AI code review: logic bugs or just style nits?

I've been using AI reviewers in CI for a couple of weeks now. They're great at flagging missing null checks and suggesting performance tweaks, but I'm starting to wonder if they actually understand the business logic. Yesterday one suggested a "simplification" that would have broken an edge case I spent hours handling. Maybe I'm prompting wrong, or is there a ceiling to what these tools can reason about? Curious how others are drawing the line between useful suggestions and noise. Are you letting AI approve pull requests, or keeping it strictly as a hint machine?