A correctness audit of an AI agent can be completed in 30 minutes through six parts: factual accuracy, handling of unknown questions, triggering of transfer to a human, resistance to jailbreaks, escalation payload content, and peak-load stress checks.