“I do a lot more reviews now because they come in very, very quickly.” That is how a developer, labelled P6, describes the new workload in a mixed-methods study on arXiv by Kaia Newman, Christian Bird, Jenna Butler and Emerson Murphy-Hill, posted on 13 November 2026. The paper, which carries Microsoft affiliations, interviewed 24 people and surveyed 2,590 at a large multinational software company. Developers say AI review makes their reviews faster and better. The people using it are no less buried in pull requests than those who don’t.
Loved by users, invisible in the backlog
The adoption numbers are lopsided. Across all 2,590 respondents, 82% have personally used AI for review and 93% have seen an AI bot comment on PRs they review. The paper reports that 78% judged their review quality better (4% worse) and 68% judged it faster (11% slower), with no base stated. Another 71% agree their review has shifted toward design and intent, asked only of respondents whose careers began before AI-assisted coding.
One interviewee, P9, explains why: “As the lower-level review gets automated and out of the way, it enables for much richer higher-level review.” The abstract is blunter about the other half. “AI review usage was not correlated with less difficulty keeping up with PR volume, indicating using AI for review alone is not sufficient for solving AI-driven review challenges.”
Personal use was in fact linked to slightly more difficulty (p=.006, r=.16, a small effect). The authors suspect the arrow runs backwards: rising volume pushes people to adopt the bot, not the bot to create the pile.
What actually grew is volume and absent authors
If that reading holds, a reviewing aid is aimed at the wrong end of the pipe. The bot makes each review cheaper. It does nothing about how many arrive, or about who stands behind them.
P6 puts the second problem in one line: “It’s more responsibility on the people who are reviewing and zero responsibility on the authors.” P15 says of AI-assisted authors that “they don’t look at the code they’re producing.” This echoes the Harvard finding that agents raise code output while reviewers send more of it back.
Here is where Pipeline goes beyond the data. The study measured self-reported difficulty and self-reported use, at one company, at one moment. It did not measure workload. Read as a whole, though, the pattern suggests faster reviewing cannot absorb more and larger PRs from authors who haven’t read them. That is a conclusion we draw, not one the survey tested.
A review bot can clear the small stuff, but it cannot shrink a queue that is growing because authors stopped reading their own code.
The helpful strategies all happen before the reviewer arrives
The most useful finding is about timing. In the paper’s ranking of helpful strategies, pre-review came out on top, and its AI version, authors running an AI review before a human sees the PR, is already used by 69% of the respondents asked. The same bot is most valued when the author points it at their own work, not when it sits beside the human reviewer.
Other practices point the same way. Self-review expectations were used by 41% of respondents, sharing author intent by 36%, and requiring PRs to be ready by 41%. GitHub says much the same in its guidance on agent PRs: “Reviewing your own pull request isn’t optional when agents are involved.” That is vendor advice, so read it as corroboration of the ownership point, not as evidence.
If you review teammates’ pull requests, the practical meaning is plain. The bot will flag the small stuff. What shrinks your pile is authors reading and annotating their own agent-written PRs before they tag you. That is a team norm, not a tool you can buy.
What the study cannot say
The objections are real, and they come from both sides. Atlassian’s ICSE 2026 paper on RovoDev Code Reviewer reports a 30.8% drop in median PR cycle time and 35.6% fewer human-written comments across more than 1,900 repositories. So AI review can show up as throughput in system data. But that result is observational and comes from Atlassian’s authors evaluating their own tool.
The Microsoft study has limits too. It is cross-sectional self-report at a single company, with a 9% response rate (6% fully completed). Only 209 respondents, 8%, had reviewed solely in the post-AI era. The authors themselves note that reported helpfulness is not demonstrated effectiveness.
That cuts against our own step as well. Respondents rated upstream strategies as helpful; nobody showed they shrink the queue. The two datasets also don’t conflict outright, since a bot can speed a single PR through while the total pile keeps growing. What they leave open is whether any tool can fix a volume problem that begins with the author.
So the sensible reading is modest. Install the reviewer if you like it, since people clearly do. Just don’t expect the queue to notice, and move the real effort to the moment before someone presses “request review”.



