The AI productivity gain is real. It is not reaching main.
I have published a free research report, Where the Gain Goes, on a question the industry keeps answering badly.
Ask whether AI has made software engineering faster and you get two confident, opposite answers. Both camps have data. Neither is lying.
The studies do not disagree. They count different things.
Measured per developer, the gain is real and larger than the sceptical reading allowed. Adopters across tens of thousands of engineers at Microsoft merged around 24% more pull requests, and the lift held for four months rather than decaying. A programme with 4,867 developers measured a 26.08% rise in completed tasks. Two years of telemetry across 22,000 developers reports task throughput per developer up 33.7%, and its authors call that real delivery acceleration.
Measured by what reaches production, the same period looks different. For the median team, feature-branch throughput rose 15% in a year while main-branch throughput fell 7% and main-branch success rates hit a five-year low of 70.8%, against a benchmark of 90% recommended by the people publishing the number.
More work than ever on the branch. Less of it landing, arriving less reliably.
Where it goes
Into verification, and the numbers are not subtle. Median time in code review up 441.5%. Code churn up 861%. The incidents-to-pull-request ratio up 242.7%. And 31.3% more pull requests merged with no review at all, human or agentic, which is what a review queue does when it cannot keep up.
The constraint is not that AI writes bad code. It is that the system for establishing whether code is safe to ship was built for a rate of arrival that no longer holds, and reading more of it does not scale.
What I did differently
Most commentary on this subject cites a number and moves on. This report publishes its working:
- The full evidence ledger is online. Every source, with its locator, method, population, quality tier, key limitation, commercial interest, and the verbatim sentence supporting the claim.
- Twelve of the 26 sources have a commercial interest in the conclusion. They are flagged on every figure that uses them, not buried in an appendix.
- Where the sources genuinely contradict each other, both readings are printed side by side in identical type, with my own reading marked separately as judgment rather than evidence.
- Every forecast carries the condition that would prove it wrong.
- The report has had no external review, and says so on page 2 rather than in the back matter.
That last one is the criticism I cannot answer. A report arguing that verification is what everyone under-invests in should not hide that it skipped its own.
Help settle what it could not
None of the 26 sources measured how much AI-generated code a developer actually reads before approving it, or how many organisations have written down what an agent must prove before it may act.
There is an eight-question survey. It is anonymous, with no email, no name and no account. Results are public and live, so nobody waits for an edition to see the findings and nobody has to take my summary on trust.
The report is free, CC BY 4.0, and the corrections page is here. If you find an error, it gets published with credit.
The ledger, the PDF, the survey and the errata are all listed together on the report’s Work page.