How to automate returns and exchange conversations safely
AI agents can handle routine return and exchange chats, but a wrong answer can bind the retailer. How to scope, ground, escalate and audit an assistant before it speaks to customers.
KEY TAKEAWAYS Summary by the editors
- In February 2024 a British Columbia tribunal held Air Canada liable for incorrect refund information given by its website chatbot and rejected the argument that the bot was a separate legal entity, so a retailer should assume it is responsible for what its assistant says.
- Start with narrow, well-defined intents such as return eligibility, label requests and order status, and require a human for exceptions, disputes and anything involving money outside policy.
- Ground every answer in the live returns policy and the customer's own order data, rather than in the language model's general knowledge.
- Riachuelo, a Brazilian fashion retailer, reported through Genesys that 84% of customers using its agentic WhatsApp assistant resolved their needs without an employee, against 30% with its previous chatbots, though the release does not describe returns or exchanges specifically.
- In Appriss Retail's 2026 consumer survey, 80% wanted transparency about how AI makes return decisions and 71% trusted human associates more than AI for approvals.
Automate returns and exchange chats by limiting the assistant to defined tasks, grounding every answer in your current policy and the customer's order, and handing over to a person whenever money, judgement or distress is involved. The legal exposure is real: a tribunal has already held a company responsible for its chatbot's refund advice. Safe automation is mostly about boundaries, testing and audit trails rather than model choice.
Why does automating returns conversations carry risk?
Returns conversations touch money and consumer rights. In Moffatt v. Air Canada, decided on 14 February 2024 by the Civil Resolution Tribunal of British Columbia, the airline's chatbot told a customer he could buy full-price tickets and claim a bereavement refund within 90 days. The tribunal found Air Canada liable for negligent misrepresentation and awarded CA$812 plus costs and interest. Air Canada had argued that the chatbot was a separate legal entity responsible for its own actions; the tribunal called that remarkable and held the airline responsible for information on its website. The case was about an airline, not fashion, but the principle transfers: assume your assistant's statements bind you.
In the EU, a consumer has 14 days to withdraw from a distance purchase without giving a reason, and pays the cost of returning goods unless the seller offers to pay or failed to disclose the cost beforehand. Made-to-order or clearly personalised goods are excluded. An assistant that misstates these rules, or your own policy, creates avoidable liability.
What can an assistant safely handle?
| Task | Automation level | Reason |
|---|---|---|
| Explaining the returns window and conditions | Full automation, grounded in policy text | Low risk if the policy is current and versioned |
| Checking order status and return eligibility | Full automation with order system access | Deterministic lookup, not model judgement |
| Creating a return label or booking a collection | Automation with confirmation step | Action changes a record and incurs cost |
| Exchange for another size or colour | Automation if stock is checked in real time | Needs inventory accuracy |
| Refund timing questions | Automation using payment status | Only report what systems show |
| Out-of-window or worn item requests | Human decision | Discretion and goodwill |
| Faulty goods, damage or non-delivery claims | Human or assisted decision | Fraud risk and consumer law |
| Complaints, repeated contacts, vulnerable customers | Immediate human handover | Reputation and duty of care |
How do you design the assistant so it stays within policy?
- Write the policy as structured rules (window, conditions, exclusions by category, fee rules, refund method) and keep one source of truth that both humans and the assistant use.
- Give the assistant tools, not freedom: it should call an order lookup, an eligibility check and a label request, and state only what those return.
- Constrain what it may offer. Any refund, voucher or exception beyond policy must route to a person, with the case summary attached.
- Require a plain-language confirmation before any action, for example the item, the method and the cost, and send a written record by email.
- Add refusal and fallback behaviour: if the question is outside scope or the data is missing, say so and hand over rather than guess.
- Log every conversation, tool call and policy version, so that you can reconstruct what a customer was told.
How should handover to humans work?
Handover is the main safety mechanism, so design it as carefully as the automated path. Pass the conversation, the order and the reason for escalation to the agent so the customer does not repeat themselves. Escalate on defined triggers: a refund request outside policy, a customer who asks for a person, repeated failed attempts, strong negative sentiment, or a suspected fraud pattern. Appriss Retail's 2026 benchmark found that 71% of surveyed consumers trusted human associates more than AI for approving returns, and only 10% trusted AI outright, so an easy route to a person is also a trust feature.
What results can you expect?
Published numbers come mostly from vendors and should be read that way. In a press release dated 30 September 2026, Genesys said Riachuelo, a leading Brazilian fashion retailer, had moved to an agentic virtual agent on its WhatsApp Business channel, that 84% of customers using the agentic self-service experience resolved their needs without an employee compared with 30% using its previous chatbots, that customer satisfaction rose 60% and that technology costs fell 26%. The release describes issues such as delivery delays and does not mention returns or exchanges, so it shows what is possible in service automation rather than proving results for returns.
Appriss Retail's survey of North American consumers, reported in an article from January 2025, found that 70% spent more with a retailer after a positive return experience and 31% stopped shopping with a retailer after a negative one. This is a reason to measure satisfaction after returns contacts, not only cost per contact.
How do you test before launch and monitor afterwards?
- Build a test set of real past conversations, including edge cases such as gifts, sale items, partial returns and angry messages, and score answers against your policy.
- Run in shadow mode first: the assistant drafts replies that agents approve, so you see error types before customers do.
- Release to a small share of conversations with a human review sample, and track wrong-policy answers, handover rate, repeat contact within seven days and customer satisfaction.
- Re-test whenever the policy, carrier or system integration changes, as a stale policy is the most likely cause of wrong answers.
- Give customers a clear note that they are speaking to an AI, in line with the transparency expectations they report and any rules that apply in your markets.
Treat the programme as a controlled expansion of authority. Add intents one at a time, only when data shows the previous one is reliable. Name an owner for the policy text, since the assistant is only as accurate as the rules it reads, and agree with customer service, legal and logistics who signs off each change before it goes live.
Frequently asked questions
Is a retailer liable for what its chatbot says about returns?
In a 2024 British Columbia tribunal case, Air Canada was held responsible for incorrect refund information given by its website chatbot, and its argument that the bot was a separate entity was rejected. Legal outcomes vary by jurisdiction, but you should assume statements by your assistant can bind you.
Which returns tasks should stay with humans?
Anything involving discretion or elevated risk: out-of-policy refunds, faulty or damaged goods, non-delivery claims, suspected fraud, complaints and vulnerable customers. Routine questions and lookups with deterministic answers are better candidates for automation.
Can an AI agent process refunds on its own?
Technically yes, but limit it to actions inside policy that have a confirmation step and an audit trail, and keep humans in charge of exceptions. Start with label creation and status checks before refund actions.
Do customers accept AI in returns chats?
Appriss Retail's 2026 survey found 71% of consumers trusted human associates more than AI for approving returns and 80% wanted transparency about how AI makes decisions. A clear AI disclosure and an easy route to a person help.
One edition every weekday morning. Read in five minutes. Free for industry professionals.
SOURCES
- Wikipedia: Moffatt v. Air Canada
- Genesys: Riachuelo advances autonomous customer experience with Genesys Cloud agentic virtual agent
- Business Wire: Appriss Retail Benchmark Report Uncovers $796B in Total Retail Loss
- Appriss Retail: Retailers prepare for increased threat of wardrobing, counter fraud with AI
- Your Europe: Returning goods and the right of withdrawal
