AI code review — the two ways teams fail at it

The main tools

Cost inverts with team size and PR volume. Many people and few PRs favours Greptile, which does not charge per seat; a small team shipping constantly is cheaper on CodeRabbit or Graphite. Our CodeRabbit vs Greptile piece works the arithmetic.

Failure 1: too many comments, and the PR stalls

In the first week a single pull request collects twenty or more comments. A conscientious team tries to address all of them and time-to-merge gets worse than it was before the tool. That is when 'AI review gets in the way' becomes the settled opinion.

Run it in two stages. For the first two weeks, state explicitly that comments are advisory and acting on them is optional, and record which categories actually proved useful. At the end of the two weeks, disable the categories that did not.

• Candidates to disable: naming preferences, suggestions to add comments, nudges toward consistency with existing style
• Worth keeping: null and undefined handling, missing error handling, type mismatches, racy async code
• Depends on the team: suggestions to add tests

Failure 2: a comment thread nobody reads

The opposite pattern. Ignoring the bot becomes normal, the PR thread fills with automated comments, and human review comments get buried. Ignoring it is not neutral — it is actively harmful.

Make the handling a rule. Every comment gets one of three responses: fix it, deliberate (write one line saying why), or defer (open an issue). Nothing is left unprocessed. Choosing between three options takes ten seconds, which is why it actually happens.

Choose the depth per pull request

Some tools let you set review depth. Greptile's credit tiers (Base 1, Plus 3, Apex 10) control how deeply it reads. Sending every PR to the deepest setting costs money and reproduces the too-many-comments problem.

• Dependency bumps and copy fixes → no review, or the lightest setting
• Ordinary feature work → standard review
• Anything touching auth, payments or data migration → the deepest review

Some tools can route this automatically by label or path. Manual selection never survives contact with a busy week, so make configurable routing part of the selection criteria.

What human review is still for

Human review does not go away. What remains is conformance to the specification, soundness of the design, and whether the change should be made at all. The AI judges correctness inside the diff it is given; it does not ask whether that diff should exist.

The change teams actually report is not that review takes less time but that review conversations move toward design. Clearing the mechanical layer first is what buys that.

Frequently asked questions