A field experiment with nearly 5,000 developers at Microsoft, Accenture, and an undisclosed Fortune 100 manufacturer measured Ai productivity.
Moreover, the most important finding isn’t about speed.
Over the years, AI coding assistants have claimed Ai productivity gains that were hard to verify. Moreover, a field experiment led by Microsoft Research, MIT, Princeton, and the Wharton School now provides real numbers. Ultimately, the finding is not about speed; it’s about what actually got shipped.
The study
Unlike lab studies that measure performance on a single scripted task, this design captured real production work over time. Additionally, it tracked actual pull requests and actual commits across teams and projects. Moreover, code shipped to actual codebases was included for analysis by researchers. This approach contrasts with controlled tasks and reflects day-to-day engineering work.
The headline result shows developers using Copilot completed 26.1% more pull requests weekly.
Additionally, they had a 13.6% rise in weekly commits, improving Ai productivity.
The number that matters more
Additionally, raw output is easy to inflate — more pull requests could just mean smaller, sloppier ones. Moreover, that’s why Accenture’s own internal tracking of the same rollout matters more than the top-line figure. Consequently, measuring 450 Copilot users against a 200-person control group, Accenture recorded an 84% increase in successful builds. In fact, that’s not a speed metric. Thus, it’s a quality one — evidence that AI-assisted code wasn’t just arriving faster, it was arriving in better shape.
Additionally, broader downstream effects were reported.
Pull request merge rates rose.
95% of surveyed developers said they enjoyed coding more with Copilot in the loop.
This satisfaction matters for retention, even if it never appears on a productivity dashboard.
The caveat: serious coverage should not be skipped.
Not every study points the same direction.
A 2025 METR analysis found experienced developers took 19% longer to complete coding tasks using Ai productivity tools.
They did this despite believing they had finished roughly 20% faster.
Now, the gap between felt and measured productivity is a closely watched thread.
It suggests gains from Microsoft/Accenture are real but not universal.
They show up most reliably in high-volume, structured work.
They are less reliable when developers are highly experienced and face unfamiliar or complex tasks.
The greater tension
Zoom out, and the individual-level gains sit awkwardly next to the organizational picture.
McKinsey’s 2026 survey of 1,719 executives across 97 countries found that 80% of individual workers report Ai productivity gains.
However, only 6% of organizations attribute a meaningful earnings impact to it.
The pattern McKinsey describes is: individual developers get faster, and individual teams ship more.
As a result, none of it reliably shows up in enterprise financials. Deliberate workflow redesign, training, and governance must be built around the tool—not just the tool itself.
The takeaway
The Copilot data settles a narrower question than the industry claims.
Moreover, does AI coding assistance measurably help developers ship more, better-tested code?
At Microsoft, Accenture, and at least one Fortune 100 manufacturer, the answer is yes.
The effect is large enough to show up cleanly in a controlled field experiment.
That bar is still not cleared by most workplace AI claims.
Whether that translates into an organization-wide productivity story is unsettled.
The honest answer right now is not yet, for most companies.
Key sources include Cui, Demirer, Jaffe, Musolff, Peng & Salz, and the MIT working paper. These cite GitHub Blog, “Research: Quantifying GitHub Copilot’s impact in the enterprise with Accenture”. Additionally, METR (2025) and McKinsey’s 2026 State of AI survey are cited.
Hi, this is a comment.
To get started with moderating, editing, and deleting comments, please visit the Comments screen in the dashboard.
Commenter avatars come from Gravatar.