AI-Powered Speed Crashes Reviews; 31% of PRs Merge Without Scrutiny
The team introduced AI coding tools, and code output doubled — but delivery efficiency didn't improve, it actually got worse. This isn't an isolated case. Faros' 2026 report shows that PR review wait times surged by 441.5%, and 31% of PRs were merged straight through without review. Opsera data backs this up: reviewing AI-generated code takes 4.6 times longer than reviewing human-written code. The pattern goes by the name "acceleration lash" — the output end sped up, but the digestion end (review, testing, release) fell behind, so the bottleneck shifted from "writing code" to "reviewing code."
Why does this happen? The root cause is a broken Batch Size. In the traditional flow, developers write a few hundred lines a day, and reviewers can get through them in about 20 minutes. AI has blown past that assumption: a single PR can now swallow thousands of changed lines, the comprehension cost spikes exponentially, and CI tests are easier to fail. DORA research finds that every 25% jump in AI usage correlates with a 7.2% drop in delivery stability. AI speeds up the things developers enjoy (writing code) while leaving untouched the things they resent (meetings, process overhead, bug fixes).
Even more worrying, mature engineering teams aren't spared either. Faros' comparison of 2025 and 2026 data shows that high-AI-adoption teams saw their PR volume climb by 98%, while organizational delivery metrics barely budged. The flood of code output crashes right up to the review doorway. Lines grow long, some people start merging straight to main without review. CodeRabbit's research adds another layer: AI-written code introduces 15–18% more security vulnerabilities, and debugging AI code takes longer than fixing human-written code. For the same REST API endpoint, a human wrote 29 lines; AI wrote 186. The nature of review itself shifts — from catching bugs to judging whether the change is necessary at all, a task that demands a much deeper systems-level understanding from reviewers.
For technical founders, the trap here is blindly piling on code volume. The actions are clear:
- Built a dedicated checklist: for AI-generated code, create an independent review checklist rather than retrofitting the old human-code standard.
- Bring in automated static analysis: let tools share the load, catching low-hanging bugs and security holes early, so human reviewers can focus on business logic and architectural soundness.
- Control Batch Size: enforce small-batch, high-frequency commits on AI-assisted PRs to prevent massive changes from paralyzing the review pipeline.
Source · Sagasu: Read the original post →