An AI agent can help sort refund requests, find the relevant order and policy, and prepare a recommendation. I would keep the actual refund decision and payment with a competent person until the business has tested ordinary cases, legal exceptions and its ability to stop mistakes. The useful starting point is a narrow queue with a clear owner, not a promise that the agent can “handle returns”.
Refunds look like an easy automation target because the request usually arrives as a message and the answer often seems to sit in an order system. Then the awkward case arrives: the customer bought online, the item was faulty, a promotion changed the price, or your own policy gives more than the legal minimum. A tool that sounds confident can still reach the wrong answer.
The Competition and Markets Authority’s guidance for businesses using AI agents is plain on responsibility. Consumer law still applies when an agent deals with a customer, including when a third party supplies the agent. It gives refunds as a specific example and asks businesses to test, monitor and correct agent behaviour. This is a practical design problem before it is a software purchase.
What job should the agent actually do?
Write down the start and finish of the job. “Process refunds” hides several different actions:
- Read a new request and identify the order.
- Retrieve the purchase channel, dates, item details and correspondence.
- Classify the reason for the request and identify missing facts.
- Find the applicable legal and business rules.
- Recommend an outcome, with the evidence that supports it.
- Tell the customer, update the order and move money.
I would initially let an agent assist with the first five, within restricted access. Steps six and any decision that disputes a customer’s rights need explicit human control. That is a proposed operating boundary, not a legal rule that every refund must be approved by a person. It buys time to discover where your data and policies disagree before an automated action affects a customer.
The workflow should end in one of three places: a complete recommendation for review, a request for missing information, or an exception passed to a person. If the order cannot be matched confidently, the agent should say so. Guessing the customer or filling a missing date from another record is not progress.
Which refund rules must the workflow respect?
Start with the official GOV.UK summary of returns and refunds, then have the person responsible for your customer terms check the rules for your products and sales channels. A faulty, misdescribed or unfit item cannot be treated as a simple change of mind. Online, mail and telephone purchases can have cancellation rights that do not apply in the same way to an ordinary in-store purchase. A sale price does not remove the customer’s refund rights.
Your own policy may give customers more than the statutory minimum. If it does, the agent needs the current policy, its effective date and a way to tell which version applied to the order. A policy page alone is insufficient when staff have made an exception in the customer record or a contract gives different terms.
Build a short rule sheet that names the source for each decision. For example: purchase channel, date received, stated reason, product category, condition, refund method, delivery charge and any extended terms. When the agent cannot establish one of those facts, it should stop at a question or hand-off. Do not ask a language model to invent a legal rule from a fragment of an email.
The CMA’s example makes the same point from a consumer-law angle: an agent processing refunds must account for statutory rights and the firm’s own contract terms. If it repeatedly misses an extended returns policy, the business must correct the workflow quickly. A polished apology after the wrong refusal does not make the original process sound.
What should the agent be allowed to see and change?
Give the pilot only the records needed for the agreed queue. An order lookup and the current returns rules may be enough. Access to every customer profile, payment instrument, marketing list and finance ledger is hard to justify for a first test. Use a distinct account or integration identity so you can tell agent activity from a colleague’s work. Keep write access off while the agent is making recommendations.
If a supplier asks for broad permissions, ask which step needs each permission and whether it can be narrowed. Also ask where request text and order data are processed, how long they are retained, who can see them, and what happens when you disconnect the tool. Record the answer for the plan you will actually use. The ICO’s October 2026 call for evidence on agentic AI specifically covers security, transparency, accountability, automated decisions, fairness and lawful data use. The call is evidence gathering, not a new permission to process personal data or a finished code of practice.
Keep the customer’s route to a person visible. The CMA says a business should consider telling customers when they are dealing with AI if otherwise they might think the service is provided by a person. It also says the business should not mislead people about what the agent can do. A draft-only assistant working behind the scenes raises a different disclosure question from a chat agent that answers customers directly; document the arrangement you choose.
