A refund and credit limits matrix sets how much an AI support agent may refund or credit, for which reasons, on what evidence, and to whom, with three caps: per transaction, per customer over a period, and a daily total across all customers. Above any cap, the agent drafts and a named person approves.
Why three kinds of cap?
A single per-refund limit looks safe and is not. An agent allowed to refund up to 50 can refund 49 a hundred times to one customer, or 49 to a thousand customers in an afternoon if something has gone wrong with a rule or a source. Each cap catches a different failure:
- Per transaction bounds the harm of one wrong decision.
- Per customer per period stops one customer, or one impostor, draining the account through repeated requests.
- Daily total for the agent limits the damage from a systematic error before anyone notices.
The method for choosing the numbers is covered in how to set refund limits for an AI agent.
When should the matrix be used?
Before an agent is allowed to issue any refund or credit, and every time products, prices or refund policies change. It is also the basis for certification: every cell should be tested at the cap and just above it. Whether an agent should hold the capability at all is answered in whether an AI agent can approve a refund: yes, within bounds like these, once certified on your own cases.
How do you fill in each field?
- Reason. The specific, checkable reason for the refund: duplicate charge, item returned, service not delivered, fee charged in error.
- Evidence required. What the agent must see in your systems, not just hear from the customer.
- Per-transaction cap. The largest single refund for this reason.
- Per-customer cap. Total refunds and credits to one customer in the period, across all reasons.
- Daily agent total. The most the agent may issue across all customers in a day.
- Exclusions. Customers or accounts where the agent never refunds: flagged for fraud, in dispute, under review.
- Destination. Original payment method only, or account credit. Never a new account.
- Above the cap. The approver, and the time target.
Goodwill credits draw on the same budget. Make the per-customer cap count both, and keep it consistent with the goodwill gesture matrix. Requests above a cap become escalations, routed as set out in the escalation matrix for customer support agents.
The refund and credit limits matrix
Global caps: per customer per period ; daily agent total ; destination rule ___.
| Reason | Evidence required | Per-transaction cap | Exclusions | Agent depth | Above the cap: approver and time target |
|---|---|---|---|---|---|
A filled sample: an online retailer
Global caps: per customer, 150 in any 90 days across refunds and goodwill; daily agent total, 5,000; refunds go to the original payment method only.
| Reason | Evidence required | Per-transaction cap | Exclusions | Agent depth | Above the cap |
|---|---|---|---|---|---|
| Duplicate charge | Two identical captures on the order | Amount of duplicate, up to 500 | Accounts flagged for fraud | 3 | Payments team, same day |
| Item returned | Return scanned at warehouse | Item price, up to 200 | Returns outside policy period | 3 | Returns team, two days |
| Item not delivered | Carrier shows no delivery after the latest expected date | Item price, up to 100 | Customer has three or more non-delivery claims in a year | 3 | Customer service lead, two days |
| Item not delivered | Carrier shows delivered; customer disputes | Not delegated | All | 2 | Customer service lead |
| Delivery fee, late delivery | Delivery after promised date | Delivery fee | None | 3 | Not applicable |
| Price adjustment | Price dropped within policy window | Difference, up to 30 | None | 3 | Customer service lead |
Look at the two "not delivered" rows. Where the carrier shows no delivery, the evidence supports the customer and the agent can refund. Where the carrier shows delivered and the customer disagrees, the evidence conflicts, so a person decides. The rule follows the evidence, not the customer's insistence.
The 5,000 daily total matters more than it looks. If a carrier's tracking feed failed and showed every parcel as undelivered, the agent would hit the cap within hours and stop, rather than refunding a full day's orders.
Where Pryme fits
In Pryme, each cap is a customer limit in the minimum chain, checked at the moment of action alongside role scope, certificate, sources and platform guardrails. When a cap decides, the refund is held for the named approver and the verdict names the cap. The stance is set out in the AI support agent that knows when to say no, and the same logic for retention in retention offers need limits, not judgement.
Questions people ask
Should the agent tell customers the limits?
It does not need to quote them. It should say when a request needs a colleague and when they will hear, so the limit never feels like a dead end.
What if a customer asks for a refund to a different card?
Refuse and explain that refunds go to the original payment method. Redirected refunds are a common fraud route, so this rule should have no exceptions at the agent level.
How often should caps be reviewed?
Monthly at first, using the rate of held requests that people approve unchanged. If most are approved as drafted, a cap may be raised through a recorded change; if many are rejected, look at the evidence rules before the caps.