (32%), AI Security Specialist (31%), AI Agent Specialist (30%), AI ROI Analyst (29%), AI Media & Content Manager (29%), AI Finance Strategist (28%), AI Customer Success Lead (28%), AI Business Process Consultant (28%), chief AI officer (27%) (Microsoft WTI 2025, footnote 4).
The acceleration is visible even in the instrument that measured it a year ago: the 2025 barometer gave the speed of skill change as scores by quartile, the 2026 barometer as the percentage gap between the extremes. What follows is how that “net skill change” is counted, what the EPOCH framework is and why PwC puts empathy and judgement on a par with AI skills.
Speed of skill change by quartile, the 2025 barometer
The values rise monotonically from the bottom quartile to the top — in the 2026 barometer the same gap grew from 66% to 116%.
116%
by this much faster the skill requirements change in the occupations most exposed to AI — 2.2 times
The gap in skill-change speed between the extreme quartiles116% — 2.2 times faster
The definition: net skill change is counted as an aggregation of the percentage-point difference between 2019 and 2025 in the share of each skill making up an occupation.
In 2025 the occupations most exposed to AI evolve their skill set 2.2 times faster than the least AI-exposed.
By quartile (2019–2025 data, PwC 2026): the values rise monotonically from the bottom quartile to the top.
For comparison, the 2025 edition gave by quartile: 2,0 / 2,3 / 2,7 / 3,3.
Pete Brown, Global Workforce Leader, PwC UK: “Key skills used to last four to six years. Now, in the age of AI, we are talking about skills that change and transform rapidly every 18 months, every 12 months”.
Carol Stubbings, Global Chief Commercial Officer, PwC UK: “There is a narrative that AI is coming for jobs, and it creates a lot of fear and anxiety. We know that every industrial revolution creates more jobs than it destroys. The problem is that the skills for the new jobs can be very different. So the problem, as we see it, is not that there will be no jobs. The problem is that workers have to be ready to take them”.
Developed by Loaiza and Rigobon (MIT, 2025–2026). It describes the abilities where the human matters especially: they are more resistant to automation and complement AI — that is, they are necessary for people to work alongside AI systems:
PCapability Presence, Networking & Connectedness · Content Physical presence, networking, navigating social contexts — critical for building trust and collaboration
Average EPOCH score of new tasks by quartile of AI exposure (2022–2025):
The ratio of the top quartile to the bottom is 2,5×.
The source's wording: new tasks added since 2022 to the roles AI affects most (and to roles of both types) rest 2.5 times more often on abilities where the human matters especially.
Method: the average EPOCH score of new tasks is calculated in each SOC-2018 occupation, then averaged across quartiles of AI exposure.
PwC's wording, placed at the end of the section: “The more AI is applied, the more distinctly human expertise is valued”.
In practice this means a reskilling programme consisting only of AI skills covers the smaller part of the change. The other half — empathy, judgement, creativity, leadership — is placed by PwC as strategically equal in importance.
The premium is not one number but a series: the barometer has counted it for three years and along the way recalculated last year's base itself. Alongside it is demand not for a skill in an ordinary candidate but for AI specialists themselves: where it grows fastest and how employers answer it.
A discrepancy in the base: the 2026 barometer says “rose to 62% from 57% in last year's barometer”, whereas the 2025 barometer gives 56%. Probably a recalculation after the sample was expanded.
11,4%
of all 2025 openings
10%
of those hired globally
70%
of the skills used today in most occupations, with AI as the catalyst
52%
of employees say their work feels chaotic and fragmented; among leaders it is 57%
Method: postings within a sector are split into AI and non-AI; the premium is estimated as the wage difference between the groups within the sector. This is a snapshot for the year, not a growth rate.
The spread between the extreme sectors is more than sevenfold. That makes using the average premium of 62% as a benchmark for a particular company meaningless.
Hiring of AI specialists — workers with advanced AI skills such as machine learning — grew in 2025 eight times fasterthan hiring overall.
It is growing in every territory of the study. Some developing economies show faster growth than many mature ones — a sign of rapid expansion of AI capability and talent demand.
It is growing in every sector; TMT leads, where 11.4% of all 2025 openings were for AI specialists.
The definition: AI postings are those requiring at least one AI-related skill.
LinkedIn via Microsoft: over the past two years employers created at least 1.3m AI-related jobs — data annotators, AI engineers, forward-deployed engineers. Five years ago these roles did not exist, but they quickly became material to digital economies.
LinkedIn via Microsoft 2025: more than 10% of those hired globally have job titles that did not exist in 2000. By 2030 LinkedIn forecasts a change in 70% of the skillsused today in most occupations, with AI as the catalyst. AI literacy was named the most in-demand skill of 2025; human strengths are also growing — conflict resolution, adaptability, process automation, innovative thinking.
LinkedIn via Microsoft 2025: the most notable startups grew headcount 20.6% year over year — almost twice as fast as Big Tech (+10.6%). A significant share of talent is flowing out of Big Tech and staying in the startup world.
52% of employees and 57% of leaders say job security in their industry is no longer a given. At the same time 81% of employees did not change jobs over the past year — the market is frozen.
The shortfall inside a company is closed in several ways at once — retrain your own, hire into new roles, redesign existing ones, send it outside — and each way has its own share. What follows is what exactly people are taught about working with agents, how expectations of human–agent work swung over three quarters, and why in the data function all of this is harder than average.
The “build, buy, bot, borrow” approachIBM CEO 2025
IBM's 2025 wording:
Build — retrain the people you have
Buy — hire the ones you need
Bot — add AI assistants and agents into workflows wherever possible; 65% CEO say they will use automation to close skill gaps
Borrow from partners — take from partners what cannot be found otherwise
Accompanying figures: 67% CEO say differentiation depends on having the right expertise in the right positions with the right incentives. 54% hire into AI-related roles that did not exist a year earlier.
57% CEO say outsourcing non-core activities delivers strategic advantages, even at the cost of some loss of control. To manage the risk, 66% say their strategy is to concentrate on fewer, higher-quality partnerships.
