A hallucinated resolution happens when an AI agent marks an interaction as resolved based on a fabricated or incorrect solution that doesn't actually address the customer's problem. The AI generates a response that sounds like a valid resolution - complete with specific steps, reference numbers, or policy citations - but the underlying information is invented or misapplied.
This failure mode is distinct from simple incorrect answers because it specifically involves the AI believing it has solved the problem. The customer leaves the interaction thinking their issue is handled, the system records a successful resolution, and every metric looks healthy. The reality only surfaces later when the promised action doesn't happen, the cited policy turns out not to exist, or the recommended steps don't work. Hallucinated resolutions compound over time - each one generates a future callback, erodes customer trust, and creates work for human agents who have to untangle what the AI told the customer. Detecting them requires analyzing what was actually said in the conversation, not just whether it ended cleanly.
