Refunds, credits and exceptions: draft and hold

An AI agent should treat refunds, credits and exceptions as three different actions. It drafts all three with the evidence attached. Account credits inside the limit you set can be applied at depth 3, because they are reversible. Refunds move money and wait for a named approver above that limit. Exceptions to policy are always held, whatever the amount.

What are refunds, credits and exceptions, and who owns them?

A refund sends money back to the customer, usually to the original payment method. Once it lands, getting it back means asking the customer. A credit reduces what the customer owes you next, on their account or as a balance. You can reverse it if it was wrong. An exception is a decision to set aside a rule: a refund outside the return window, a fee waived that the terms say applies, a second goodwill gesture in a month.

Finance owns the limits and the money. The support lead owns the process and the approver rota. Each tier of approver is a named person, not a team mailbox. Why drafting first is the right starting point for any new support agent is argued in draft and hold is the right default for new support agents. The wider support position is in the AI support agent that knows when to say no.

How does the agent draft and hold, step by step?

1. Agent, depth 1: establishes what happened. The order or account, what the customer was charged, what they received, and what they say went wrong. Customer statements are declared; order and payment records are sealed.
2. Agent, depth 1: finds the policy clause. Return window, fee terms, service commitments. It cites the clause by version.
3. Agent, depth 2: decides which action fits. Refund, credit, or exception, and the amount the policy supports. If the policy does not support it, the action is an exception by definition.
4. Agent, depth 2: drafts the action. Amount, type, reason code, policy clause, evidence, and the reply the customer will receive.
5. Platform: checks the draft against the minimum chain. The lowest bound wins. The verdict names it: the customer's limit, the certificate, the role scope.
6. Agent, depth 3: applies allowed credits and sends the confirmation.
7. Agent, depth 2: tells the customer plainly about held actions. What was asked for, that a named colleague will decide, and when they will hear.
8. Person, approver for the tier: decides held actions. Approve, change the amount, or decline with a reason that goes back to the customer.

What goes in a held draft?

Approvers work fast when everything is on one screen. A held draft should show:

• the customer, the order or account, and the amount;
• the action type and why the agent chose it;
• the policy clause and whether the request falls inside or outside it;
• the evidence, each item labelled with its source and provenance class;
• previous refunds, credits and exceptions for this customer;
• the drafted customer reply, ready to send on approval.

Writing the reply is the same skill as drafting replies to support tickets for staff to send, with one addition: the reply must state the amount and where it will go.

Why are exceptions always held?

An exception is a precedent. If the agent waives a late fee for one customer because their story is persuasive, the next customer with the same story expects the same. Applied at scale, a few kind exceptions become an unwritten policy nobody approved. The longer argument is in can an AI agent make exceptions to company policy.

Holding every exception also gives you data. After a few months, the exceptions approvers grant most often show where the written policy is out of step with what the business is willing to do. Change the policy; then the agent can apply it inside its limits.

Which controls matter, and what evidence do they leave?

Refunds go to the original payment method only. A customer asking for the money to go to a different account is making an external claim, and an external claim cannot authorise a payment. Evidence: the destination checked against the payment record.

Duplicate check before every action. Has this order already been refunded or credited? Evidence: the check and its result in the verdict.

Limits enforced at the moment of action. Evidence: the verdict naming the limit and its version.

Approver is named. Evidence: the person, the time and the decision, in the append-only record.

Cumulative limits. Ten credits just under the limit are still ten credits. Set a per-customer total over a period. Evidence: the running total on each draft.

How do you staff the approval queue?

Held actions are only safe if someone works them. Give each tier a named approver and a deputy, a service level measured in hours, and a daily view of what is waiting. If the queue grows, look at which holds approvers always approve; those are candidates for a higher limit on Attest evidence. The staffing model is covered in how do you staff the queue of actions AI agents hold.

What should you measure?

• Held actions approved unchanged, by tier. High rates suggest a limit is set too low.
• Time held. The customer's experience depends on it.
• Credits later reversed, as a check on depth 3 accuracy.
• Exceptions by reason, to inform policy changes.
• Repeat claimants reaching the cumulative limit.