AI made your team ship 2x the code. Your reviewers didn't get 2x the eyeballs

Two numbers from the same telemetry dataset (22,000 developers, two years) that explain most of the AI-productivity debate: in high-AI-adoption quarters, median code review time rose fivefold — and merges with no review at all rose 31.3%. A queue that quintuples while more people climb over the turnstile means the bottleneck moved. It didn't disappear.

The output side is everything the demos promised: tasks per developer +34%, code-heavy tasks +210%. The absorption side is where it curdles: bugs per dev +54%, incidents per PR more than tripled, commit-to-production lead time +480% where organizations measured it.

Our read: this is arithmetic before it's culture. Generating a PR now costs roughly one sentence of intent. Accepting it still costs a human, reading at human speed, with human working memory. When one side of a pipeline drops its marginal cost by orders of magnitude and the other side doesn't move, the constraint relocates to the expensive side. It has nowhere else to go.

Every fix we've seen sorts into two buckets. Compressing acceptance — lighter or no review — converts review cost into incident cost, payable later with interest. Leveraging acceptance is the unglamorous bucket that works: smaller PRs (size is the axis authors control, and it moved +51% the wrong way), artifacts that arrive carrying their own evidence — test output, the exact command to re-run — so review shifts from reconstructing intent to checking claims, and machine passes before human attention.

One honest caveat: nearly all public numbers are creation-side telemetry from vendors with something to sell. The acceptance-side data that would settle this — revert rates by review depth — mostly doesn't exist yet.

What does your team do when the review queue backs up?