AI sped up coding and moved the bottleneck to code review: what the Harvard study says
Data from 655 firms and 720,000 workers: AI agents raised lines of code by 44% while finished work rose only 13.7%. The gap piles up in code review, where time to approval grew 56%.
AI made writing code faster. Do software teams actually ship more? A Harvard working paper by Fiona Chen and James Stratton, "Artificial Intelligence in the Firm: Bottlenecks in Software Production", looks at this with the largest dataset so far: 655 firms, 720,542 workers and 445 million work events (GitHub activity, Jira issues, calendar events). In short: code went up a lot, shipped work went up a little, and the gap piled up in code review.
More code, not the same amount of shipped work
The paper measures two generations of tools separately. AI assistants raise lines of code by 30.4%, commits by 16.8% and pull requests by 20.9%. AI agents add a further 44.0% in lines of code, 36.0% in commits and 35.7% in pull requests opened.
Work that reaches production tells a different story. The effect of assistants on resolved Jira issues is near zero (−1.5%). For agents it is 13.7%, which the authors call borderline significant. Code grows 44% while finished work grows by about a third of that at best. Issues did not get bigger either; the authors checked issue descriptions to rule that out.
Where the gap goes: code review
After agent adoption, the time from review submission to approval rises by 56.1%. It has two parts: code waits 59.9% longer to reach a reviewer, and active review takes 58.5% longer. Comments per change rise by 38.6%.
The first part is a queue: more code for the same reviewers. The second matters more: each review takes more effort, so the incoming code is harder to get to an approvable state. Reviews mostly go to senior engineers who know the code, so the load lands on the most expensive people. In the authors' model, agents raise coding productivity by 73.7% and lower review productivity by 7.0%, and firms shift labor from coding to review.
Why AI review tools did not fix it
79.1% of firms use an AI code review tool. The bottleneck persists because a human still signs off, and humans write the overwhelming majority of reviews and comments. A tool can produce comments; it does not take responsibility for "this can go to production".
The authors' five-year simulation points the same way: if review productivity stays fixed, output rises 57.8% by 2031 and then levels off; if review keeps pace with coding, the rise is 341.4%. Review capacity, not coding speed, sets the size of the gain.
What to take from it
A caveat first: this is a working paper, not yet peer reviewed, and the data comes from customers of Jellyfish, which sells engineering analytics. The numbers moved between versions (the August version put the review-time increase at 49%, the November 14 version at 56.1%). The direction did not.
Practical consequences for teams:
1. Change what you measure. Lines of code and pull request counts no longer show success. Look at finished work, time from open to approval, and the share of changes sent back.
2. Budget review capacity. When you buy agent seats, plan who reviews that code. Counting how many people give 80% of approvals is a good start.
3. Give the reviewer more evidence. What stretches review is doubt: the reviewer has to work out alone what the change might break. If the tests touched by the change have already run, and the pull request shows which changed lines no test reached, the reviewer reviews design instead of "does it work".
The third point is where we work. CoverOps shows, on every pull request, which changed lines are covered by a test; TestingOps generates and runs tests for the change. We do not claim testing alone removes the bottleneck the paper measures, and neither do the authors. But as code volume grows at this pace, the difference between a change that arrives with evidence and one that arrives without it grows every month.