Do AI code reviewers actually catch real bugs for you?

We rolled out an AI reviewer on our PRs last month. It's great at flagging style nits and missing null checks, but it keeps missing the stuff that actually bites us: race conditions, bad assumptions about retries, migrations that lock a table. Curious how others are using these tools. Are you running them as a first pass before human review, or just letting them comment freely? And has anyone found prompting tricks that get them past the shallow stuff? I'm starting to think they're better as a rubber duck than a gatekeeper, but I'd like to be proven wrong.