I saw a customer support tool once route urgent complaints into the wrong queue for almost two days. The system did not crash, so nobody treated it like an emergency at first, but real users were affected. That made me think companies need a better plan for AI mistakes, not just a technical fix after the damage is done. What should a company do first when an AI system makes a serious mistake?