The budget that success blew up
Uber's chief technology officer, Praveen Neppalli Naga, gave The Information the number that should stop every owner mid-sentence: the company had burned through its entire planned 2026 AI coding budget in four months. Uber had rolled the tool out to roughly 5,000 engineers in December 2025, and adoption climbed from 32 percent of engineers in February to 84 percent classified as heavy agentic users by March 2026. By spring, reporting put monthly cost per engineer between 150 and 250 dollars on average and between 500 and 2,000 dollars for power users. Nothing had gone wrong. The tool worked, engineers used it, and using it is what emptied the budget.
The same pattern surfaced at Microsoft. Multiple reports, led by The Verge, describe Microsoft cutting most internal Claude Code access in its Experiences and Devices division - the group behind Windows, Microsoft 365, Outlook, Teams and Surface - by 30 June 2026, moving engineers to GitHub Copilot CLI. The official framing is toolchain unification, but the reporting is blunt about the driver: at consumption pricing, heavy use of an agentic coding tool blows past the annual budget, because the tool is metered by tokens consumed, not by seats occupied. The better it works, the more it is used, and the more it is used, the higher the bill.
Why it matters: this is a base-rate error, not a tooling problem
Why it matters: the mistake here is not choosing the wrong tool. It is budgeting a new tool on the old base rate. Companies priced AI coding as if it were cheap autocomplete - a small per-seat cost that barely moves with usage. Agentic tools break that assumption because they run continuous reasoning cycles, and they are priced for it: cost scales with how much work the tool does, which is to say with how successful the rollout is. An owner who budgets the pilot - a handful of curious engineers running light queries - is modelling the wrong base rate for a tool whose whole value proposition is that everyone will use it constantly.
Yes, but: this is not an argument that agentic AI is too expensive to use. Uber still deploys it, and the productivity is real - by spring, roughly 70 percent of Uber's committed code originated from AI tools. The point is narrower: consumption pricing turns adoption into a cost curve, and the success case, not the pilot, is the one that has to be affordable. Microsoft's response, moving to a different tool, is one answer; metering and per-user caps from day one is another, and usually the cheaper one.
The bottom line: budget the fully-adopted case and meter from day one
The bottom line: before you buy a consumption-priced AI tool, model the cost of the scenario where it works. Take the per-user cost of your heaviest expected user, multiply by the headcount you actually intend to roll out to, and check whether that number fits the annual budget. If it does not, the tool is unaffordable at the success case regardless of how cheap the pilot looked, and the disciplined move is to cap and meter per user from the first day rather than discover the ceiling in month four the way Uber did.
The wider decision principle outlives this one product category. When a vendor prices per unit of the exact behaviour you are paying them to increase, the pilot cost is meaningless and the success cost is everything. Model the outcome you are hoping for, not the trial you are running, and if you cannot afford the tool working perfectly, do not buy the tool that charges you more for working.
Read next: The Man Spending 145 Billion Says It Is Not Working Yet | Google Runs at 9% Overhead. Power Use Rose 37%
