The failure mode support leaders will see
A customer writes in after a billing issue. The AI support agent reads the ticket, finds the account, decides a refund is warranted, and prepares a Stripe refund. The answer may be correct. The control problem is that the amount is above policy.
Without an approval gate, support velocity turns into a finance incident: the refund executes, the ticket closes, and later finance has to reconstruct why the AI sent money.
What AgentGovernance demonstrates
In the live control plane, `Billing-Agent-03` attempts a Stripe refund for `$640` to Acme Corp. The policy `POL-SPEND-LIMIT-004` requires human approval for charges or refunds over `$500`, so the action is routed through the approval queue and recorded as an approved high-risk action.
Threshold before Stripe
The policy evaluates the refund amount before the payment action runs, not after money moves.
Human approval
The reviewer sees the amount, customer, ticket context, policy, and reason before approving.
AI audit trail
Approval and final Stripe result are separate facts, so finance can see both the decision and the outcome.
Narrow rollout
Start with refunds over limit before expanding to credits, cancellations, chargebacks, or retention offers.
The first policy to ship
- Refunds under threshold can be auto-approved when customer identity and order state are verified
- Refunds above threshold require a named support or finance approver
- Disputed transactions, repeat refunds, and fraud-flagged accounts always route to review
- Every approval stores ticket ID, customer, amount, policy, approver, timestamp, and Stripe result
- Timeouts default to no refund, not silent execution