Microsoft just cancelled its own Claude Code licenses. What is your May board's productivity number now?

The setup

On May 14, Tom Warren at The Verge reported that Microsoft is cancelling most of its internal Claude Code licenses and pushing thousands of developers in its Experiences and Devices group, the team behind Windows, Microsoft 365, Outlook, Teams and Surface, to GitHub Copilot CLI by June 30. That date is not random. It is the last day of Microsoft’s fiscal year.

Six months ago Microsoft was the most enthusiastic Claude Code customer on the internet. In December it gave thousands of engineers access. By spring it had become so popular internally that it was, in the words of one of the follow-up writeups from TechBuzz the same day, “undercutting internal usage of Copilot CLI.” Microsoft loved Claude Code so much it decided it had to stop loving Claude Code.

That is the event. Now the question for any exec sitting down to write a May board update. What is the productivity number that goes on the slide?

What they tried

Microsoft’s six-month Claude Code rollout was not a pilot. It was a real production deployment across one of the largest engineering organizations on earth. The Experiences and Devices group is not a research lab. It ships the calendar app on a billion phones.

Through that deployment Microsoft learned three things that matter for the rest of us, because Microsoft has more telemetry on engineering productivity than most boards will ever see.

First, developers picked Claude Code on its own merits when given the choice. That is a real productivity signal, not a marketing line. According to the internal memo attributed to Rajesh Jha, the EVP of Experiences and Devices, reconstructed in the Harianbasis writeup on May 15: “Claude Code was an important part of that learning… at the same time, Copilot CLI has given us something especially important: a product we can help shape directly with GitHub for Microsoft’s repos, workflows, security expectations, and engineering needs.”

Read that twice. The reason given for the switch is not that Claude Code was slower or worse. The reason is that Microsoft wants to own the harness it can govern. The phrasing is careful. Anthropic’s models stay accessible through Copilot CLI. What goes away is the parallel commercial license, the parallel telemetry, and the parallel line item.

Second, Microsoft converged on a date that lines up with budget cycles, not with quarterly releases. Three of the four writeups across The Verge, TechBuzz, Harianbasis and Let’s Data Science name June 30 specifically. When a Fortune 50 picks a fiscal-year-end as a harness cutover deadline, it is telling its CFO and not its CTO.

Third, the same week this story broke, Jellyfish released its 2026 State of Engineering Management survey on May 12, sampling more than 600 engineering leaders.

Eighty percent in that same survey say AI improves team productivity. Seventy-five percent say AI frees up time for higher-value work. Jellyfish closes with the line: “AI is no longer optional, and the organizations that are embracing it most aggressively are already pulling ahead.”

So in the same five-day window we have a vendor-leaning survey saying AI is now mandatory and a Microsoft procurement decision saying the harness layer is now consolidatable. Both are real signals. Neither is a board number.

Where it broke and where it worked

Here is what I keep returning to. The Jellyfish 64% number is genuinely impressive. It is also the kind of number that does not survive a serious CFO question.

A 25% increase in what? Measured against what baseline? Reported by whom? Compared with which control? Pulled from which system?

The GitKraken DORA velocity guide on May 13 walks straight into the same paradox. The 2025 DORA report it summarizes shows that developers using AI complete 21% more individual tasks and merge 98% more pull requests. And yet, as Faros AI’s research dataset surfaces in the same piece, organizational delivery metrics stay flat. The exact quote DORA pulled into its 2025 report, also surfaced in that same May 13 GitKraken writeup, is hard to walk past: “AI adoption has a negative relationship with software delivery stability.” Stack Overflow’s own 2025 number that the GitKraken piece cites is just as awkward. Sixty-six percent of developers called “AI solutions that are almost right, but not quite” their biggest frustration.

This is the productivity paradox in one image. The author of the PR ships faster. The reviewer audits longer. The organization moves at roughly the same pace, but spends more on harness seats, more on tokens, and more on review hours. The board sees a “+64% velocity” survey on Monday and a “we cancelled Claude Code to save money” memo on Wednesday. Both can be true. They are both true.

What Microsoft seems to have concluded is that when two harnesses run on the same engineer, the right denominator for productivity is not gross velocity. It is delivered shipping per total harness spend per engineer.

GitHub Copilot CLI v1.0.48, which shipped May 14 the same day as the cancel story, made a small change that lands directly on this point. The model picker now shows actual token prices instead of generic dot indicators. The dollar cost of each model selection is now visible in the terminal at the moment a developer picks a model. That is what owning the harness means. Microsoft put the price tag in front of the person spending the money.

The pattern

I have been sitting with this for two days and the pattern that keeps surfacing is this. The productivity question on a May 2026 board update is no longer “which harness is winning?” It is “are we paying twice for the same code?”

When the same Fortune 50 that owns the most aggressive internal AI rollout in the industry decides at fiscal year-end that paying for two parallel command-line agents is indefensible, every other board hears the same question land on its own desk by Q3.

The number that survives that conversation has three properties. It is measurable from systems the company already owns, not from a vendor dashboard. It has both a shipping numerator and a spend denominator. And it can be re-pulled in any month without depending on a survey.

The pair I keep coming back to is the same one a CFO would write on a napkin. Merges per active engineer per week, from the CI/CD log, at loaded engineer cost. Harness seats plus harness token spend per active engineer per month, from the procurement and metering systems. Same engineer, same week, two numbers on one slide.

If those two numbers are moving in opposite directions over a rolling quarter, that is the Microsoft problem. If they are moving in the same direction, you have the budget defense.

The Jellyfish 64% line belongs in the appendix, not on the slide. It is directional, useful, and not CFO-verifiable. The Microsoft decision belongs on the slide because it has a date attached and a named EVP.

What I would tell you over coffee

Three things, said the way I would actually say them in a coffee shop in Barcelona.

First, do not waste a board cycle trying to be definitive about which harness is “best” right now. Microsoft just spent six months trying to do that with telemetry most companies will never have, and the answer they landed on was governance, not preference. Pick one governance plane for your fiscal year and let the model choice live underneath it.

Second, if your current productivity slide has a vendor logo on it, replace the logo with a same-engineer ratio. The ratio is harder to game than a vendor adoption percentage and easier to recompute next quarter. The Microsoft decision was made on the ratio side of that slide, not on the logo side.

Third, name the date by which the team will pick its single harness control plane. Microsoft picked June 30 because that is when their fiscal year ends. A company fiscal year-end is the right date for the same reason. It gives the team a deadline that is not negotiable and a budget conversation that already has a room on the calendar.

One last thing, mostly to lower the heart rate. Nothing about Microsoft cancelling internal Claude Code licenses says Claude Code stopped working. The models are still going through the Copilot CLI pipe. The harness was the dollar conversation. That is a calmer story than the headlines, and it is the one you can actually take into a board meeting.