Over the past week, several technology outlets picked up a working paper by two Harvard economists, Fiona Chen and James Stratton, titled Artificial Intelligence in the Firm: Bottlenecks in Software Production. The paper is a job market paper in its August 4 version and has not been peer reviewed, and its data comes from Jellyfish, a company that sells engineering analytics to software teams and reviewed the manuscript only to check that it disclosed no proprietary information. With those two caveats on the table, the dataset is unusually large: around 300 million work events, including code activity on GitHub, Jira issues and calendar entries, from 718 firms between January 2021 and March 2026, with usage data for GitHub Copilot, Cursor and Claude Code. Comparing firms by when they adopted the tools, the authors find that AI coding agents raise lines of code per worker by 30%, commits by 20% and pull requests by 23%. Earlier AI assistants, the ones that only suggest code as you type, moved the same measures by 12%, 9% and 5%, and only the effect on commits was statistically significant.
Then comes the part that matters to anyone who signs a budget. Measured by finished work, meaning Jira issues resolved once the code is deployed and epics completed, the effect of agents is small, positive and statistically indistinguishable from zero; the estimates rule out any gain in output larger than 12%, well below the 30% gain in code. Employment does not move either: the authors can rule out a decline in overall headcount larger than 2.9%. The missing output is sitting in the review queue. After firms adopt agents, the average time to review a pull request rises 49%, the share of pull requests sent back for changes nearly doubles, comments per pull request rise 35%, and the share of workers doing code review grows 14%. The bottleneck persists after firms adopt AI code review tools, because human reviewers keep the central role. A survey of 100 engineers the authors ran to validate their time estimates explains why: engineers spend only 30% to 40% of their time writing code, and the rest goes to planning, reviewing, testing and deploying, steps the agent does not do for them.
Why this is not only about software
The mechanism the paper describes is general. Any production process has a step the AI speeds up and a next step that checks the result: the partner who signs the brief, the accountant who closes the books, the manager who approves the proposal, the supervisor who reads the reply before it goes to the customer. If the first step produces 30% more and the second step keeps the same capacity, the extra volume waits in a queue, and when the drafts are also of a different quality, as the near doubling of revision requests suggests, the second step gets slower per piece as well. The paper also shows what companies actually did about it: they moved people toward review. That is the investment that turns more drafts into more finished work, and it rarely appears in the business case for the licenses.
What to do before buying more AI
- Measure finished work, not activity. Lines of code, drafts, pages and messages produced are measures of effort. Count what leaves the building: features deployed, matters closed, invoices sent, tickets resolved. If that number has not moved since the AI arrived, the gain is still inside the company.
- Size the review step before the writing step. Before adding seats, ask who will check the extra output and how many hours a week they have. If the honest answer is nobody, the licenses buy a longer queue. Make review planned work with named owners and time on the calendar; the firms in the study ended up moving people toward review in any case.
- Make the agent deliver smaller, checkable pieces. A reviewer takes less time on ten small changes than on one large one, and sends fewer back. Ask the agent to work in bounded tasks, to run the tests or the checklist itself before handing over, and to state what it did and did not verify.
- Grade review by risk, not by habit. Not every output needs the same scrutiny. Low-risk changes that pass automated checks can go with one approval; anything touching payments, access, personal data or a client deliverable keeps two pairs of eyes. Track rework alongside speed, so a faster lane does not become a leakier one.
On our own team, the output we report on is finished work accepted by the person who supervises each agent, and the first thing we size with a client is who will review what the agent produces, because that person, not the license, sets the pace.
Sources
- Artificial Intelligence in the Firm: Bottlenecks in Software Production — Fiona Chen and James Stratton, Harvard University (working paper, version of August 4, 2026)
- AI coding agents write 30% more code and resolve zero additional issues, Harvard study finds — Forkast (October 11, 2026)
- Harvard study: AI coding agents lift code 30%, shipped output flat — AI Weekly (October 12, 2026)