When a company rolls out a new AI tool, the instinct is painfully corporate: licenses, training, documentation, management announcement. A study of Claude Code and GitHub Copilot CLI inside Microsoft tells a messier story.
The authors studied tens of thousands of engineers during the first months of 2026. The headline result is roughly 24% more merged pull requests among adopters than the model estimates they would otherwise have produced.
Interesting. It is not "24% better software."
The researchers make that limitation explicit. A merged PR is a proxy for output. It does not measure the value of the code, its difficulty or its long-term quality.
Adoption looked more like diffusion than deployment
Another finding is easier to reuse in a small team.
First use spread primarily through internal social networks. In practical terms, seeing nearby coworkers actually use the tool mattered to whether people tried it.
Retention was also associated more strongly with coding activity than with the demographic characteristics studied. People with enough real coding work to hand to the tool had more reason to keep coming back.
That shifts the question of how teams adopt agents.
Perfect documentation cannot show which tasks are worth delegating. A coworker taking a real bug, starting an agent, stepping in when it wanders and then showing the diff can.
The practice becomes visible. Other people can judge the cost, failures and payoff on work they already understand.
Put brakes on the 24% number
This study covers one unusually large organization, Microsoft, with its own tools, repositories and engineering practices. The observation window is four months. Its main output metric is still merged pull requests.
It does not justify promising another team a 24% productivity increase after installing a CLI agent. That claim would be much stronger than the evidence.
The narrower result is useful enough: in this large rollout, usage did not vanish after the novelty period, and weeks with agent use were associated with more merged PRs. The paper reports that the lift persisted across the window it studied.
For a small team, the diffusion result is what I would steal. Do not start by telling everybody to become "AI-first," a phrase almost engineered to make people close Slack.
Take two or three real tasks instead. Use the agent where coworkers can see the process. Leave the failures in the demo. Show what still needed review.
If the tool is useful, the practice can spread. If it is not, you saved everyone a mandatory training session and probably several slides.