BudgetPixabayMicrosoft is winding down most internal Claude Code licences across its Experiences and Devices division, the team behind Windows, Microsoft 365, Outlook, Teams and Surface, with access ending 30 June. The pilot launched only in December. Six months in, a company whose chief executive has said AI writes roughly 20% to 30% of the code in some of its repositories is reportedly standardising developers on GitHub Copilot CLI. The stated rationale is toolchain unification, though reporting also points to cost control and the fact that 30 June is the last day of Microsoft's fiscal year.Uber walked into the same wall earlier and harder. Chief Technology Officer Praveen Neppalli Naga told The Information in April that the ride-hailing company had burned through its entire planned 2026 AI coding budget in four months. His words, as reported, were blunt. The budget he thought he would need was already gone.ForbesUber Burns Its 2026 AI Budget In Four Months On Claude CodeBy Janakiram MSVTwo large, AI-forward engineering organisations just hit a budget failure mode that no finance team modelled. That is the part worth a CXO's attention. The tools did not break. They worked well enough that engineers would not stop using them, and consumption-based pricing converts "loved by the team" directly into "over budget by the second quarter".How A Loved Tool Becomes A Runaway CostAgentic coding tools like Claude Code are not priced like the software most procurement teams know. A traditional seat licence is a fixed number. You buy a thousand seats, you know the annual cost, and usage above or below the plan changes nothing. Token-based billing inverts that. The bill tracks how hard the tool is worked, and an agent running for hours against a large codebase consumes far more than a developer accepting the occasional autocomplete.MORE FOR YOUUber's reported numbers show how fast that compounds. According to The Information, adoption of Claude Code jumped from 32% to 84% of Naga's roughly 5,000-engineer organisation between February and March. AI usage now runs deep into the engineering workflow, with 95% of engineers using AI tools monthly. Reported figures suggest a large share of committed code is now AI-assisted, though the definitions vary and a metric for code that is fully agent-generated with no human in the loop sits much lower. Average monthly spend per engineer was reported between $150 and $250, with heavy users said to reach $2,000, and Naga reportedly spent $1,200 in a single two-hour demo.Uber did not stumble into this passively. The company encouraged adoption and ran internal leaderboards ranking engineers by Claude Code activity. Every incentive pointed toward more usage, and on token pricing more usage is more cost. The leaderboard worked. The budget did not survive it.A Forecasting Problem, Not An Isolated OneIt would be comfortable to treat Microsoft and Uber as outliers, two companies that simply moved too fast. Survey data argues otherwise. A 2025 study from cost governance firm Mavvrik and Benchmarkit, covering 372 enterprises, found that only 15% of companies forecast AI costs within 10% of actual. A majority miss by 11% to 25%, and nearly one in four miss by more than 50%. Mavvrik sells AI cost-governance tooling, so its survey is relevant evidence rather than neutral academic data, but the figures track closely with the pattern Microsoft and Uber have now made visible. The firm's chief executive predicted the real reckoning would arrive in the first half of 2026 as pilots flipped to production.The mechanism is consistent across all of them. Flat licences kept token spend invisible because the price did not move with usage. The moment a tool is billed by consumption, every prompt, every long agent session and every large context window shows up on an itemised invoice, and the total becomes both visible and hard to predict. Quarterly engineering output and quarterly AI cost are now the same curve. Microsoft and Uber are among the first high-profile cases to make that cost-governance problem visible.The Migration Target Is Moving TooMicrosoft's chosen exit is GitHub Copilot CLI, a tool it owns outright. Owning the vendor lets Microsoft negotiate internal economics, retire duplicate tools and standardise controls in ways an external supplier would never allow. There is a real saving available there, and it has little to do with the per-token price of either tool.What ownership does not do is repeal consumption economics. GitHub is moving all Copilot plans to usage-based billing on 1 June, replacing premium request units with token-linked AI credits priced at one cent each. GitHub's own explanation is candid. Copilot is no longer the lightweight autocomplete it once was, and agentic workflows consume far more compute than a flat seat can absorb. A CXO reading the Microsoft retreat as "switch to the cheaper tool" has misread it. Vendor consolidation and unit-cost reduction are separate questions, and as analysis of GitHub's shift suggests, token economics still follow the workload whichever vendor sends the invoice.What Belongs In Place Before The Next RolloutThe fix is not slower adoption. Engineers who lean this heavily on AI are not giving it up, and a CFO who tries to claw it back is fighting productivity. The fix is treating agentic coding spend as a metered utility rather than a software subscription, and a mature control model deserves a place before the next deployment.Blunt per-engineer ceilings are a starting point but a poor finish, because a hard cutoff in the middle of critical work trades one problem for another. A better model layers softer instruments. Team-level budgets with soft alerts catch drift before it becomes an overrun. Anomaly detection flags a runaway agent loop in hours rather than at quarter close. Role-based model access keeps expensive frontier models out of routine work, and approval thresholds gate the long autonomous runs that drive the largest bills. Showback or chargeback ties the cost back to the team that generated it, which changes behaviour faster than any memo.None of that works without a denominator. Spend figures alone tell a board nothing about whether the money was well used, so the controls should sit alongside ROI metrics such as cost per merged change, cost per accepted suggestion or cost per resolved ticket. A renegotiated contract structure matters too, whether committed-use discounts, spend caps or hybrid pricing, because a pure pay-as-you-go model leaves the enterprise carrying all the volatility.For boards, the harder lesson is cultural. Microsoft and Uber show that adoption velocity and cost exposure have become a single number, and an engineering culture optimised purely for AI usage is also optimising for an unbounded bill. The companies that scale agentic coding without a Q2 surprise will be the ones that put a meter on the tool before handing it to the team that loves it most.