The interaction will be more equal, with neither side permanently in chargeQ3 2025 49%*
02
People will predominantly manage and direct agentsQ3 202576% · Q1 202657%
03
Agents will take leading roles in managing specific projects with people on the teamQ3 2025 9% · Q4 202544%
04
There will be no significant changesQ3 2025 0%
KPMG calls the Q3 → Q4 shift “large”: the share expecting agents to take leading roles grew from 9% to 44%. In Q1 2026 the pendulum swung back: 57% again expect people to predominantly manage agents.
A separate crisis in the data functionIBM CDO 2025
Every hiring indicator in the function worsened:
01
Hire for data roles that did not exist a year earlier in connection with generative AI82%
02
Have difficulty attracting or retaining people for key data roles77%
03
Recruiting and retention deliver the needed skills and experience53%
04
Attracting and retaining talent with advanced data skills is a top challenge47%
The source's explanation: team composition is evolving too fast, and recruiting effectiveness turned out to be below half.
The wording: “Business teams need people with data skills to bridge the gap between what they do today and what will be required of them tomorrow. But the growing talent shortage is now a barrier to building broader data literacy across the organisation”.
The three answers are worth seeing separately, with the data under each: at PwC, headcount and wages by company quartile; at Microsoft, how executives set priorities for the next eighteen months; at Stanford, age. And the verbatim caveat about survivorship bias that PwC writes itself and that is worth quoting alongside the conclusion that “AI is a job multiplier”.
1. PwC — market data. Headcount grows faster at the AI-exposed.
Growth in headcount by quartile of companies' AI exposure, base 2018:
Headcount growth against the 2018 base
At the superstars headcount growth is roughly in line with the rest of the top quartile — 44%.
The source's wording: “AI-exposed companies, and especially the superstars, are not using AI only to find efficiencies or cut headcount. They appear to be using AI to fuel growth, which in turn drives expansive hiring and wage growth. Far from being a job killer, AI may turn out to be a job multiplier”.
Confirmation from PwC's AI performance study: 32% of leaders on AI performance expect headcount growth of 5% or more, against 17% at the laggards.
Wage growth across the same quartiles against the 2018 base: least AI-exposed — 16,6%, most AI-exposed — 24,4%. At the superstars wage growth is 68% — markedly above the rest of the top quartile.
2. Microsoft — executives' intentions. A third is considering cuts.
Workforce strategies for the next 12–18 months, share putting it in their top 3:
78% of leaders are considering hiring into AI-specific roles; among pioneering companies — 95%.
3. Stanford — the effect is concentrated in the hiring funnel.
“AI's labour-market effects show up unevenly, concentrating in hiring funnels and among the youngest workers in AI-exposed occupations. Employment of software developers aged 22–25 fell by almost 20% since 2024. Employer surveys point to further change: a third of respondents expect changes”.
How to reconcile this
The three answers do not contradict each other if read as different levels of aggregation:
What is happening
01
Total headcount at large firmsis growing, and faster at the AI-exposed
02
Executives' intentionsa third is considering cuts, but the priority goes to training and maintaining headcount
03
Entry to a professionis contracting in the most AI-exposed roles
04
Inside the entry levelit diverges: the “grown-up” ones +35%, the rest −10%
The aggregate conclusion that “AI creates jobs” is true — and useless for a twenty-two-year-old specialist in a highly AI-exposed occupation.
Orbis company data is not intended to represent economy-wide employment. The analysis is based on larger formal firms with available financial and headcount data for 2018 and 2024/25, with an initial filter for companies with annual turnover of at least $50m; firms that left the market over the period are excluded, which creates survivorship bias. In some countries financial reporting requirements also differ, which means firms with consistent data may be larger, more formalised or more successful than the wider business population. The key interpretation is the relative difference in headcount growth between more and less AI-exposed companies, not the absolute growth rates compared with the wider economy.
This caveat has to be given every time the conclusion “AI is a job multiplier” is quoted.
The source frames its own conclusion cautiously: “Our analysis of more than a billion job postings worldwide does not support fears of an AI-driven job apocalypse — but it does reveal a deep rebuilding of the labour market and of the skills needed to succeed”.
Everything on the section is gathered here, plus the non-matching definitions.
The spread of regular use across functions is 27 points, half the spread in agent deployment across the same functions: people are ahead of organisations. The series across eleven functions is given in the deep dive “Who uses AI: levels, countries, functions”.
“Regular user” — the definitions do not matchBCG 2025
01
BCG (both years)Definition uses AI daily or several times a week · Consequence 72% of all in 2025, 74% of frontline employees in 2026
02
Microsoft WTI 2025Definition uses generative AI at least several times a week · Consequence comparable with BCG
03
Microsoft WTI 2026Definitionan AI user = uses generative AI for work at least occasionally (from less than once a month to more than once a day); those answering “never” were screened out · Consequencenot comparable: the whole 2026 sample consists of AI users only
The practical implication: all WTI 2026 percentages are shares of AI users, not of all workers. They cannot be compared directly with BCG's percentages or with WTI 2025.
2025. All data is aggregated and anonymised, through 15 February 2025, excluding education and European tenants:
Interruptions: a rolling 28-day sum of pings (meeting invitations, emails, chats) per unique user per working day. The two-minute figure reflects the average time between pings in an eight-hour working day; 275 is for a 24-hour day. For the top 20% of users by volume of pings received.
Last-minute PowerPoint edits: a rolling 28-day sum of views and edits per meeting participant, measured in fixed time windows before meetings.
Spontaneous meetings: a rolling 28-day volume of unique meetings per user per working day; the top 20% of users by meeting volume.
Off-hours chats: a rolling 28-day sum of chats sent outside Monday–Friday 9 to 5.
Late meetings and cross-time-zone work: a rolling 28-day sum of meetings starting between 20:00 and 23:59 in each participant's local time.
