A confident wrong answer occurs when an AI agent provides incorrect information with high apparent certainty, delivering the response in a way that gives the customer no reason to question its accuracy. Unlike obvious errors or hedged responses, confident wrong answers are dangerous precisely because they're convincing.
This pattern is one of the hardest failure modes to detect through standard metrics. The interaction appears successful from every quantitative angle - the customer didn't escalate, the conversation flowed naturally, and the customer may even express satisfaction because the agent 'sounded knowledgeable.' The error only surfaces when the customer acts on the incorrect information and discovers it doesn't work. Containment rate, CSAT, and handle time all miss it. Detection requires evaluating the substance of what the AI actually said and verifying it against ground truth - which is why evidence-based conversation analysis matters more than outcome metrics for AI agent evaluation.
