Should AI Auto-Send Refunds?
No. An AI support agent should not auto-send refunds. Moving money, issuing refunds above a threshold, approving chargebacks, and similar high-risk actions belong on a forbidden-action list that always routes to a human. The AI can draft the response, gather the context, and recommend an amount, but a person should approve any action that spends money or that a company would be legally bound by.
Why refunds are a hard line, not a confidence question
Some actions are too risky to gate on a confidence score alone, because even a well-calibrated model is occasionally confidently wrong. Refunds, fraud decisions, chargebacks, legal commitments, and account-ownership changes fall in this class. These are not "escalate if unsure" cases; they are "always escalate" cases, regardless of how sure the model seems.
Compute the money in code, never in the model
When a refund is warranted, the amount, the eligibility window, and the policy should be computed deterministically in code from real system data, never authored by the language model. The Air Canada tribunal case, where a chatbot invented a bereavement refund policy and the company was held liable, is the precedent every support team should keep in mind. You own every word your bot says.
How Resolve handles refunds
Resolve keeps a hard forbidden-action list, refunds over a threshold included, that always produces a decision card instead of an auto-send. Business logic like refund amounts and return windows is computed in code, not by the LLM. Reva presents the human the request, the relevant policy, its recommended action, and its reasoning, and the human approves, edits, or rejects. The AI never moves money blind.