2026:
Classification of goals in Copilot: one week of February 2026. The percentages show the share of each activity among classified user goals — not time and not the number of sessions. Each interaction is classified by the user's intent against the O*NET Intermediate Work Activities taxonomy and then mapped to Generalized Work Activities.
Active agents: the annual change in the number of unique active agents on the Microsoft 365 Copilot Agents and SharePoint platform over a rolling 28-day window.
Industry agent telemetry: March 2025 – March 2026, all metrics expressed as shares and ratios, with no absolute values.
What is covered in this section
A ceiling that a year ago looked structural came off in a single year — here is the whole series by level, with BCG's wording from both editions, including that dip among leaders which the source does not explain. Alongside it, the same picture by country across two years, by function and by what people feel about it.
2025 — the “silicon ceiling”. Overall use 72%, but among frontline employees it stalled: 52% a year earlier and 51% in 2025, that is minus one percentage point after a 32-point jump. BCG's wording: “AI is now part of our everyday working life — but frontline employees have hit an adoption ceiling”.
+23 pp
the fastest change in any behavioural indicator over the year
33 pp
the gap between the leading and the trailing country on regular use, 2026
75%
of frontline employees in HR use AI themselves — while agents are at 2% of organisations
67%
of regular users report a rise in job satisfaction
2026 — the ceiling is broken. Frontline employees +23 pp over the year — the fastest change in any behavioural indicator over the year. Managers +10 pp. BCG's wording: “No more silicon ceiling”.
An anomaly the source does not explain: leaders 88% (2024) → 85% (2025) → 93% (2026). The 2025 dip is shown as it is.
Share of regular AI users. In both years the Global South is ahead of the North, but the readings were taken on different samples: in 2025 all respondents were surveyed, in 2026 only frontline employees (n=4,040). What is worth comparing is the order of countries, not the levels.
The Middle East in the 2025 reading is the UAE, Saudi Arabia, Kuwait and Qatar.
A persistent pattern: the Global South leads the North in both years. The gap between the leader and the laggard in 2026 is 33 percentage points.
Set against agent deployment this gives an unexpected contrast: in HR 75% of frontline employees use AI themselves, but agents are deployed at only 2% of organisations. People in HR use AI themselves, but the organisation does not put agents there.
Attitudes to AI, share putting a feeling in their top 2 of seven options:
The source's wording: “As AI goes mainstream, confidence is rising and concerns are falling”. In the 2026 edition BCG measures something else: job satisfaction — a rise is reported by 67% of regular users — and cognitive load (41%).
The two waves measured savings differently, and both are more honest read alongside their definitions — while the guidance gap breaks down by level far more sharply than a single overall figure. What follows is where the time actually goes, what a redesigned process does with it and what BCG and IBM advise about it.
Who saves at least eight hours a week
Frontline employees, 2026. The spread between functions is 12 points: the savings are distributed far more evenly than the guidance on what to do with them.
The organisation gives limited or no guidance on what to do with the time savedFrontline employees66% · Managers 66% · All 61% · Leaders 52% · Frontline-employee–leader gap+14 pp
02
Do not reinvest the time saved into more strategic workFrontline employees58% · Managers 43% · All 45% · Leaders 36% · Frontline-employee–leader gap+22 pp
2025 for comparison: “only a third receive guidance on how to reallocate that time, which mutes the impact”.
Shares of respondents doing this with the time saved (multiple choice, BCG 2025):
The source's wording: “Time savings are reallocated to very diverse tasks”.
The most frequent answer is intensification. That is the mechanics of the leak: without explicit guidance, time goes by default into increasing the volume of the same work rather than reallocating towards more valuable work.
Employees at companies redesigning workflows save more: 55% against 29% save more than an hour a day (2025); 53% against 31% save at least a day a week (2026).
BCG's first imperative for CEOs: “Change the scoreboard: measure value, not reach. The time individuals save leaks out of the organisation unless it is tracked and deliberately reinvested. Look at business outcomes, not at usage”.
IBM's recommendation with a specific number: before the next budget cycle closes, agree a fixed share of productivity gains to be reinvested — typically 60–80% — and commit to funding faster cycles of experimentation and scaling. Stop local efficiency improvements that unlock no new capabilities.
Here is the whole curve — including what in-person format and a live coach add on top, and why the lever has lain untouched for a second year. And the barriers through frontline employees' own eyes: no skills given, no tools given, no support given — hence shadow AI.
Share of regular AI users depending on the volume of training received:
Share of regular users
The jump between “no training” and “1–5 hours” is +45 percentage points. Everything above ten hours adds little.
The source's note on the same slide calls the 67/79 pair the five-hour threshold, although on the panel itself it is labelled as the in-person format. The source shows the five-hour threshold in the adjacent panel — that is the 18 / 63 / 82 / 89 series above.
The independent effects of format
The source's wording: “At least five hours of instruction, in-person sessions and coaching are the key components of effective training”. These three components also significantly raise employees' confidence in AI and improve the quality of results from working with AI.
Expect to need major reskilling within five years88%
03
Say skill expectations for their role have changed72%
BCG notes both figures as unchanged from 2025. The source's headline: “Demand for reskilling is loud and persistent. The response is still insufficient”.
The 2025 wording: “Training is often too short or superficial — only 36% of employees say they have been trained on the skills needed for AI transformation”.
The three adoption barriers as frontline employees see themBCG 2025
The top 3 barriers named by frontline employees (n=3,537):
Lack of skills or training. Training is often too short or superficial; only 36% believe they have been trained adequately.
Limited access to the right tools. Almost four in ten (37%) say the company does not supply the tools they need. When corporate solutions fall short, 54% say they would use unauthorised AI tools, raising security risks.
Lack of leadership support. Only 25% of frontline employees say they received sufficient leadership support on how and when to use AI at work.
The source's definition: shadow AI — the use of AI tools, systems or models inside an organisation without the explicit knowledge, approval or governance of central IT or data teams.
Would use AI without company support
The generational gap +19 pp.
