Microsoft is winding down most of the internal Claude Code licences in its Experiences and Devices division, the group that ships Windows, Microsoft 365, Outlook, Teams and Surface, with access ending on 30 June. The pilot started in December. It lasted six months. The official reason, per a Forbes summary of the company's internal messaging, is toolchain unification onto GitHub Copilot CLI, which Microsoft owns. The fact that 30 June is also the last day of Microsoft's fiscal year is the kind of coincidence that nobody in finance treats as a coincidence.
Uber, per a report Praveen Neppalli Naga gave to The Information in April, walked into the same wall earlier and harder. Its chief technology officer said the company had burned through its entire planned 2026 AI coding budget in four months. That was not because the tools failed. It was because they worked. Claude Code adoption inside Naga's roughly 5,000-engineer organisation jumped from 32 per cent to 84 per cent between February and March, with average monthly spend per engineer landing somewhere between $150 and $250 and heavy users hitting $2,000. Naga himself, by his own account, spent $1,200 on a single two-hour demo. The company was also running internal leaderboards ranking engineers by Claude Code activity, which means the incentive system was pointing in exactly the direction the budget could not afford.
The shape of the problem is what makes it interesting, and what makes it not a Microsoft or Uber story. Traditional developer-tool pricing is a flat seat licence. You buy a thousand seats, you write a fixed cheque, and individual usage above or below the plan is invisible. Agentic coding pricing inverts that. A long-running coding agent chewing through a large codebase consumes orders of magnitude more tokens than a developer accepting the occasional autocomplete suggestion. The bill scales with how hard the tool is worked, and the tool is worked hard precisely when the engineers love it.
A 2025 survey by the cost-governance firm Mavvrik, covering 372 enterprises, found that only 15 per cent of companies forecast AI costs within 10 per cent of actual. A majority miss by 11 to 25 per cent. Nearly one in four miss by more than half. Mavvrik sells tooling to solve exactly this problem, which is worth declaring, but the figures track too closely with the public Microsoft and Uber cases to dismiss. The firm's chief executive predicted the visible reckoning would arrive in the first half of 2026 as pilots flipped to production. It is the first half of 2026, and the reckoning has arrived.
The temptation will be to read Microsoft's pullback as "switch to the cheaper tool." That misses what is actually happening. GitHub is moving all of its 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 too much compute for a flat seat to absorb. Microsoft is not escaping consumption pricing. It is consolidating onto a vendor it owns, which lets it negotiate the internal economics and retire duplicate tools, while accepting that the underlying meter does not go away.
For everyone else, the lesson is more uncomfortable. Hard per-engineer caps are the obvious instinct, and the wrong one, because a cutoff in the middle of critical work trades a budget problem for a productivity problem. What actually works, judging by the firms that have not yet had to walk back their pilots, is a layered set of softer controls: team-level budgets with soft alerts, anomaly detection that flags runaway agent loops within hours rather than at quarter close, role-based gating on the most expensive frontier models, and chargeback that connects the cost to the team generating it. None of those controls work without a denominator, which means agentic coding has to be measured in cost per merged change or cost per resolved ticket, not just total spend. The companies that scale this 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. The companies that don't will become next quarter's case study.