Sam Altman said AI adoption has gone slower than he expected. There has been no "iPhone moment", in his words — no shift from treating the technology as one more tool to rebuilding the work around it — and he put the blame on economic inertia: people resist change. The next day it emerged that Meta had quietly abandoned a plan to reorganise its workforce around AI agents. The internal project, OT, short for office-work transformation, would have cut some teams by 60% and replaced them with small groups supervising AI systems. Mark Zuckerberg stopped it.
Two of the loudest voices in the industry conceded the same thing inside 24 hours: deployment is harder than it looks in the slides.
That is not a verdict on the technology. AI produces more material and more code every week. Delivery Hero says its HeroGen coding agent team delivers output comparable to roughly 130 developers. But volume and value are different quantities. AI writes more text, more email, more code, and mostly does not make decisions — the system carries no conviction about which option is right. The companies that win, on the author's argument, are the ones converting messy work into dependable processes, not the ones with the most impressive demo.
The numbers around this are more revealing than either announcement. Only 6% of companies are prepared to hand key business processes fully to AI agents, according to a 2026 Harvard Business Review survey. In Germany, the Ifo Institute found the share of companies using AI rose from 40.9% to 54.5% in a year. Set those two figures side by side and the shape of the market is obvious: adoption is broad and delegation is almost nonexistent. Over half of German firms use AI; fewer than one company in sixteen anywhere trusts it with something that matters. That is not a wave of staff replacement. It is a wave of supervised assistance, which is a different product with different economics.
Altman's diagnosis of the cause is the right one. The difficulty is usually not the model. It appears when AI has to be carried into an actual workflow, because nearly every process is full of undocumented exceptions — the informal workarounds staff rely on that never made it into the written procedure. Making AI useful means mapping what a company actually does against what the system can actually execute, line by line. That is a large amount of manual work, and it never looks impressive.
The author's analogy is self-driving cars, and it holds up better than most. Early autonomous vehicles learned the traffic code quickly and still could not merge into real highway traffic, because people do not drive by the rules alone — habit, courtesy and a long tail of small exceptions govern the road. So engineers sat in the car, over and over, watching what drivers really did, until the system could be taught the difference. Driving data by itself was not enough: the gap between the rules and the situations was wide enough that the engineers had to change the system for each rare case.
That job now has a name. FDEs — forward deployed engineers, who work inside the customer's business — have seen job postings grow 729% in a year. The role is essentially the solutions engineer adapted to machine learning; Palantir was among the first to formalise it, and OpenAI, Anthropic, Google, Salesforce and OMMAX are now hiring for it in bulk. The mandate is not to install a general-purpose product but to sit with a real process long enough to see what the org chart and the vendor contract do not record.
Here is what that 729% actually describes, and it is not a hiring trend. The industry that promised to automate the integration layer away is rebuilding it, by hand, and staffing it with its most expensive engineers. An ML engineer embedded at a client for months is not a software cost structure; it is a consulting cost structure, the exact thing enterprise software spent two decades claiming to make unnecessary. Model capability per dollar keeps falling. The cost of the person required to install that capability in one specific company is going the other way. Notably absent from every announcement quoted here is what an FDE costs, how many customers one can serve, or what gross margin looks like on a deployment that needs one.
The author expects the role to follow the driving curve, and there is a debate already running about how many FDEs a company needs relative to ordinary engineers. In the car, the trained engineer behind the wheel was not driving but watching, correcting and feeding the system what people really do. As the technology matured, the seat needed less engineering and more ordinary human judgement. The same framing came out of a conversation with Alex Fink of Swarmer about autonomous combat systems: the question is never AI versus no AI, it is what level of autonomy the system needs — true for business agents and for attack drones alike. If FDEs travel the same path, the job stops requiring an ML engineer brought in from outside and starts requiring domain expertise with AI training on top: a marketer who knows the tooling, a finance specialist who knows the tooling.
That trajectory is the optimistic reading, and it is an assumption, not a result. Whether FDEs or whoever replaces them can actually close the gap between what AI can do and what gets deployed is unknown. What the Altman and Meta week established is the negative case: you cannot assume the model will absorb the work and the org chart will rearrange itself around it. Meta is the strongest available evidence on that point, because Meta was not short of models, money or agents — it was short of a process the agents could be dropped into, and the person who cancelled the plan was the same one who could have forced it through.
The goal itself — lower cost, more channels, more products than manual work allows — stays reachable. The realistic route to it currently runs through a human being sitting inside one specific process, doing the configuration the platform will not do on the customer's behalf. Which means the near-term winners in AI deployment may not be the labs with the best models, but whoever can industrialise that human step and stop paying for it one customer at a time.