The practical implication: the share of unauthorised use is not an indicator of employees' indiscipline but an indicator that corporate tools are inadequate. With 37% not getting the tools they need and 54% prepared to go around restrictions, shadow AI is a predictable consequence, not an anomaly.
Each of the six changes is given not only as the share who agree but also with those who shrugged or disagreed; you can see where agreement is genuine and where it is fragile. What follows is what makes up half of the conversations with Copilot, how employees in the leading group differ from the rest, and how WTI describes the four modes of working with an agent.
Skill expectations for the role changed\*Neutral 16% · Disagree 13%
02
AI took the simpler tasks, leaving the complex, high-stakes workNeutral 17% · Disagree 16%
03
The bar for what counts as “good enough” roseNeutral 22% · Disagree 17%
04
Time spent checking and correcting AI output grewNeutral 23% · Disagree 25%
05
Roles shifted towards managing and directing AINeutral 23% · Disagree 30%
06
The number of decisions taken grewNeutral 24% · Disagree 35%
\* Share reporting that AI will significantly or moderately change skill expectations for the role.
The source's headline: “AI is reshaping work, impacting the very nature of jobs and management”.
A row that is easy to miss: 52% report an increase in time spent checking and correcting AI output. Time savings and a rise in the verification burden coexist.
A privacy-preserving analysis of more than 100,000 chats in Microsoft 365 Copilot: 49% of all conversations support cognitive work — helping analyse information, solve problems, evaluate and think creatively. The rest: work with people 19%, producing output 17%, finding information 15%.
Four mutually exclusive categories of one classification, summing to 100%.
The detail within the categories: decision-making and problem-solving 28% · obtaining information 13% · communicating with supervisors and peers 8% · thinking creatively 6% · analysing data 5% · interpreting information for others 5% · documenting and recording information 5% · processing information 3% · evaluating the qualities of objects and services 3% and others.
The source's wording: “Almost half of Microsoft 365 Copilot chat use supports analysis, decisions and problem-solving — the kind of high-value work that used to require deep expertise”.
Deliberately do part of the work without AIto keep their skills in shapeThe rest 30% · The leading group43% · Δ +13 pp
02
Deliberately stop before starting work to decide what AI should do and what the human shouldThe rest 33% · The leading group53% · Δ +20 pp
This is the only practice that directly answers Stanford's warning that heavy reliance on AI may carry long-term costs for learning and slow skill development.
A conceptual framework WTI 2026 on two axes — how the human is involved in the work (directing or supervising) and how intensively the agent is used (assistant or partner):
Mode
01
DelegatingThe axes the human directs + the agent as partner · What it looks like turning raw notes into a structured summary or draft · assembling a regular report from standard inputs · preparing a research summary with sources cited after the scope is defined
02
AskingThe axes the human directs + the agent as assistant · What it looks like looking up a fact, a date, a definition · rewriting a phrase for clarity, tone or length · reformatting a small table
03
CollaboratingThe axes the human supervises + the agent as partner · What it looks like refining a proposal through several rounds of feedback · building an analysis where each result changes the next question · preparing a communication where tone and delivery require judgement
04
ExploringThe axes the human supervises + the agent as assistant · What it looks like checking whether Copilot can handle a new workflow · testing different prompting strategies · probing the limits of an agent's autonomy
The source's key observation: “What distinguishes employees in the leading group is not which mode they use but knowing which mode a task requires”. Routine execution, research and synthesis are delegated. As AI does more of the work, people stay involved by setting direction and taking responsibility for how the results are used.
The source's caveat: this is a conceptual framework, not a data visualisation; the placement of each mode is qualitative.
The source's wording: “As AI use matures, the most effective users will not be those who do more things faster. They will be those who redefine their value around what only humans can do: setting clear intent — defining the desired outcome and the quality bar — and designing how the work is done by people and AI together. The question stops being “which tasks define my job” and becomes “which outcomes am I now able to deliver””.
The paradox is named that way by BCG itself, and both its halves are at a maximum on the top floor: whoever gets the most enjoyment also carries the greatest load. What follows are the levers on which value and enjoyment rise together, the reshaping companies that get both at once, and the first warning of what the load will turn into a quarter later.
The top 5 organisational levers, ranked by combined uplift in employee enjoyment and measurable impact:
The definition of measurable impact: an improvement in key business metrics attributed to AI.
Not one lever improves one at the expense of the other. The source's headline: “The actions that drive business impact are the same ones that make employees thrive”.
Confirmation through a comparison of groupsBCG 2026
Reshape and invent companies against deploy ones deliver simultaneously both more value and a better experience:
Save at least a day a week: 53% against 31%
Rise in job satisfaction: 68% against 48%
See measurable improvement in business metrics: 67% against 43%
The rise in cognitive load among 41% is a direct precursor of what KPMG will show a quarter later: resistance caused by increased load and complexity doubled from 28% to 51%.
Cognitive load is not in itself bad — in the first months it is an engine of enjoyment. The problem arises when it persists without strategic clarity.
Every factor here appears twice: as it was in the first six months and as it became a year on — and it is the distance between the two points, not their height, that answers the question. From it comes the forecast for an adoption programme that rested on enthusiasm.
A comparison of regular AI users with under 6 months' and over a year's tenure (n=6,998):
A year on it is clarity, not novelty, that holds
Values for tenure over a year. To the right is what adds to enjoyment of work, to the left what takes it away. A programme built for the enthusiasm of the first six months sags in the second year.
A clear AI strategy< 6 months +24 pp · > 1 year+30 pp · Change strengthens
05
Clear guidance on the time saved< 6 months +23 pp · > 1 year+25 pp · Change strengthens
06
Difficulty of showing your unique value< 6 months −29 pp · > 1 year −26 pp · Change barely changes
07
Insufficient training< 6 months −15 pp · > 1 year −16 pp · Change does not change
Reading the source's note: “+/− pp is the additional share of employees enjoying their work when the factor is present (+), or the share lost when the obstacle is present (−)”.
Three conclusions
Novelty and intellectual challenge work at the start and almost halve within a year. For the first six months people like that the work became harder and more interesting. After that the effect runs out.