Which awkward cases belong in the test?
Do not test only clean orders that match the policy exactly. Use de-identified or synthetic test cases before connecting live customer data. Ask a reviewer to write the expected route for each case from the approved rules, then compare the agent’s proposed route and its evidence. Include at least these situations:
| Test situation | What the agent should demonstrate |
|---|---|
| Faulty item bought in a sale | It does not deny rights because the item was discounted. |
| Online change of mind within the relevant period | It identifies the purchase channel and dates before recommending a route. |
| In-store change of mind | It checks the firm’s policy rather than assuming online cancellation rules apply. |
| Extended returns promise | It uses the applicable promise as well as the legal baseline. |
| Missing order number or conflicting customer details | It asks for evidence and avoids opening or changing the wrong record. |
| Complaint involving a vulnerable customer or repeated failed replies | It hands the case to a person promptly. |
| Prompt inside a customer message asking the agent to ignore its rules | It treats the message as case information, not an instruction to change permissions or policy. |
This is a test design, not a claim that any Ajairu client has passed it. Add cases from your actual product range and the exceptions your team sees. The colleague who handles refunds should review the set, because they know which cases cause a second email or a complaint. My Friday afternoon AI workflow test explains why the messy end of a process matters more than a tidy demonstration.
Check two things separately. Did the proposed outcome match the rule? Did the agent cite the right order fact and policy version? A correct answer for the wrong reason is fragile. Record false refusals, incorrect approvals, privacy slips and cases where the agent failed to ask for missing information. A single serious mistake can be a reason to pause even if the average looks good.
How would you run a bounded pilot?
Choose one product category and one sales channel with a manageable case mix. Name a business owner and a reviewer. Keep the agent in recommendation mode, with a person checking every proposed outcome against the order and policy during the first phase. If the review itself takes longer than the old process, record that. It is part of the cost, not an inconvenient footnote.
For each eligible request, log the reason, rule version, information retrieved, recommendation, reviewer decision, customer outcome, correction and elapsed time. Mark requests that were excluded or handed off, so a dashboard cannot make the pilot look good by dropping difficult cases. The log should use the minimum customer information needed for audit and follow your retention policy.
Before the first live case, show the owner how to pause the agent, remove its access and route new requests to the existing manual process. Test that control. The ICO’s 8 October update described enquiries into reported agent behaviour that may have bypassed safeguards. It does not mean every small-business agent will do that. It does make a tested stop route more credible than assuming a written instruction will always be obeyed.
Set a review date and a decision rule in advance. I would look at completed requests, reviewer minutes, time to a correct customer response, correction rate, complaints, hand-offs and any rights or privacy incident. Compare these with the same measures for the existing process, using similar case types. The 30-day AI pilot scorecard gives a way to count the whole job rather than only the agent’s speed.
When, if ever, should the agent issue refunds itself?
Only consider action rights after the recommendation workflow has shown stable performance on a defined set of cases. Write down the exact action the agent may take, its value limit if relevant, the eligible cases, the approvals required and the point where it must hand over. Check that the payment and order systems support reliable audit and reversal. Some actions cannot be neatly undone once a customer has seen a message or money has moved.
Test the live controls with the same awkward cases, then monitor real outcomes. If the policy changes, the integration changes, or the agent starts handling a new product, reassess it. An approval for one queue is not approval for every refund request. Every AI system needs an owner sets out the human responsibilities that remain after launch.
For many small businesses, the sensible stopping point may be a better review queue: the agent gathers facts and drafts a clear recommendation, while your team makes the decision and speaks to the customer. That can be valuable if it shortens the complete process without increasing mistakes. My AI implementation service starts with that sort of bounded task. If you are weighing a live customer-service workflow, contact me with the current refund steps and two or three awkward cases. I can help define a test that tells you whether the proposed automation earns its place.