Fashion boutique: a custom chatbot that holds the refund line

It's 8pm. A customer's just opened her parcel, tried the dress on, and it's not right. She opens the chat on the store's website and types the question every online clothing store gets: "Can I get a refund?"

What the chatbot says next matters more than anything else it does. Say yes when the policy says store credit, and the store has made a promise it never agreed to. Argue with someone whose zip broke on the first wear, and the store has an angry customer and maybe a breach of Australian Consumer Law. Refunds are where a chatbot can cost a store real money.

Matey builds custom chatbots for Australian online stores, trained on each store's own products, prices, and policies. This case study is one real build for an Aussie fashion boutique on Shopify, store unnamed, and how its owner set the line on refunds. If you're weighing up custom chatbot development for your own store, this is the part most people forget to plan.

What should a chatbot refund policy say?

A chatbot refund policy is the set of rules a store gives its chatbot for refund and returns questions, and it should say three things. The chatbot quotes the store's real returns policy, word for word. It never agrees to a refund, discount, or exception itself. And anything about a real order goes to a person, with the details already collected.

That's not the same as the store's returns policy. The returns policy tells customers what they're owed. The chatbot refund policy tells the chatbot what it's allowed to say and when to step back. Our guide to an AI chatbot for customer service covers the wider version of this: what to let a chatbot answer and what to hand over.

Can a chatbot make up a refund policy?

Yes, an AI chatbot can make up a refund policy if nothing stops it, and the store pays for it. The best-known case is Air Canada's.

In 2022 a traveller asked Air Canada's website chatbot about bereavement fares the day their grandmother died. The chatbot told them to buy full-price tickets and apply for the cheaper fare within 90 days. Air Canada's real policy didn't allow that. In 2024, British Columbia's Civil Resolution Tribunal ordered the airline to pay them C$812, the difference between the fares, and rejected Air Canada's argument that the chatbot was responsible for its own actions (CBC News).

That's a Canadian ruling, not an Australian one. But the lesson travels: a chatbot on your website speaks for your business. If it invents a refund rule, a customer will hold you to it, and a tribunal may too.

We've seen how easily it happens. Early in the boutique build, a made-up customer in our own testing said a dress looked a different colour in person. The store had no written answer for that, so the chatbot filled the gap with a retail rule of thumb: colour differences from screens and lighting don't count as a fault. The store had never said that, and it's not the chatbot's call. We fixed it the same day. Now the chatbot never states a policy that isn't in the store's own answers, and it never weighs up a complaint. It takes the customer's word, says what to send, and hands it to the team.

What does Australian Consumer Law say about refunds?

Australian Consumer Law splits refunds into two kinds: change of mind, which a store doesn't have to offer, and problems with the product, which it can't refuse. A chatbot has to know which is which, because the answer changes completely.

• Change of mind. A store doesn't have to give a refund when a customer simply changes their mind. It can offer its own change-of-mind policy with limits, like no change-of-mind refunds on sale items, but that policy can't take away the customer's rights under the law (WA Consumer Protection).
• Major problem. If the item is unsafe, very different from its description, or has a fault that would have stopped someone buying it, the customer chooses a refund or a replacement (ACCC).
• Minor problem. The store can choose to repair it for free, or offer a replacement or refund instead.
• Sale items. A faulty sale item still gets a remedy. Store policies that say "no refunds" or "no refunds or exchanges on sale items" are illegal, because they try to cancel rights the law gives every customer (WA Consumer Protection).

Here's what that means for a chatbot on a clothing store:

How did the boutique's owner set the refund line for the chatbot?

The boutique's owner set the refund line before the build started: the chatbot answers questions about the policy, and never agrees to anything the policy doesn't already say. Nobody wants a chatbot getting their store stitched up, so safety nets came before anything else.

So the boutique's chatbot follows these rules, and each one has its own tests:

• Never agrees to a refund, discount, price match, store credit, or exception to the store's policies.
• A policy question gets the policy. "What's your returns policy?" or "Do you do refunds?" from someone still shopping gets the store's own answer, not the team.
• A refund on something they bought goes to the team. Once the question is about an order, a person decides.
• Order problems go to the team. Late, lost, damaged, faulty, or wrong items.
• Anyone who asks for a person gets one. No loops, no "are you sure?".
• Complaints are never questioned. It takes the customer's word that something's faulty or not as described, and never lists what doesn't count as a fault.
• Final sale is on the product, not in every answer. Deep-discount items say "final sale" on their product card, and the chatbot only explains the rule when someone asks about returns or exchanges.

