The danger in enterprise AI is no longer just a wrong answer - it is an autonomous action that nobody can fully reconstruct after the fact.