I think the real challenge is catching hallucinations before the customer does. Testing AI with realistic scenarios, limiting what it can act on, and having a clear fallback when information is uncertain can prevent small mistakes from becoming bigger trust issues.