The most useful number in the Harvard working paper by Fiona Chen and James Stratton is not the 30 percent. It is the standard error: 0.17.
Chen and Stratton, both Ph.D. candidates at Harvard, measured AI coding agents across 725,938 workers at 718 companies. Lines of code added rose by 1,494 per worker-month against a baseline of 5,017, ComplexDiscovery reports on the paper. Commits rose 20 percent. Pull requests rose 23 percent. Finished work, counted as Jira issues resolved once their code reaches production, moved 0.12 issues per worker-month on a 3.67 baseline. That is statistically indistinguishable from zero, and the 95 percent confidence interval rules out any increase above roughly 12 percent of the baseline.
So the industry shipped a third more code and finished the same amount of work. The extra output went somewhere. The paper finds it in the review queue.
The queue is where the code went
Time from pull request submission to merge rose 3.45 days on a 7.03-day baseline, a 49 percent jump. The share of pull requests drawing a formal change request climbed about 12 percentage points from a 13 percent baseline, nearly doubling. Review comments per pull request rose 35 percent. The share of workers doing any review rose 14 percent from a base near 29 percent, while the share writing code did not measurably change.
That last pair is the whole story in miniature. Firms did not add reviewers. They pulled existing engineers into review, which is the labor-shifting channel the authors model. Faster drafting pushed more code at a fixed pool of approvers.
The authors checked the obvious escape hatch. Maybe engineers took on bigger tasks, so a flat issue count understates real output. They labeled a 0.5 percent sample of issues by size with GPT-4o, extended the labels to the full set with a text-embedding model, and found no shift in average issue size. Pull requests did not grow larger either. The simplest explanation for the review slowdown, that each unit of work got heavier, does not hold.
AI review tools did not clear it
Here is the part vendors should read twice. By March 2026, nearly 80 percent of the firms in the sample had used an AI code review tool. Those tools generated 23.3 percent of review comments. Only 10.8 percent of pull requests received even one AI comment. The bottleneck persisted after adoption.
The authors’ interviews put a face on it. A junior software engineer described a project where a single senior engineer held approval rights and could look at the interviewee’s code once a week. Then came outages traced to AI-generated commits, and new rules requiring senior sign-off on AI-written code. “Which created even more review bottlenecks,” the engineer said, in a quote the authors say they lightly edited for clarity.
That is a governance response, and it is rational. It is also a tax on the exact throughput the agents were bought to deliver. The paper’s two-channel model matters for anyone writing the check. In one channel, faster drafting floods a fixed reviewer pool, and moving labor into review mostly relieves it. In the other, the rise in change requests and comments signals lower-quality drafts, and that needs better technology, not more reviewers. The change-request and comment data, which faster drafting alone does not explain, points at the second channel. Buyers should care which one dominates, because the two call for different spending.
What the method can and cannot carry
The provenance deserves scrutiny before anyone quotes the 30 percent in a board deck. Chen and Stratton built the study on Jellyfish analytics data joined to LinkedIn employment records from Revelio Labs, covering January 2021 to March 2026. The tables list 68.3 million Jira issues, 200.9 million commits, 49.1 million pull requests and 21.3 million calendar events, though only 213 of the 718 firms supplied calendar data. The design is a staggered difference-in-differences comparison, and its validity rests on why adoption dates differed. The authors say timing was driven mainly by when firms learned about the tools and by procurement, which ran from about a day at one interviewee’s firm to six to 12 months at another. They acknowledge larger firms adopted earlier and control for size. Any residual bias, they argue, would make AI look more productive than it is, which makes the flat output result generous to the technology.
The paper is dated Aug. 4, 2026, with a first version dated Jan. 7. It is Chen’s job-market paper. Nothing in the version reviewed indicates journal acceptance or completed peer review. That is a real limitation and it does not dissolve the finding.
A third more code, the same finished work, and a review queue 49 percent slower. The agents did what they were sold to do. The organization around them did not move.
A second study points the same direction. A National Bureau of Economic Research working paper revised in September 2026, by Mert Demirer, Leon Musolff and Liyuan Yang, covers over 500,000 GitHub developers and finds a 240 percent cumulative rise in commits from autonomous agents shrinking to 30 percent for actual releases. That paper carries its own disclosure: Demirer and Musolff each held a postdoctoral position at Microsoft, which owns GitHub, and each now works as a paid research consultant for the company. Two datasets, two methods, one shape. Drafting accelerates. Delivery does not.
The legal detour is the tell
The ComplexDiscovery piece reads the software result through legal AI, and the parallel is sharper than it first looks. Harvey’s motion-to-dismiss workflow, co-developed with LexisNexis and announced Oct. 7, says the multi-step process of reading a complaint, researching and drafting “commonly takes 40 to 50+ hours of work” and that its tool produces research and a first draft “in a matter of hours.” The announcement describes “human checkpoints at the moments where a lawyer’s judgment matters most” but reports no measure of time spent at them. Cooley’s Oct. 8 redaction agent, still in development with Google Cloud, keeps attorneys making every redaction call before a filing goes to court, and carries no error rate. Federal Rule of Civil Procedure 11(b) makes that checkpoint a legal duty, not a product choice.
The software evidence says a first draft arriving in hours can leave the review step absorbing much of the savings. Legal buyers should ask vendors for approved work product, revision rates and reviewer time, not drafting speed. Chen and Stratton studied no legal work, and document review is not code review. The measurement gap is the same.
What builders should take from it
The agents work. That is not in dispute here, and the 30 percent rise in lines written is real. What the paper shows is that the constraint moved to the human checkpoint, and the tooling aimed at that checkpoint has not cleared it. Nearly 80 percent adoption of AI review tools, and a 49 percent slowdown in merge time. That gap is the product opportunity, and it is bigger than another code-completion model.
The other lesson is for information governance. The researchers identified agent use partly from traces inside the firms’ own systems: bot accounts tied to AI tools, “Co-authored-by” lines in commit messages and pull request descriptions. AI authorship leaves metadata in collaboration systems, and that metadata is a record a retention schedule or legal hold may need to reach. Most firms have not thought about that yet.
Watch for a revised version of the Chen and Stratton paper, and for any vendor that publishes review-stage figures instead of drafting-speed claims. The first one to do it will be telling you something the rest of the market is still hiding.