In 30 seconds
- Every AI output is a draft until a human approves it.
- Review shifts from "is the syntax right" to "is this the right thing, done safely".
- Review capacity becomes the team's real limit.
What reviewers check
- Intent: does it solve the actual problem?
- Fit: does it match the design and conventions?
- Safety: security, data handling, edge cases
- Tests: do they check the right things, or only what the code does?
- Understanding: could the team maintain this?
Keeping review from becoming a bottleneck
- Small changes, so each review is quick
- WIP limits on work waiting for review
- Automated checks for style and known issues
- Risk-based depth: more care for sensitive areas
Accountability
The person who approves AI-generated work owns it. That principle belongs in your Definition of Done and working agreements.
Myth vs reality
Myth · If tests pass, review is optional.
Reality · AI can write tests that confirm its own mistakes.
Myth · Review is quicker for AI code.
Reality · It can be slower: the reviewer didn't write or discuss it.
Myth · Humans just rubber-stamp eventually.
Reality · That's the risk to design against, not an inevitability.
FAQ
How much AI output can one reviewer handle?
Less than people expect; measure review time and adjust.
Should two people review AI code?
For high-risk areas, yes.
Can AI review AI code?
As a first pass, yes; a human still decides.