A product manager deploys a customer-facing support agent that can draft refunds and update billing records for a SaaS product. Finance requires approval for refunds over $50 and a block on direct production writes. Which control should the manager configure to implement agent guardrails?

The correct control requires human approval for large refunds—configured as refunds over $50 to match finance’s threshold—and routes billing writes through a review queue. Together, these controls prevent unapproved refunds and block the agent from writing directly to production, supporting segregation, reversibility, and least privilege. Requiring approval over $50 while logging direct billing writes gets the refund threshold right, but audit logs are detective, not preventive: direct writes can take effect before anyone reviews them, so the production-write block is missing. Requiring approval over $200 and queuing billing writes preserves the write restriction, but the higher threshold lets refunds between $50 and $200 proceed without the approval finance requires. Allowing refunds over $50 with post-action manager sign-off also queues billing writes, but retroactive sign-off cannot prevent a refund from being issued before approval. An exam caveat is to distinguish preventive controls from detective controls, and pre-execution approval from post-action review; logging or later review does not meet an explicit block or approval gate. As a practical check, test a $51 refund in staging and confirm it pauses for approval, then verify an agent billing-write request creates a queued proposal rather than changing production, including when approval is denied or expires.