Microsoft Claude Code Cuts Expose an AI Spending Governance Gap

What happened

Microsoft is pulling most internal Claude Code licenses from its Experiences + Devices division. That group builds Windows, Microsoft 365, Outlook, Teams and Surface. Engineers have been told to move their workflows to GitHub Copilot CLI by June 30, which is also the last day of Microsoft's fiscal year. 3 The directive reportedly came from Rajesh Jha, the executive vice president who runs the division. Microsoft had only opened Claude Code to the group in December 2025, when it invited thousands of developers, program managers and designers to use it. 3 The rollout lasted about six months before it was reversed.

One report says Microsoft is also cutting per-employee AI budgets by 90%. The same outlet says Meta has roughly halved its Claude Code users, from a peak of about 60,000 engineers to around 30,000. 1 That outlet dates its reporting to early October, while others place Microsoft's move in May 2026. The timelines don't line up neatly, so the Meta figures and the budget-cut percentage should be treated with some caution until they are confirmed elsewhere.

Lock-in, or the bill coming due?

The simplest explanation is competitive. Microsoft owns GitHub Copilot and works closely with OpenAI. Paying a rival, reportedly on the order of $1 billion, to equip its own engineers is hard to defend when it has a comparable product in-house. 1 One analysis frames the cut as part of a wider lock-in story. It reports that Microsoft cut off enterprise Claude Code licenses to push customers back onto Copilot. 4 Other accounts describe the cancellations as internal. That difference matters, because steering your own staff is a very different act from steering paying customers.

The lock-in story also runs into a contradiction. Microsoft has been adding rival models, not shutting them out. GitHub Copilot's model picker runs Claude, GPT-5.x and Gemini 3.1 in one interface. Microsoft 365 Copilot turned on Claude by default alongside GPT-5 and Gemini in January 2026. 4 So Microsoft is not rejecting Anthropic's models. It is rejecting Anthropic's standalone tool, where it doesn't control routing, metering or the customer relationship. 4

The cost problem behind it

The broader context points less toward strategy and more toward money going out with nothing in place to control it. One report describes a single enterprise client that burned through $500 million in one month. It had put no usage controls on employee AI licenses. 2 Uber reportedly used up its entire 2026 budget for AI tools, spent on Claude Code and Cursor, in four months. 3 Anthropic's Opus 4.8 arrived in the middle of all this. 3

The same reporting lists four friction points that keep enterprise AI from paying off [2]:

• strategy that doesn't match how the tools are used
• costs that spiral from routine queries
• slow adoption
• reluctance to feed proprietary data into the tools

It argues that adoption metrics alone can no longer justify these deployments. If even Microsoft, the company most committed to embedding AI everywhere, found Claude Code costs unjustifiable without tighter controls, other buyers are likely to audit their own spend. 2

One account also mentions that Microsoft's legal team paused some Claude access. It hints at legal exposure that could shape how AI companies hire and build. 1 The details in that account are incomplete.

Reading the signal

Both explanations are probably partly right, but governance is the more important one. Lock-in explains why Microsoft chose Copilot as the destination. It doesn't explain why the move happened now, or why it was timed to the fiscal year-end. The more convincing story is a company facing the same problem as Uber and the unnamed $500 million client. Agentic coding tools bill by usage, and developers who are encouraged to experiment will use them heavily.

Performance-based pricing works well when the tool succeeds. It becomes a problem when nobody has set limits. A 90% cut to per-employee AI budgets, if accurate, is not a fine-tuning adjustment. 1 It is an emergency brake.

For Anthropic, the risk is more about where its product sits than how good it is. Its models stay embedded in Microsoft's products, but as one choice in someone else's model picker. 4 That brings lower margins and less direct contact with developers.

For enterprise buyers, the lesson is practical. The decision is no longer which assistant writes the best code. It is whether usage caps, cost attribution and routing controls are in place before thousands of engineers get a usage-billed tool. Platforms that bundle those controls, as Copilot's multi-model approach does, have an advantage that has little to do with model quality. 4

The 2026 pullback doesn't show that AI coding tools failed. It shows that many organizations adopted them faster than they put spending controls in place, and are now cutting back to bring costs under control.