A support ops team pilots an agent that refunds orders, updates billing records, and closes tickets from chat requests, with strict limits on financial loss and audit needs. Where should human judgment be required as a critical decision point for validation?

The correct response places human judgment before irreversible financial state changes because refunds and billing updates directly create monetary loss and compliance exposure that cannot be fully undone by closing a ticket or writing an audit entry, so an above-threshold approval gate enforces least privilege, segregation of duties, and preventative control while preserving autonomy for low-risk chat and closure work. A daily retrospective review of audit logs after completion is a sound detective control commonly used for quality and forensics, but it fails as a preventative control for this pilot because money has already moved and strict limits require blocking excess loss before execution rather than detecting it afterward. Approving refunds while automating billing record updates is also common to reduce friction, yet it fails the stated scope because billing records carry equal financial and audit impact and leaving that write path autonomous preserves a direct route to ledger inconsistency and unapproved loss. Limiting approval to VIP customers only reflects a real prioritization practice for reputation risk, but it fails the threshold concept by substituting customer segment for monetary risk, leaving high-value non-VIP refunds unguarded despite the explicit loss limit. Exam caveat: distinguish preventative versus detective controls and segment-based versus threshold-based gates when the stem emphasizes strict financial limits. Practical check: verify the agent cannot call the refund or ledger-write tool above the amount without a recorded human approval token and immutable log entry.