Two Harvard researchers, Fiona Chen and James Stratton, measured what happened inside about 700 software companies after they started using AI coding agents, and Ars Technica wrote it up last week. Their working paper uses data from Jellyfish, a company that sells engineering analytics, and runs from January 2021 to March 2026. After agents arrived, lines of code rose 30% and pull requests, the batches of changes a programmer submits for someone else to approve, rose 23%.
Those are the numbers a sales deck would use. The finished work didn't follow them. The study measured tasks resolved per worker per month, counting a task only once the work was live, and found no effect it could tell apart from zero. The authors' estimates rule out a rise in finished work of more than 12%.
The extra code piled up at review. The time from submitting a pull request to merging it into the main code rose 49%, from a starting point of about a week (7.03 days). The share of pull requests sent back for changes nearly doubled, and comments per pull request rose 35%. The authors' own summary is that "code review becomes the bottleneck." AI made producing the work cheap. It hasn't made checking it cheap.
These were not small companies. The firms in the sample averaged 1,652 workers. A company that size has more people to put on review. At these firms, the share of workers doing reviews rose 14%. My inference, which goes past what the study measured, is that in a company of 30 the reviewer is usually the owner, or the one senior person the owner trusts. I'd expect the same thing anywhere AI drafts the work. Every quote and customer email the tool produces lands on that person's desk before it goes out. A Microsoft Research survey of 319 knowledge workers reported that generative AI moves the thinking in their work toward checking what the tool gives them.
So the owner who buys an AI tool to get out of the weeds can end up deeper in them. The business already ran through one person's judgment, and now it produces more work that needs that judgment. That moves you further from a business that can run without its owner.
AI may learn to check its own work, but it isn't there yet. Nearly 80% of firms had used AI tools in code review by March 2026, yet AI wrote only 23.3% of review comments. Agents are new, and that share may grow. For now, people do most of the checking, and somebody has to answer for what ships.
The practical move this month is to settle the checking before you add the tool. Before you put AI on any drafting work, name who reviews its output, and write down what "done" means for that work in terms a second person could check. Then count finished results, like jobs quoted and won or invoices out the door. If the only reviewer is you, fix that before you buy anything, because a tool on top of unfixed work just makes more of it. RVRBN builds AI systems for small businesses, so we have an interest in owners adopting them and in those systems working.
Watch whether AI starts carrying more of the review in the next round of studies. Until then, AI multiplies whatever review process you already have. If you're the only one who can say a job is done, the extra work will bury you.
If you want to talk through what this means for your company, get in touch.