Strategic clarity and clear guidance, by contrast, strengthen over time. They are what sustains the effect once the novelty passes.
The obstacles do not weaken at all. Neither the difficulty of showing your value nor insufficient training softens with experience. That means they cannot be “waited out” — they require intervention.
“Want the AI honeymoon to last? Strategic clarity beats tools in achieving sustained impact”.
And further: “AI's novelty and cognitive stretch fuel enjoyment at the start. But sustained enjoyment comes from strategic clarity. Employees thrive when the direction is real and the message reaches them with strong CEO engagement”.
The practical implication
An organisation relying on the enthusiasm of the early phase will see a sag about a year after mass adoption. The only factors that work on that horizon are clarity of strategy and guidance on using the time saved; both belong to the operating model and leadership, not to employee experience.
The quarter's full series — with all the rows that did not fit in the article, and with the caveat about how these numbers were extracted from mixed-up PDF columns. Then comes employees' pullback from agents against a backdrop of calm corporate deployment, and the token race KPMG warns against in the same report.
growth in the share expressing resistance to agents over the quarter — from 5% to 20%
41%
would consider introducing a token race — incentives for maximum token consumption
The Q2 2026 breakdown: mixed reaction 37%, mild resistance 19%, significant resistance 1% (20% expressing resistance in total).
A caveat on extraction. This page extracts from the PDF with its columns mixed up; all values were resolved by the coordinates of the text blocks. The direction of change is set by the source's own verbs: resistance because of load “almost doubled”, anxiety about training “fell by almost half”, concerns about skills and jobs “fell quarter on quarter”.
What the change of nature means
Fear went away, fatigue grew.
The concerns “I will be replaced” (jobs) and “I do not know how” (skills, training) all fell — all three.
The concern “it got harder” almost doubled and became dominant.
KPMG's wording: “Resistance caused by increased workload or complexity — where AI makes work harder rather than easier — almost doubled”.
This is exactly what BCG's joy paradox predicted: cognitive load rises (41% report an increase), and without strategic clarity it turns from a challenge into exhaustion.
The employee agent adoption rate fell from 55% to 43% in a quarter. This happens while organisations' overall agent deployment holds steady (53%) and chaining agents together doubles.
The organisation moves forward, employees pull back. This is one of the most recent and most alarming signals in the report.
The series of employees' perception of agents over the yearKPMG Q3’25
41% of leaders would consider introducing a token race — incentives for maximum token consumption, often with internal leaderboards — to encourage AI use.
KPMG's own warning: “Efforts that put usage metrics above meaningful outcomes risk locking in the wrong behaviour”.
And yet 47% agree or strongly agree that employees who use AI effectively and productively outperform the rest. So the performance difference is real — the question is which indicator people try to reproduce it with.
This directly contradicts BCG's first imperative for CEOs: “Change the scoreboard: measure value, not reach”.
“Successfully scaling agent adoption will require a sharper focus on employee experience, clearer alignment with outcomes and more deliberate approaches to engagement”.
The state AI arrived into was measured by Microsoft with telemetry, not a survey: here is what the capacity gap is made of and what a day torn into pieces looks like. And along with it, why people go to AI rather than to a colleague for help, and how pioneering companies differed from everyone else even then.
Of the global workforce — both employees and leaders — say they lack the time or energy to do their job80%
02
Of leaders say productivity must rise53%
03
Of leaders are confident they will use digital labour to expand workforce capacity over the next 12–18 months82%
Microsoft calls this the capacity gap: the business demands more than people can sustainably deliver.
By region (leaders who believe productivity must rise / workforce without time and energy): Asia-Pacific 61% / 84% · Europe 49% / 75% · Latin America 44% / 78% · North America 43% / 72%.
Spontaneous meetings (not scheduled in advance)60%
02
Spike in PowerPoint edits in the last 10 minutes before a meeting versus three hours before+122%
03
Growth in meetings starting after 20:00+16% year over year
04
Meetings spanning several time zones30%, +8 pp since 2021
05
Of employees say their work feels chaotic and fragmented48%
06
Of leaders say the same52%
The source's wording: “AI delivers real productivity gains, but that is not enough. The pace of business still outruns how we work today. To keep up, companies need more than adding AI to existing workflows — they need to rethink the very nature of knowledge work”.
Reasons for choosing AI over a colleague or manager:
+Reason8
01
Availability 24/742%
02
Machine speed and quality30%
03
An endless flow of ideas on demand28%
04
Unlimited capacity — infinite time and energy for repetitive or complex tasks23%
05
Fear of human judgement — with AI it is more private, it does not judge17%
06
The demands of working with colleagues — AI requires less explaining16%
07
Taking the credit — you can take all the credit yourself15%
08
Friction with colleagues — AI does not get irritated or lose patience8%
The source's conclusion: “People use AI for its unique capabilities rather than for advantages over colleagues”. The reasons connected with avoiding human qualities are at the very bottom of the list.
52% see AI as a command tool — they give direct, simple commands to get work done. 46% see AI as a thought partner — they hold a conversational exchange to challenge their thinking, sketch ideas, spark creativity. The remaining 2% do not know.
The wording: to work effectively with agents, every employee will have to adopt the thought-partner mindset and build the related skills — iterating with AI, knowing when to delegate, prompting with context and intent, refining outputs instead of accepting first drafts, spotting weak reasoning and gaps, knowing when to push back and redirect the conversation or the plan.
28% of managers are considering hiring AI workforce manager to lead hybrid teams of people and agents; 32% plan to hire AI agent specialists to design, develop and optimise agents over the next 12–18 months.
Within five years leaders expect their teams' scope to include: building multi-agent systems to automate complex tasks 42% · training agents 41% · redesigning business processes with AI 38% · managing agents 36%.
51% of managers say AI training will become a key responsibility of their teams within five years; 35% are considering hiring AI trainers.
83% of global leaders say AI will let employees take on more complex strategic work earlier in their careers.
