Returns and exchange questions are one of the largest and most repetitive categories of ecommerce support. Automating returns questions means letting an AI agent handle the policy lookups, eligibility checks, and status updates, while a human stays on the exceptions. Done carefully, it removes a slice of routine volume without making the return feel like a fight. This guide covers what to automate, what to keep human, and how to set it up.
Why automating returns questions is worth it
Returns are not a rare event. The National Retail Federation, in its 2025 Retail Returns Landscape report with Happy Returns, estimated that 15.8% of all retail sales would be returned in 2025, totaling $849.9 billion. For online sales the rate is higher. The same report put 19.3% of online purchases as returned. Close to one in five online orders comes back.
Each of those returns generates questions, often several. Can I return this. How long do I have. Where is my refund. Can I swap the size instead of refunding. The questions are predictable, they repeat across thousands of customers, and they arrive in waves after every sale and every holiday. That is exactly the profile of work worth automating: high volume, low variation, answerable from policy and order data.
There is a second reason, and it matters more than ticket count. The return is a loyalty moment. A customer mid-return is deciding whether to buy from you again. A fast, clear, accurate answer keeps them. A slow or contradictory one does not. Automating the routine part is how you answer quickly without burning out the team that handles the hard part.
What returns questions can you safely automate?
Not every returns question is equal. Some are pure lookups. Some need judgment. The first group is safe to automate today. The second is not.
| Question type | Automate or escalate | Why | |---|---|---| | "What is your return policy?" | Automate | Static policy text, no judgment | | "Is this order still in the return window?" | Automate | Date math against order data | | "Where is my refund?" | Automate | Status lookup in connected systems | | "Can I exchange this for another size?" | Automate | Eligibility check plus stock check | | "How do I start a return?" | Automate | Guided steps, then label handoff | | "My item arrived damaged" | Escalate | Needs a judgment call and goodwill | | "I am past the window, but..." | Escalate | Policy exception, human discretion | | "I want to return a final-sale item" | Escalate | Sensitive, often a complaint |
The pattern is clear. If answering the question only needs policy text and order data, an AI agent can do it accurately and instantly. If answering needs discretion, empathy, or an exception to the written policy, it should go to a person. Fit and sizing questions sit on the automatable side and matter a lot, because sizing and fit are consistently the single biggest reason customers return apparel. An AI agent that can confirm the window and check stock for the right size turns many of those into exchanges instead of lost sales.
What should still go to a human?
Drawing the line wrong, in either direction, costs you.
Three categories belong with a person. Damaged or wrong items, because the customer is already unhappy and the reply needs a tone an apology script cannot fake, plus often a goodwill decision. Policy exceptions, because the moment a customer is outside the written rule, someone has to decide whether to bend it, and that is a judgment call with revenue attached. Anything that reads as a complaint, because a customer who is angry about a return wants to feel heard by a human, not processed by a flow.
The NRF report is also a reminder that returns are not all honest. It found 9% of returns are fraudulent. You do not want an AI agent silently approving every refund request with no checks. The point of automation is to handle the routine majority fast while flagging the unusual case to a person, not to wave everything through.
A good rule: automate the answer, escalate the exception. The AI tells any customer what the policy is and where their refund stands. The human handles the cases where the policy needs to bend.
How do you set up returns automation without annoying customers?
Automating returns questions badly is worse than not automating them. A few rules keep it on the right side.
Connect real order data. A returns answer is only as good as the data behind it. An AI agent guessing at a return window from generic policy text will be wrong for half your customers. It needs the order date, the items, and the refund status from your store and payment records. Keloa's integrations connect that data so the agent answers from the customer's actual order, not a generic template.
Write the policy as one clear source. The AI answers from your content. If your return policy is spread across three pages that contradict each other, the agent will pick one and sound confident about the wrong answer. One clear, current policy page is the single highest-value thing you can fix before automating.
Make the human path obvious. A customer who wants a person should reach one in a sentence, not after an argument. The AI should hand off the moment a return turns into a complaint or an exception, carrying the full conversation so the customer never repeats themselves.
Be proactive where you can. Many returns questions never need to be asked. A clear shipping confirmation, a refund-issued notification, and an honest return window stated at purchase all remove the question before it becomes a ticket. Automation is not only answering fast. It is answering before the customer has to ask.
A flow builder helps here. You can route a straightforward refund-status question to the AI and a damaged-item report to a human automatically, by the type of return, without an agent triaging each one.
How Keloa approaches this
We built Keloa's AI agents to answer returns questions from your real data, not from a script. The agent checks the order date against your return window, looks up refund status in the connected systems, and confirms whether the size the customer wants is in stock before it offers an exchange. Every answer is grounded in your policy and the customer's actual order, and when the agent is not sure, it says so and hands off.
The escalation rules are yours to set. Damaged items, policy exceptions, and anything that reads as a complaint go to a human with the full thread attached, so the customer does not start over. For Shopify brands and other ecommerce teams, this means the routine three-quarters of returns volume gets a fast, accurate answer, and the team keeps its attention on the returns that actually need a person. Our pricing is billed per reply, so automating the routine questions lowers your cost rather than hiding it.
Frequently asked questions
Which returns questions should an AI agent handle? The ones answerable from policy text and order data: what the return policy is, whether an order is still in the window, where a refund stands, how to start a return, and whether a size exchange is possible. These are high-volume, low-variation questions an AI agent can answer accurately and instantly.
Which returns questions should go to a human? Damaged or wrong items, requests for an exception to the written policy, and anything that reads as a complaint. These need discretion, empathy, or a goodwill decision that automation should not make on its own.
Does automating returns mean approving every refund? No. Automation handles the routine questions fast and flags the unusual case to a person. The NRF found 9% of returns are fraudulent, so the AI agent should answer policy and status questions, not silently approve every refund without checks.
How does an AI agent know my return window? It needs your real order data connected, so it can compare the order date to your policy. Without that connection it is guessing from generic policy text and will be wrong for many customers. Accurate returns answers depend on the integration, not just the policy page.
Can automating returns questions actually keep customers? Yes, if it is fast and accurate. A customer mid-return is deciding whether to buy again, and a clear, correct answer protects that. An AI agent that confirms a size is in stock can also turn a refund into an exchange, keeping the sale.