Preventing duplicate refunds in AI agent integrations with signatures, idempotency and limits

AI assistants are beginning to call business systems on users' behalf, such as requesting a return or rescheduling an appointment. Reading order data is low-risk, but refunds move money and require careful engineering. The author outlines four key failure modes to design for, with TypeScript examples using the open-source allowfy package, which applies to any webhook-based integration. Requests are verified against a signed timestamp over the raw body with a constant-time comparison to block replays from the internet-facing endpoint. Agent actions are split into a quote step, where terms are computed without changes and approved by the customer, and an execute step, so that changed terms invalidate prior approval. Stable operation IDs and idempotency keys, passed down to the payment provider, ensure a retry after a dropped connection does not refund twice.