Pioneering companies (844 of 9,037 leaders surveyed) against the global average:
The gap on fear of job loss is half as high at pioneering companies. BCG will later confirm the same observation on a different sample: employees at companies genuinely redesigning work less often see agents as a threat if they understand how they work.
A general note: none of the five shifts is complete. Each has counter-examples, and they are given below. A shift is a change in the dominant frame, not the disappearance of the previous one.
The five shifts in figures
What is covered in this section
The one shift that is not on the picture at the start of the section: no comparable pair of readings was found for it, but both formulations were — about data locked inside functions and about infrastructure that cannot absorb change fast enough. Both in full, and alongside them what the data foundation actually was by 2025, what architectural inflexibility cost by 2026 and what in the same data contradicts the shift.
2025 (IBM CDO 2025, foreword by Ed Lovely, VP and CDO of IBM):
“This year I personally spoke to more than 150 enterprise clients, and one problem came to the fore: data is locked inside functions. Finance has its data. HR has its own. Marketing, supply chain, legal — every function's data operates in isolation. No shared taxonomy. No common standards. No end-to-end visibility. This is not merely an operational inconvenience. It is the Achilles heel of enterprise AI transformation. When data lives in disconnected stores, every AI initiative turns into a drawn-out six-to-twelve-month data-cleaning project. Teams spend more time finding and reconciling data than generating meaningful conclusions”.
2026:
“The key constraint is adaptability. Most organisations were built for control, standardisation and efficiency. But in an AI-driven landscape it is decisions on architecture, governance and investment that set how fast they can move, what level of risk they can carry and which opportunities remain reachable”.
“Today the main constraint on business performance is neither ambition nor technology. It is the ability of IT infrastructure to handle change at machine speed without losing control, resilience or economic discipline”.
Two more quantities from the same study do not go on the share scale: cloud spend exceeded the original forecast by 48% on average, and early architectural replaceability delivered +10% on return on AI investment.
The data foundation is objectively solid. All three indicators above — 81%, 75% and 74% — speak of mature work with data, and the share of the IT budget on data strategy grew over two years from 4% to 13%. The source does not claim “data has stopped being a problem”, and neither do we: the bottleneck moved because requirements grew faster than progress.
Data returned to the top of the barriers in 2026. At KPMG in Q2 2026 data readiness and availability is the main barrier to deploying agents (58%), ahead of the complexity of agentic systems (38%) and technical skills (37%). So at the level of specific projects data has not gone anywhere.
Data quality remained a barrier even with a stronger foundation: it was named by 82%.
Confidence in one's own data stayed at a quarter, despite all the progress: 26% on new revenue-generating streams, 26% on unstructured data, 19% on architecture completeness.
An honest wording of the shift: not “data is no longer a problem”, but “data stopped being the only bottleneck; the adaptability of architecture and the operating model was added to it and came to the fore”.
The sign changed above all in words: governance moved from the column “what slows things down” to the column “what nothing scales without”. Taken in order: both formulations verbatim, the oversight measures by quarter, the economic argument for designed-in control — and the gap between what is acknowledged and what has been built.
2025. Governance is present as both control and constraint: only 22% of organisations have clear rules and guardrails for AI in automated decision-making. The source's comment: “Organisations are investing in AI capabilities but most are underinvesting in the frameworks”.
Recognition versus implementation
The economic argument exists: ×16 agents at a quarter of the budget share. The governance architecture does not.
2026. KPMG puts it in the heading of an executive-summary section: “Governance is the condition for scaling”. IBM translates this into an engineering framing: “Control stops being a permissions problem. It becomes a design problem”.
Governance is recognised but not built. 77% acknowledge governance lagging adoption; 70% — that teams deploy faster than IT can track; only 11% are ready for the expected scale.
Half of employees see no rules. 50% say their company has no clear guidance on managing “human × AI” teams.
A third of organisations override AI outputs with no formal criteria.
Incidents are growing. 362 documented incidents over the year; 54 agent incidents per organisation per year.
Responsible AI is not a single scale. Improving safety can worsen accuracy, which means “governance maturity” as a single number misleads.
An honest wording of the shift: governance moved from the category “what slows things down” to the category “what nothing scales without” — in rhetoric and in economic calculations, but not yet in practice. The gap between recognition and implementation is larger here than in any other dimension.
It split in two even in the headlines: a year ago the same source ran “AI is making people more valuable”, now it is “Two futures for jobs in the age of AI”. Let us look at what changed — the frame or the series themselves, what an independent reading on the youngest developers adds, and why “two tracks” does not mean “one grows, the other falls”.
2025 (PwC 2025, cover and p. 2): the report's title is “AI is making people more valuable”. The key findings: three times greater growth in revenue per employee, wages growing twice as fast, the number of jobs growing practically everywhere, including the most automatable occupations.
2026 (PwC 2026, cover and p. 3): the title is “Two futures for jobs in the age of AI”. “Rather than simply replacing jobs, AI is rebuilding them in fundamentally different ways”.
The charts side by side
Three rows measure different things — productivity growth, the wage premium and the gap in the speed of skill change. Comparison is possible within a row; between rows it is not.
01
The organising frame2025 AI exposure: most / least · 2026 mechanics: professionalisation 22% / democratisation 52% / AI barely touches it 26%
02
Wages2025 ×2 faster in AI-exposed industries · 2026 +42% faster for roles with rising expertise requirements
03
Number of jobs2025 growing everywhere, more slowly in the AI-exposed (+38% against +65%) · 2026 roles with rising requirements +39% against roles with a falling entry bar +17%
04
Entry to a profession2025 not singled out · 2026 −16% in AI-exposed fields; +35% for roles with raised requirements against −10%
Employment of software developers aged 22–25 fell by almost 20% by 2024, while older ones grew. The wording: the effects concentrate “in hiring funnels and among the youngest workers in AI-exposed occupations”.
Both tracks are growing. Occupations with a falling entry bar deliver +17% openings and +26% wages — that is growth, not decline.
