Guardrails for AI Customer-Support Agents That Issue Refunds, Credits and Account Changes
An agent that can refund, credit or change an account needs controls outside the model: value limits enforced by the payments system, verification before any account change, eligibility computed from the written policy, escalation that gives a person the whole case, and a record of every action with the approval and policy behind it. A prompt can describe the policy. It cannot enforce it.
On this page
A support agent that only answers questions can be wrong. One that can issue refunds can be wrong with your money.
Customer-service agents often start with answers and grow into actions: a goodwill credit for a late delivery, a refund on a damaged item, a new delivery address, a changed email on the account. Each step makes the agent more useful, and each moves a mistake from the transcript into the ledger.
What the Air Canada chatbot decision does and doesn't settle
In Moffatt v. Air Canada, 2024 BCCRT 149, British Columbia's Civil Resolution Tribunal rejected the airline's argument that it could not be held liable for information from its chatbot. "It should be obvious to Air Canada that it is responsible for all the information on its website," the tribunal member wrote. "It makes no difference whether the information comes from a static page or a chatbot." We covered what that means for what customer-facing agents say.
That chatbot only described a policy. Air Canada's staff, not the chatbot, handled the refund request, and declined it. An agent that issues refunds itself removes that step. What it gets wrong is not a statement to argue about later. It has already happened.
What can go wrong when an AI agent can issue refunds
An over-broad tool. OWASP's Top 10 for Agentic Applications uses this case under ASI02 Tool Misuse: "A customer service bot intended to fetch order history also issues refunds because the tool had full financial API access."
A persuaded agent. Customers argue, and some argue well. Suppose a customer insists the policy allows a full refund on an opened item, quotes a promotion that ended last year, and stays polite and persistent. A model that weighs the conversation can talk itself into agreeing.
An account takeover dressed as a support request. Someone claiming to be the account holder asks to change the email address, then resets the password through the new one. The agent saw a reasonable request. The attacker now owns the account.
Duplicates and splits. A refund call that times out and is retried can pay twice. Ten refunds just under an approval threshold add up to one large one.
Five guardrails for refunds, credits and account changes
1. Value limits enforced by the payments system
The limit belongs in the system that moves the money, not in the prompt. Set it per transaction, per customer per period and per agent per day. Give the agent a tool that can only refund an order it has looked up, up to the amount paid, to the original payment method.
That is OWASP's Excessive Agency (LLM06:2025) guidance applied: minimise what each tool can do and what it may touch. LLM06 also recommends rate limiting, to reduce the number of undesirable actions "that can take place within a given time period", which is what catches split refunds. The general case is in why the model shouldn't decide what it's allowed to do.
2. Verification before any account change
The agent should never be the one to decide that a caller is who they say they are. An identity system should tell it, and the bar should rise with the action. OWASP's AI Agent Security Cheat Sheet calls for step-up authentication for critical actions, naming account recovery, payment initiation and privilege changes.
For contact details, an older rule still holds. OWASP's Transaction Authorization Cheat Sheet says a change of authorisation token should be authorised with the current one: when a user changes the phone number for SMS codes, the code goes to the current number. By the same logic, a new email address should be confirmed through the old one.
3. Eligibility computed from the written policy
Most refund rules are data: days since delivery, whether the item came back, what was paid, how many refunds the customer has had. Compute eligibility from the order record and the policy, outside the model, and let the agent explain the result rather than decide it.
Some decisions need judgement, such as a goodwill credit or an exception for a long-standing customer. Check those against the policy text as written. A justification that relies on a rule the policy does not contain is a fail, however fluent it sounds.
4. Escalation that hands over the whole case
Escalate above the limit, when verification fails, when eligibility is unclear and when the customer disputes the outcome. The person who picks it up should see the conversation, the order, what the agent proposed and why, not a one-line summary.
Bind the approval to that exact action. The AI Agent Security Cheat Sheet asks for the approval record to include the actor, tool, target, normalised parameters, timestamp and expiry, and says changes to the target or parameters need a new approval. Without that, an approval for a £20 credit could be spent on £200. Route only the cases that need a person, because approvals that arrive by the hundred stop being read.
5. A record of every action
For every refund, credit and account change, keep who asked, how they were verified, what the agent proposed, which limit and policy version applied, who approved it and what the payment or identity system returned. The cheat sheet's list of decision metadata for high-risk actions covers most of it, down to "approval identifier, execution result, and policy version". Our post on agent audit trails covers how long to keep it.
Make the action idempotent too. Stripe's API, for instance, supports idempotency keys so that a request repeated after a connection error does not perform the same operation twice.
Suppose a retailer's settings looked like this:
| Action | Enforced limit | Verification | Escalate when |
|---|---|---|---|
| Refund | Up to the amount paid; caps per customer and per agent | Signed-in account holder | Over a cap, outside the policy window, disputed |
| Goodwill credit | Small fixed ceiling | Signed-in account holder | Above the ceiling, or a repeat credit |
| Delivery address change | One change per order | Step-up authentication | Order already dispatched |
| Email or phone change | Never by the agent alone | Confirmed through the current contact | Confirmation fails |
A checklist before your support agent can move money
- Could the refund tool pay more than the order's value, or to a different payment method?
- Where is each limit enforced, and could the agent exceed it by asking several times?
- Does any account change depend on the agent's own view that the customer is genuine?
- Is eligibility computed from data, or decided by the model?
- Is each approval tied to one action and one amount?
- Could a retry pay twice?
Where Syntrox fits
Syntrox sits outside the agent. Our judge, Nucleus, applies domain rubrics, fintech among them, that name specific violations and their severity, and returns pass, fail or uncertain with a short reason that quotes the fragment that drove the verdict. When a grounded check has no policy text to check against, it returns uncertain rather than guessing.