Faros AI ran telemetry across 22,000 developers this year. Median time a pull request spends waiting in review is up 441.5% year over year. Pull requests that merge with zero review, human or automated, are up 31.3%.
Read those together and the story isn't "review got harder." It's that the queue got long enough that people started routing around it.
Most tools sold as "code review automation" don't touch that. Static analysis, AI-generated review comments, merge gate checks in CI: all real, all useful, all upstream of the actual bottleneck, which is a person deciding whether a change is safe to ship. A bot leaving a comment on a 400-line diff doesn't make the diff smaller. It adds more text to read alongside it.
LinearB's numbers explain why the diffs got that big in the first place: AI-assisted PRs run about 2.5x larger than unassisted ones, and reviewer pickup time on them averages over 16 hours against roughly 200 minutes for the rest. Only 32.7% of AI-assisted PRs merge within a month. Writing code got faster. Reviewing it didn't.
The full piece digs into why "automating the comment" and "automating the review decision" are different jobs, and what a queue sorted by actual risk (not diff size) would need to look like: full post → https://tomosu.ai/blogs/code-review-automation.html
Top comments (0)