Uber has capped what each employee can spend on each agentic coding tool at $1,500 a month. Microsoft questioned the cost of Claude Code licenses before cancelling their use in its Experiences and Devices division. Duolingo abandoned a plan to factor AI usage into employee performance reviews after staff objected to using the tools for the sake of using them. Three different companies, three different mechanisms, one shared problem: token consumption went up and nobody could show what it bought.
The behavior has a name now — token-maxing, meaning a sharp rise in token consumption without demonstrated return. It is arriving at exactly the moment the spending curve steepens.
$207 billion expected spend on AI agent software in 2026 $86.4 billion spend in 2025
Adoption rates and token burn are still treated as inputs to the question of whether AI is helping, not answers to it. You can get nearly every engineer in the company using AI and still be unable to demonstrate that the money produced more shipped features, more fixed bugs or more resolved customer problems.
Noe Ramos, VP of AI operations at Agiloft — a platform for enterprise contract lifecycle management — thinks "token-maxing" points at the wrong culprit. Teams are not spending out of a love of waste, he says; the default infrastructure pushes them into it. At most companies model selection is left to whoever writes the prompt, so a frontier model gets pulled in for work an inexpensive open model would have handled. He calls that an infrastructure problem, not an employee behavior problem.
Dmytro Palaniychuk, head of engineering at the language company Promova, blames the settings shipped with most AI tools. On enterprise and team plans, premium models are typically preselected, high reasoning depth is on by default, and on Promova's plan Opus sessions automatically get a 1 million token context window. Employees rarely change any of it. After a few months Palaniychuk discovered the premium model was handling tasks as trivial as checking email.
Promova is aiming to split its load across Opus, Sonnet and Haiku at 40/50/10. Palaniychuk calls that a reference point rather than a destination — the ratio is not the real question. The real problem, he says, is habit: even if you make Sonnet the organization-wide default, employees without a working sense of which model fits which situation will act on inertia. They have to pick deliberately and re-check the choice before each new conversation with a model.
Rick Spencer, general manager for technology and product at SUSE, argues the first reaction should not be "use less AI." High token burn can mean several different things. If an employee spent $16,000 on tokens and saved the company $100,000, he says, that usage should be encouraged. Start with diagnosis, not prohibition.
SUSE sorts AI use into three buckets managers handle differently: daily work, meaning employees' routine tasks; autonomous agents, systems that carry out tasks by themselves; and jumping the curve, one-off strategic projects. For an autonomous agent the best choice is not necessarily the newest frontier model. The company is trying to connect usage to outcomes rather than suppress it. There is no automatic routing layer today — developers decide for themselves — and a proxy layer is on the roadmap.
Everlaw, which builds AI tools for litigation and investigations, sets per-person token limits the way Uber does. But CTO Max Christoff treats them as a feedback mechanism rather than rationing: when an engineer hits the ceiling, the tooling team gets a short email, and the limit is usually doubled the same day.
Everlaw was also the only company in the story willing to put numbers on returns. One core Java infrastructure project consumed $3,500 in tokens and cut delivery from 9.5 to 2.5 engineer-months. On another, still unreleased product, token spend reached $27,000 and will probably hit $40,000 — against an estimated development time that dropped from 90–100 engineer-months to 19.
Agents do not handle everything. Everlaw spent thousands of dollars migrating front-end code from Dojo to React and then threw the result away: the two frameworks rest on different assumptions about how state relates to rendering. The team now asks an agent to describe the old system's behavior first, then builds the new one from that documentation.
Agiloft went the other direction and dropped limits entirely. Ramos says 74% of employees never came close to the old ceilings; the caps obstructed heavy users without touching the actual cause of the spend. The company set cheaper default models instead, reserving frontier models for work that genuinely needs them, with routing handled at the infrastructure layer rather than in the prompt, plus caching. The fastest improvement available to most enterprise teams, in his view, is a single LLM gateway with inexpensive defaults and smart routing. Don't ration the tool, fix the architecture underneath it: managing scarcity papers over the problem, intelligent routing removes its cause.
A lot of vendors agree, including Merge, Databricks, AWS Bedrock and Azure AI Foundry, all of which now sell automatic routers that score task difficulty, pick a model and weigh its cost at the same time. Databricks introduced Smart Routing inside its Unity AI Gateway at the Data + AI Summit in June, alongside hard spending limits and cost allocation across hosted models, coding agents and user-facing agents. It is in beta and runs in two modes: recommendations only, or automatic routing.
