Scenarios

I'd want three things resolved before the tool goes live. First, failure ownership: who eats the loss if the agent approves a fraudulent claim under $500, and how fast does the fraud pattern surface — this is a T3 data-side accountability question about the agent's access to expense history and vendor-master data, not just a policy question. Second, the audit trail: is a human-reviewable log kept for at least the statute-of-limitations window, and can a compliance auditor trace any single approval back to the specific model version and prompt at the time of decision. Third, the escalation threshold: some categories — travel, client entertainment, cash advances — should route to a human regardless of dollar amount, because dollar cap is the wrong axis of risk for those categories. Then I'd propose a two-week shadow-mode pilot where the agent runs but a human still approves, so we can compare rates and override patterns before flipping the switch. The override rate — how often the human catches something the agent missed — is the signal I'd track for the T5 override-rate anchor.