We asked three engineering teams to tag every automated review comment as useful, noise, or wrong. The distribution was more consistent than any of them expected.
Where the value concentrates
Null handling, forgotten error paths, inconsistent naming and missing test cases dominated the useful bucket. Design feedback — the thing senior reviewers actually provide — barely registered.
That is a reasonable division of labour. Reviewers stop spending attention on mechanical issues and arrive at the human conversation with the small stuff already resolved.
The failure mode to watch is confident wrongness on unfamiliar internal APIs, which teams learned to spot within a couple of weeks.
Reader comments 1 comments
Sign in to comment.
Bench numbers or it did not happen.