Field note
AI Refunds: Keep Human Approval in the Loop
Let deterministic rules calculate eligibility and require an authorised person to approve exceptions or high-value refunds.
Let deterministic rules calculate eligibility and require an authorised person to approve exceptions or high-value refunds.
Why refunds are different from other chatbot tasks
A wrong answer about opening hours costs a phone call. A wrong refund costs money, and a pattern of wrong refunds is either a fraud channel or a consumer-law problem. Refunds also sit on legal rules that do not bend to a persuasive customer: the Consumer Rights Act 2015 gives a short-term right to reject faulty goods, and the Consumer Contracts Regulations 2013 give a cancellation period for most online and distance sales. GOV.UK sets out the basics for businesses.
A language model is good at explaining those rules in plain English and bad at applying them consistently. It can be talked into a refund it should not give, and it can refuse one the customer is entitled to. Neither outcome is acceptable, and the CMA's guidance on AI agents makes clear the business, not the vendor, carries the consequences.
Three jobs, three owners
Split the refund journey into three parts and give each to the component that is good at it.
| Job | Owner | Why |
|---|---|---|
| Explain the policy and collect details | Chatbot | Language work; low risk if grounded in the published policy |
| Decide eligibility and amount | Deterministic rules in your order system | Same input, same answer, every time; auditable |
| Approve exceptions and high-value cases | A named person | Judgement, fraud awareness, accountability |
The chatbot never calculates the amount and never issues the payment. It gathers the order number, the reason and the evidence, hands them to the rules engine, and reports the result. If the rules say "exception", it tells the customer a colleague will review it and gives a timescale.
What the rules should encode
Write the rules down before anyone configures a bot. At minimum:
- Automatic refund. Within the statutory or published window, item returned or return initiated, order value under a threshold you choose, no refund on the same account in the last N days.
- Automatic partial refund. Only where the published policy already allows it, such as a missing item from a multi-item order.
- Always to a person. Faulty goods claims outside the short-term window, anything over the threshold, any request to refund to a different card or account, repeat claimants, and any message that mentions a complaint, a chargeback or a regulator.
- Never through the bot. Goodwill payments, compensation, and refunds on orders with a fraud flag.
Pick thresholds you would defend to an auditor. Many small retailers set the automatic limit at a figure their finance lead is comfortable losing to a single mistake, and review it quarterly.
A realistic example
A 15-person online homeware retailer in Bristol connects a chatbot to its order system and lets it "handle refunds". Within a month, two patterns appear. Customers learn that a firm tone plus "I know my rights" produces a refund on items well outside the window. Meanwhile, a customer with a genuinely faulty lamp is refused because the bot decided the fault was cosmetic.
The fix was not a better prompt. The retailer moved eligibility into its order system, gave the bot permission to look up orders and create a refund request only, and routed anything the rules could not decide to the customer service lead. Refund value fell, disputes fell, and the faulty-lamp case was resolved in a day because it reached a person immediately.
Approval that means something
An approval step is only a control if the approver can see what they are approving. Show the order, the customer's history, the reason, the rule that sent the case for review, and the amount. Log who approved, when and why. Give the approver a way to send it back to the customer with a question rather than a yes or no.
Do not give the bot a permission it must be asked not to use. If the bot cannot call the refund function above the threshold, no conversation can move it past that line. The AI agents and UK consumer law note has the wider pre-launch test pack for customer-facing agents.
Monitor after launch
Track weekly: refunds issued automatically, refunds sent for review, review outcomes, refund value by channel, and complaints that mention the bot. A rise in automatic refunds after a prompt or model change is the signal to pause and check. Set a stop condition in writing, for example "all refunds go to human review if automatic refund value exceeds twice the monthly average".
What to write into policy
- The chatbot explains refund policy and collects details. It does not decide eligibility or issue payments.
- Eligibility is decided by documented rules maintained by the finance or operations lead.
- Exceptions and refunds above the threshold require approval by a named role, with the decision logged.
- Refunds to a different account or card are never automated.
- Thresholds and rules are reviewed quarterly and after any complaint about an automated refund.
The customer support AI workflow page describes how we set the approved sources, rules and handover points for a small support team.