Headcount at AI-exposed companies grows faster: 52.2% against 35.7%. The aggregate claim of 2025 remains true.
The category does not predetermine the fate. PwC's own counter-examples: nursing assistants (expertise requirements rise, but growth is limited by regulation and the supply of people) and childcare managers (the entry bar falls, but openings grew 111% because of latent demand).
Survivorship bias. The headcount charts are built on large formal firms with turnover from $50m that continued to exist; those that left the market are excluded (PwC 2026, notes).
Causality is not proven. PwC notes this for every series; for the entry level directly: “this is not a claim that AI causes these effects”.
An honest wording of the shift: the overall dynamic is still positive; what changed is that inside the overall picture two diverging tracks were found, and entry to a profession is contracting. The optimistic conclusion of 2025 is not refuted, it is refined.
Over the year the suspect really did move from presentations into operation — and along with work came everything work involves: incidents, unwritten rules, undocumented processes. What follows is how this move is measured, what exactly fell behind, and why an order of magnitude still lies between “organisations are deploying agents” and “agents do the work”.
The independent reading gives single-digit percentages. Stanford: agent deployment stays in single-digit percentages across almost every business function. This is the strictest and most objective reading in the report, and it is an order of magnitude below the survey figures.
Overall deployment stalled. KPMG Q1 → Q2 2026: 55% → 53%. Only the chaining of agents together is growing.
Employees' adoption of agents fell. 55% → 43% in a quarter.
KPMG's series has a definition break between Q3 and Q4 2025, which means it cannot be treated as an exact time series.
Agents' importance in employees' eyes fell slightly: those considering them important within the next 2–3 years fell to 72% against 77% a year earlier, −5 pp.
An honest wording of the shift: agents really did leave the conceptual phase — the ×15 telemetry and the doubling of integration into workflows confirm that. But an order of magnitude lies between “organisations are deploying agents” and “functions are executed by agents”, and operating models, rules and documentation lag at every level.
The office where they looked for someone accountable was disbanded over the year: the agenda is owned by several perimeters at once, and the chief AI officer role travelled that long path from the picture at the start of the section. We take apart how the analysts' own cut changed, how the top team's attitude to risk inverted — and why this rebuild has not yet reached down to the doer.
IBM IBV2025 CEO · CMO+CSO · CDO · CAIO · 2026 CEO (business) · Tech Leader — CIO/CTO (technology)
02
The logic2025 functional C-suite roles · 2026 the perimeters of taking and executing decisions
The series on ownership of the AI agendaKPMG Q2’25
Who owns it
01
Q2 2025CIOs lead enterprise-level AI strategy — in the view of 87%
02
Q2 2026distributed: a named senior executive 34% · the CEO or executive committee 32% · the business-unit head 14% · a governance committee 10%; 67% — the CEO actively owns AI as a strategic priority
This is an observation about the analysts, not only about the market. The change in the structure of IBM's studies is the source's decision. It is backed by data (77%, 85%), but is not in itself a market fact.
Down below they do not see the changes. Only 26% of AI users consider leadership clearly and consistently aligned on AI; 33% of frontline employees consider that leadership communicates clearly; 28% see words matching actions.
The “skills exist — they are not used” gap did not close. 86% of CEOs are confident in employees' skills; the same CEOs put regular use at 25%.
Decentralisation without clarity of rights is dangerous. The source itself notes: 79% are decentralising decision-making, but you first have to explain who decides what, otherwise there will be no acceleration.
At 76% penetration the CAIO role may be nominal. IBM separately recommends clarifying the mandate of existing CAIOs — which implies that for some it is not clarified.
An honest wording of the shift: the top level of management rebuilt itself quickly and visibly. Whether that reached the doer — the data says not yet.
The article names five limitations. Here each is taken apart on one scheme: which figures are affected, which conclusions weaken as a result and by how much, and what would need to be measured to remove the limitation.
$5.8bn
the median revenue of a company in the IBM CEO 2026 sample; the average is $12bn
$14.4bn
the average revenue of a company in the IBM Tech Leader 2026 sample, average headcount about 33,500
84%
of technology leaders have not operationalised financial management of AI
85%
do not have full real-time visibility of AI spend
Whom these samples see at all: the threshold for inclusion in the sample
The scale is in millions of dollars. A company on $50m of revenue applying this report's benchmarks to itself is comparing itself with something other than itself.
For each: which figures are affected · which conclusions weaken and by how much · what would need to be measured to remove the limitation.
Limitation 1. Not one source provides independently verified ROIIBM Tech 2026
Which figures are affected. Every return metric except two PwC series and one HAI series. The full list — 17 ROI metrics in D-4.0.
What weakens.
Finding 1 (return stratified) rests on technology leaders' self-assessment and on PwC's external proxy. The first is weakened, the second is not — which is why the finding held.
All the ROI premiums (+10% for a CAIO, +36% for the operating model, +10% for replaceability, +38% for three pillars) are comparisons of self-assessments between groups. The direction is probably right; the magnitudes should be treated as upper estimates.
The claim “97% see a return” is unusable for any conclusion at all: it measures the choice of a metric.
What needs to be measured. The return of an AI portfolio from management accounting data on a sample of companies with a comparable method of recognising costs and effects, with a control group. Not one source did that.
The closest to objectivity:
PwC — revenue per employee from Orbis financial reporting;
HAI — the soberest series on the size of the effect: savings under 10% and revenue gains under 5% as the most common values by function.
Limitation 2. Small and mid-sized business is practically unrepresentedKPMG Q2’26
The sample profiles.
Threshold
01
KPMGUS only, only companies with revenue of ≥$1bn
02
IBM CEO 2026median revenue $5.8bn, average $12bn; ~80% public
03
IBM Tech 2026average revenue or budget $14.4bn, average headcount ~33,500
04
PwCa filter for companies with turnover of at least $50m; those that left the market are excluded
05
BCGthe revenue distribution skews large; 23% of the sample is $500m–$1bn
What weakens. All the benchmarks in the indicator panel apply to large business. For mid-sized they are untested, for small almost certainly wrong: the economics of an AI portfolio, the need for management infrastructure and the cost structure at scale are fundamentally different.
