Microsoft pulled most Claude Code licenses inside Experiences and Devices, effective June 30, 2026. That’s the division that builds Windows, Microsoft 365, Outlook, Teams, and Surface hardware, one of the highest-volume engineering organizations at the company. The tool had been live for six months. Engineers got access in December 2025.

The official line is toolchain unification, pushing teams toward GitHub Copilot CLI, Microsoft’s own product. Rajesh Jha, the executive who runs Experiences and Devices, was more direct: Claude Code had become “perhaps a little too popular” among his engineers. Popular enough that token spend blew through the annual AI budget faster than anyone planned for.

That same week, Microsoft announced Frontier Company: $2.5 billion and 6,000 people dedicated to deploying AI inside other companies’ operations, with London Stock Exchange Group, Unilever, Land O’Lakes, and Accenture as launch partners. Notice which story broke first. The same company building a $2.5 billion business to teach Unilever how to run AI with discipline didn’t have that discipline over the tool its own engineers had already voted for with their usage.

Six months, one off switch

What’s missing from the public record says as much as what’s in it. Microsoft hasn’t published a dollar figure, an engineer count, or the specific budget ceiling that broke. What Jha confirmed is the shape of the story: adoption was real, and cost outran it. June 30 isn’t a random cutoff either. It’s the close of Microsoft’s fiscal year, the point where any spend line that ran hot has to get resolved before the new book opens.

That detail matters more than it looks. This wasn’t a quality call. Nobody on record said Claude Code produced worse code or broke in production. If anything, the public signal runs the other way: engineers adopted it with more enthusiasm than Microsoft’s own internal tooling had managed to earn in its own hallways. When the tool your engineers prefer and the tool your own company bills through are two different products, and spend spikes, there’s a structural tension between what serves the engineer and what serves the balance sheet. Microsoft resolved that tension in favor of the balance sheet, right as the fiscal year closed.

GitHub Copilot CLI, the tool teams are now pushed toward, is the option Microsoft controls end to end: the invoice, the contract, and the margin all stay in-house. That doesn’t make it the worse tool. But it does change what question Microsoft was actually answering when it pulled the licenses, and that question wasn’t “which tool do our engineers prefer.”

Uber ate the overrun. Microsoft cut the cord.

I covered Uber’s version of this in April, when CTO Praveen Neppalli Naga admitted Claude Code had torched Uber’s full annual AI budget in the first months of the year. Same root cause: adoption outran the plan, token billing replaced seat pricing, and an annual budget cycle wasn’t built to track exponential consumption. Same surprise landing on finance’s desk.

The difference is what happened next. Uber couldn’t cut the tool. 11% of its pull requests were already opened by agents, and more than 90% of its diffs ran through AI review before shipping. The engineering flow had absorbed the tool, so Uber chose to govern consumption instead of switching it off. Microsoft, running a division of comparable size and, after six months of real adoption, probably comparable dependency, took the other path: kill access and move people to an in-house product, even if it wasn’t the one they wanted.

Neither call is automatically the right one. What both companies share is the failure that preceded the decision. Neither Uber nor Microsoft had spend instrumented before the bill arrived. Both found out when it was already too late to plan, not while there was still room to choose.

The problem is the meter, not Microsoft

When a company buys software by the seat, budgeting is a straightforward multiplication: price per license, times headcount, times twelve months. That number gets set once a year and doesn’t move on its own. It’s a purchase decision, made once, with a ceiling known in advance.

Token billing doesn’t work that way. Every agent run, every sub-agent a task spawns, every automated review cycle draws down budget in real time, with no one approving that specific execution. Gartner measured agentic workflows consuming 5 to 30 times more tokens per task than a standard chatbot. That’s not a pricing difference. It’s a difference in kind: a tool decision that used to get made once a year became a spend exposure that runs every single day, whether or not finance is watching it.

Most enterprise AI procurement and spend-control processes are still built for the world before this: approve a license, set an annual budget, review it at the next quarterly meeting. That process has no way to catch, in week three, that week-twelve consumption is about to triple the forecast. Microsoft, with its cloud infrastructure, its deepest-in-the-industry relationship with OpenAI, and billions committed to compute, didn’t have that instrument installed inside its own walls. If Microsoft didn’t have it, most companies deploying Claude Code, Copilot, or any other agent today don’t have it either.

What IQ Source does with this

It’s not a coincidence this happened at the same company that just announced Frontier Company. Selling deployment discipline to Unilever or London Stock Exchange Group is a services business, priced against the client’s outcome. Measuring consumption across your own engineering fleet is an internal instrumentation problem, and the two don’t get solved by the same team or the same budget. Microsoft built the first before it built the second.

When we open an AI Maestro engagement, the first question we answer isn’t which model to run or which agent to install. It’s what sustaining it in production will actually cost, measured per workflow, not per seat or per license. The Process Reality Map we build in the first weeks carries that number, because it’s the number that decides whether the project clears the Go/No-Go gate before real budget gets committed. We’d rather a company learn its spend ceiling in week three of discovery than in the July board meeting, after the fiscal year has already closed and someone has to explain a budget that broke without anyone seeing it coming.

Talk to us about AI Maestro discovery before your budget surprises you