AI can write 10x more PRs now and the team is still checking them at the same pace as last year.
That gap? It’s growing every sprint. And almost nobody is tracking it.
The Metric Everyone’s Celebrating Is Wrong
Teams are excited about the number of PRs. “We delivered a PR volume 40% higher this quarter.” Nice! But did someone check them?
AI coding assistants write more code. They don’t write extra reviewers. The bottleneck didn’t vanish. It just shifted downstream.
DORA Doesn’t Capture This (Yet)
DORA metrics measure lead time, change failure rate, and more to quantify engineering productivity.
However, there is a blind spot here. Lead time is the clock from first commit to production. If a PR sits in a review queue for three days because every senior engineer is buried under a pile of AI-generated diffs, that shows up as slow lead time — but nobody’s attributing it to the right cause.
The way the system fails is not obvious:
- AI generates more PRs, faster
- Review queue balloons
- Reviewers skim instead of reading carefully
- Change failure rate creeps up
- Team blames “quality issues” instead of recognizing a capacity problem
You’re not shipping faster. You’re just creating work faster. That’s not the same thing.
AI Review Debt Is Real
The term “AI Review Debt” was recently coined by Sumant Thakur on his Substack, and it couldn’t be more accurate. Any AI-generated PR that enters a review queue without an associated expansion of review capacity is debt. It grows unnoticed.
Most teams are already feeling this. Senior engineers are overwhelmed. Juniors are generating PRs with Copilot but don’t yet have the context to review each other’s work meaningfully. Three team members end up being the bottleneck for everything.
And unlike regular tech debt, nobody’s putting this on a roadmap. There’s no Jira ticket for “our review pipeline can’t keep up with our generation pipeline.”
What Actually Helps
There’s no clean solution yet, but some things appear to be heading in the right direction:
- Measure review queue depth and review cycle time separately from lead time. If you can’t see the bottleneck, you can’t fix it.
- Stop celebrating PR count as a productivity metric. Merged PRs matter. Open PRs are inventory, not output.
- Invest in review tooling with the same energy you invested in generation tooling. AI-assisted review is coming, but most teams haven’t even explored what’s available today.
- Set explicit review capacity limits. If a human can thoughtfully review 4–5 substantial PRs per day, that’s your throughput ceiling. Plan around it.
The uncomfortable truth is that adding AI coding tools without reconsidering review workflows simply shifts the pressure from writers to reviewers. That’s a problem with team design, not the tooling.
The Real Risk
When reviewers are overloaded, they tend not to decline more PRs — they merge them faster.
That’s the worst possible outcome. You get the illusion of velocity with the reality of degraded quality. The change failure rate increases. People lose confidence in the system. And six months later, everyone’s wondering why the product feels fragile.
The bottleneck moved. The org chart didn’t.
If you doubled your PR output this year because of AI tools, one question worth asking: did your review capacity also double? If the answer is no, you’re accumulating debt you haven’t named yet — and the longer it goes unmeasured, the more expensive it becomes to address.