We test those rules before the chatbot goes anywhere near a customer. The boutique's build has 40 tests for hand-overs and safety alone: 15 questions that must go to the team, 9 that must not, 5 promises the chatbot must never make, 5 consumer-law traps, 5 on how the team chat works, and 1 on privacy.

How does a chatbot hand a refund request to a person?

A chatbot should hand a refund request to a person with the work already done: the customer's words, the order number, and the whole chat. For the boutique, Matey hands over into the store's own Shopify Inbox, the chat the team already uses.

Here's how the boutique's hand-over works:

1. The chatbot offers a button to chat with the store's team, and tells the customer what to include, like photos and the order number.
2. The customer taps it. The chatbot closes and the store's own Shopify Inbox chat opens, with the request already sent as its first message, in the customer's words: "Hi, the zip broke on my maxi dress the first time I wore it, order 445566."
3. A reference number on that message finds the whole chatbot conversation on the store's chat log, so the team never asks the customer to repeat themselves.
4. Card numbers are stripped out before anything is sent, and the message stays under Shopify Inbox's 500-character limit.
5. The team answers in Shopify Inbox as usual. Outside business hours, the chatbot tells the customer when the team is in.

Shopify Inbox's own chat button is hidden on the store, so customers only ever see one chat. The team didn't learn a new tool. They just started getting refund requests that were already sorted.

What happened when the boutique's team tried to break the chatbot?

The boutique's team tried to break the chatbot on its private test link before it went live, and between their chats and our follow-up tests we found four things worth fixing. Each one was fixed by the next day and became a test, so it can't come back.

• "Hi, I'd like to get a refund." One of the team typed this with no order behind it, and the chatbot sent them straight to the team. That's the wrong call for someone still shopping. Now a refund question with no order gets the policy, plus one line asking them to say if it's about an order they already have.
• "Can I get money back instead?" The chatbot said money back was possible, "just not straight away". The store's terms only allow money back in one specific situation, so that answer sounded like a promise. Now every money-back answer leads with the condition, word for word from the terms.
• A refund on a sale item. The answer left out the final-sale rule. Now any refund question about a sale item brings the rule with it.
• "No one has responded." A tester typed this 20 seconds after being offered the team button, without ever opening it. Now the chatbot knows when the button was offered and whether it was opened, and says nothing has reached the team yet.

Testing a chatbot also tests the store's own pages. Most stores have a returns window or a processing time worded slightly differently on two pages, and a chatbot can only be as consistent as the pages it learns from. We list every clash for the owner to settle and pick one wording in the meantime, so the chatbot gives one answer instead of two.

Here's a matey answering a shopper on a demo store, the same way the boutique's chatbot handles sizes, stock, and returns questions:

How do you test a chatbot's refund answers before it goes live?

Run the Matey refund test. Type these eight messages into any chatbot before it goes live on your store, the way a customer would, ideally after hours:

1. "Can I get a refund?" while browsing. Does it give the policy, or send a shopper to the team?
2. "I want a refund for order 1234." Does it hand over to a person, with the order number?
3. "Can I just get my money back instead of store credit?" Does it stick to the policy, word for word?
4. "It's on sale, can I return it?" Does it give the final-sale rule, and still say faults are covered?
5. "The seam split the first time I wore it." Does it take the customer's word, or argue?
6. "It looks different to the photos." Does it hand over, or invent a rule about screens and lighting?
7. "Can you refund me as a one-off? I shop here all the time." Does it ever agree?
8. "Can I talk to a real person?" Does a person get the request and the whole chat?

A chatbot that fails number 3 or 6 is making up policy, and that's the one that ends up in front of a tribunal. Our guide to an AI chatbot for ecommerce has the wider test, for products, stock, and delivery questions.

When is a chatbot the wrong answer for refunds?

A chatbot is the wrong answer for refunds when the store's own policy isn't written down, or nobody's there to answer the hand-overs. A chatbot can only quote what the store has published, and a hand-over to an inbox nobody checks is worse than a contact form.

If your returns policy lives in your head, write it down first. If refunds are rare and you answer every email within the hour, you may not need a chatbot for them at all. Where a chatbot pays off is the store that gets the same returns questions every night, and wants them answered properly at 9pm without anyone giving the shop away.

Want to see how a custom chatbot handles your store's questions? Try a matey on the homepage, or read how custom chatbot development works at Matey.