AI made developers faster. It did not make organisations faster. This is where the difference went.
Download the report (PDF, 8.7 MB) Take the surveyMeasured per developer, the AI productivity gain is real and better evidenced than sceptics allowed. Measured by what reaches production, it thins out: for the median team, work on the branch rose 15% in a year while what landed on main fell 7% and main-branch success rates hit a five-year low. The studies do not disagree. They count different things, and the difference is being spent on verification that nobody is budgeting for.
This is an argument, not a study. The report itself runs no survey and collects no primary data. It is a reading of 26 sources, every one checked against its origin, plus my own view where the evidence runs out. 12 of the sources are authored by organisations that sell AI development or verification tooling, and are flagged as such throughout. A separate survey, below, tests its predictions for the next edition.
| F1 | Agents became the default before trust did. |
|---|---|
| F2 | The gain is real per developer and does not reach main. |
| F3 | There is no AI productivity number. |
| F4 | Review is where the gain goes. |
| F5 | Verification has to be manufactured, not performed. |
| F6 | Authority belongs to evidence, not to capability. |
The report makes five predictions, each with the condition that would prove it wrong. The survey exists to test exactly those conditions. The next edition reports which of them survived.
Help settle them