David Nasi, director of product management at Databricks, says the router scores each request using deterministic signals plus model-based classification, weighing prompt intent and length, the files a request references, stack traces, the size of the change, the reasoning depth required and execution complexity. It also routes on more than the model: Databricks found the same model performs differently depending on the harness it is used through. Decisions are made transparently at runtime rather than inside a black box. When spend hits the budget ceiling, an administrator picks between a hard block and a downgrade to a cheaper model that still meets the task's requirements. Agentic workloads are harder — a single task can spawn dozens of model calls with no human approving each one — so the router evaluates those at execution-step boundaries instead of per call.
Here is what the whole discussion keeps sliding past: routing is a cost intervention, not an ROI intervention. Every fix described above makes the denominator smaller. None of them measures the numerator. A company that cuts its token bill 40% by sending boilerplate to Haiku has learned how to buy the same unproven output more cheaply, which is progress in procurement and no progress at all on whether the output was worth buying. It is also worth noticing who is selling the remedy: Databricks, AWS and Azure make money on the inference these routers are trimming, and a cost-control feature that keeps a nervous CFO from cutting the whole line item is, from the vendor's side, a retention product.
The ROI gap holds under both manual selection and automatic routing — almost nobody can calculate the savings precisely. Promova never named the spend difference before and after its changes, because the rollout coincided with other work; telemetry did show that Opus with the 1 million token context window accounted for roughly a third of monthly spend. SUSE, with its three categories and its planned proxy layer, has no exact metrics either, only individual examples: one project cut CVEs in dependencies from several hundred to zero, and since May the company's agents have classified nearly 10,000 CVEs in the VEX database.
Databricks can at least see which way behavior is moving. Nasi says companies that used to route everything to frontier models have started shifting ordinary work — boilerplate generation, simple bug fixes, small changes — to cheaper ones, with no noticeable drop in task success rates. To let teams see the effect and catch quality regressions after a route change, Unity AI Gateway ships with unified tracing and LLM-as-judge evaluation frameworks, along with evaluation datasets, trace analytics and automatic feedback loops.
Domain-specific accuracy checking stays the customer's job. Palaniychuk calls that the hard part: companies are trying to apply deterministic checks to non-deterministic output, and new model versions land roughly quarterly, so an evaluation tuned to one model does not carry over to the next unchanged.
Everlaw's approach is to offer engineers a wide model selection and a budget in dollars, with choice available after a security review — on the theory that the person reviewing the code is fastest to recognize which model produced output worth keeping. Even so, a model can pass every test and still leave behind code that cannot be maintained. Christoff notes that a coding agent can propose twenty surface-level fixes instead of addressing the common cause.
Notice, too, how the one company with hard ROI numbers arrived at them. Everlaw's figures — 9.5 engineer-months down to 2.5, 90–100 down to 19 — are comparisons against estimates of work that was never done the old way. That is the least falsifiable measurement in software, and the fact that it is the best evidence on offer in an industry heading toward $207 billion in agent software spend says more about the state of the discipline than any of the routing architectures do.
Palaniychuk and Christoff both expect today's specific optimization techniques to age out, while the discipline of working with AI survives them. Palaniychuk does not think Promova's exact model ratio will matter for long; matching the tool to the task will, and more options make that choice more important, not less. Christoff expects companies to start planning around token spend directly: within one to two years, he predicts, tokens will be treated more like an annual headcount or production cost than like ordinary IT and software spend, with division heads arriving at annual planning carrying separate numbers for people and for tokens, each with its own business case. One team might budget several million dollars in tokens and barely add staff. Another will do the reverse.
Executives disagree on where to start. Ramos says build a reliable routing layer first. Spencer starts with training and mentoring for managers. Christoff's first step is understanding what outcome the company is chasing. Palaniychuk watches how his team actually uses the tools and where optimization is needed. His advice is to hand the team a tool and a visible experimentation budget across models and providers, then spend a short period collecting real usage data through telemetry and feedback, rather than deciding the correct ratio in advance — your own team's statistics should produce it. The people who show the most discipline, he notes, are not necessarily the most technical ones.
Which points at the uncomfortable shape of the next budget cycle. If Christoff is right that tokens get planned like headcount, then someone will have to defend a token line the way they defend a hire — and unlike a hire, a token line has no performance review, no manager and, so far, no agreed way to tell a productive one from a wasteful one.