"AI reviewed this PR" gets said the same way "it's insured" gets said, like the phrase closes the question of risk on its own.

It doesn't. AI review genuinely catches a real category of problems: syntax slips, inconsistent naming, obvious gaps in error handling. That coverage is real. What it structurally can't see is anything living outside the diff, a business rule that's internally consistent but wrong, a downstream service assuming something the changed files never mention, a missing test that produces no visible defect because it's an absence, not a flaw.

The failures that actually hurt aren't the obvious bugs the AI missed. They're the ones that trace back to a category the review was never covering in the first place, discovered at the worst possible time, the same way an exclusion clause gets discovered when a claim is denied.

I wrote up where that boundary actually sits, with a real example of how a "clean" PR can still ship a production incident: https://medium.com/@nagatomopedro05/i-thought-the-pr-was-covered-it-wasnt-2e7164da47b9