What needs to be measured. The same indicators on a sample of companies with revenue of $10–500m. No source has such data.
The only indirect evidence to the contrary. Microsoft 2025 gives examples of very small organisations: a co-founder on the way to $2m in revenue with an AI-based staffing firm; a five-person startup, ICG, using AI from construction simulations to marketing research, with a 20% margin increase; an entrepreneur running budgeting and forecasting without a CFO thanks to a single AI tool. These are anecdotes, not a sample.
Limitation 3. There is no longitudinal data on the same organisationsMicrosoft WTI 2025
What this means. Every series is a new sample each year or quarter. When we write “integration of agents rose from 13% to 30%”, that is a comparison of two different samples.
Which figures are affected. All the “before → after” series: BCG 2025→2026, PwC 2025→2026, Microsoft 2025→2026, KPMG by quarter, IBM 2025→2026.
What weakens. Above all causality. All the “leaders versus the rest” differences are correlations:
Companies at the “reshape/invent” level deliver better results on six indicators — but they may have been stronger to begin with and therefore took on the redesign.
Organisations with a CAIO have +10% ROI — but they may have appointed a CAIO precisely because they were already more mature.
Pioneering companies thrive more often — but the definition of a pioneering company includes “saw ROI from AI adoption”, which is partly tautological.
Microsoft notes this directly: “the values show a statistical association, not a causal effect”. PwC notes it for every series: “we cannot prove causality with confidence”.
What needs to be measured. A panel of the same organisations over three or more years with the moment of adopting each practice recorded. No source runs one.
The practical rule: read all differences between groups as a description of the leader's profile, not as a recipe. “Leaders do X” ≠ “do X and you will become a leader”.
4.1 KPMG, agent deployment, Q3 → Q4 2025. The series 11% → 33% → 42% is incompatible with 26% in Q4. Affects: all citations of the agent maturity curve. Weakens: the ability to treat the series as an exact time series. Compensation: it is used as an indicator of direction; the chart spec requires the line to be broken.
4.2 PwC, the base of the AI skills premium. 56% in the 2025 barometer against “57% in last year's barometer” in the 2026 edition. Affects: indicator 38 of the panel. Weakens: minimally — a 1 pp discrepancy. Significance: it shows that PwC's series are recalculated retrospectively.
4.3 BCG, regular users among leaders. 80% → 88% → 85% → 93%. The 2024→2025 dip is unexplained. Affects: indicator 43 of the panel. Weakens: the interpretation of monotonic growth in adoption among leaders.
4.4 Extraction from decorated slides. The large figures in KPMG's and BCG's decks are laid out as separate blocks, and the extraction order does not match the visual one. Affects: potentially every figure from these decks. Compensation: disputed values were resolved by the coordinates of the text blocks. This check found and corrected two errors in the report's first edition (both on KPMG Q2’26). The full breakdown is in the appendix “Divergences between sources”.
4.5 Chapter numbering AI Index 2026. Chapter titles extract reliably, ordinal numbers do not. Affects: only the description of the AI Index's structure. Requires a manual check before publication.
4.6 The reconstructed figure of IBM's forecasts. In the figure of forecasts for 2026 and 2030 the correspondence of columns and rows was recovered from the source's textual wording; intermediate values are marked with “~”. Two anchor figures — 49% and 10% — are given directly in the text.
Limitation 5. What is not covered at allIBM CEO 2026
5.1 Russia and the CIS are absent from every sample without exception. Not one of the 20 sources includes these markets. All country conclusions come from the lists in the methodologies.
5.2 The public sector is represented fragmentarily. It appears in BCG's industry breakdowns (reshape 31% — last place among industries, BCG 2025) and IBM's (three priorities and three challenges, IBM CEO 2026), but nobody has a separate study of it. The AI skills premium in the public sector is the lowest of the eight sectors shown in the source (16% against an average across all sixteen of 62%).
5.3 Industry depth. There are cuts, there are no full industry studies. At most, three priorities and three challenges per industry at IBM (consumer, banking, insurance, energy, public sector, healthcare, telecom, manufacturing).
5.4 The economics of running AI. The topic appeared for the first time in KPMG Q2’26 and is measured by one wave of 204 respondents. The only figures: 26% full visibility of costs, a set of four cost-management practices. Confirmation from IBM's side — two indicators: 84% have not operationalised financial management of AI and 85% do not have full real-time visibility of spend. That is not enough for a robust conclusion.
5.5 The long-term effect on skills. Stanford mentions that heavy reliance on AI may carry long-term costs for learning and slow skill development, but there is no systematic data. The only counter-practice is the self-report of employees in the leading group: 43% against 30% deliberately do part of the work without AI to keep their skills.
5.6 The effect on decision quality. These reports measure speed and volume and barely measure quality. The only approaches: IBM's recommendation to measure success by “decision quality and speed” and the metric “improved analytics for senior management decisions”, which grew from 62% to 83% — but that is self-assessment again.
5.7 The cost of failures. These reports count the return of successful initiatives in detail and barely count the cost of the unsuccessful ones. The only figure: more than a quarter of AI initiatives were cancelled, postponed or failed to scale because of security problems.
How these limitations change the reading of the report
Read any return figure from an executive survey as sentiment, not as fact. The exceptions: PwC's series and HAI's series on the size of the effect.
Check any maturity figure against a figure from an employee survey and against an independent source. The divergence is itself informative — that is gap 1.
Treat “leaders versus the rest” differences as correlation. It is a profile, not a recipe.
Apply all benchmarks to large business. For mid-sized, with an explicit caveat; for small, do not apply them.
Check the base before comparing figures between slides from one source. At BCG the bases change from slide to slide.
Do not compare metrics with different question wordings. The register of incomparable pairs is in “We see” versus “we measure”, the divergences between sources in “Where the data diverges”.
@andre_dataist
Cover · cover.jpeg
