Two Agents. One Return. Two Refunds?
Hypothetical scenario: a small retailer receives a return at the counter while a support agent handles the same request online. Both systems see “not refunded.” Both issue money. Congratulations, the business has automated generosity.
This is a concurrency problem: two workers act on the same old information. Telling an AI to “check first” does not solve it. Both workers can check before either finishes.
Make the Database Say No
A forward deployed AI engineer starts beside the returns desk, tracing receipt verification, inspection, refund approval, payment, and ERP reconciliation. The model can interpret a messy customer message. It should not own the money button.
Assign each approved refund operation a permanent identifier. Enforce a required unique key in the integration database. PostgreSQL documents uniqueness constraints that reject duplicate values. Use the same operation identity across channels, not a fresh identifier for every incoming message.
Reserve the approved amount and record the pending payment task inside one database transaction: both changes commit, or neither does. Lock the order balance while checking remaining refundable value so different partial refunds cannot collectively exceed it.
That transaction does not include the payment provider. Send the persistent operation key through its supported duplicate protection mechanism. If a response times out, query the existing payment before retrying. Unknown is not failed.
Reuse the Failure Tests
The reusable agentic asset includes the receipt parser, policy lookup, constrained refund tool, approval record, and reconciliation worker. Its most valuable component may be the test suite: duplicate messages, simultaneous requests, partial refunds, and payments completed just before a connection drops.
Keep customer text as evidence, never as instructions to bypass approval. Post the confirmed refund reference to the ERP, then reconcile payment totals against credit entries. Flag mismatches for staff.
Illustrative assumptions: 300 monthly returns with six minutes of manual reconciliation reduced to three minutes release 15 hours. At $40 per hour, that represents $600 of capacity before software and support costs, not realized profit.
Measure duplicate payments, unresolved refunds, reconciliation time, and customer complaints. Faster mistakes are not transformation. They are mistakes with better distribution.