As LLMs get integrated deeper into real workflows, one bad prompt could misroute a customer, corrupt a ticket, escalate the wrong case, or trigger a human handoff that didn’t need to happen ...