Why chatbots invent return policies, and how to stop it

Of all the mistakes an AI assistant makes, the invented policy is the one that costs most. A wrong delivery estimate annoys a customer. A return window that doesn’t exist, a free return that isn’t free, or a refund promise the store never made turns into an argument, and the customer has a screenshot.

It happens even on well-known platforms with good settings, and the cause is almost always the same.

What it looks like

A made-up electronics shop sells laptops, phones and accessories. Its returns page says:

You can return most items within 14 days of delivery. Opened software, SIM cards and earphones can’t be returned for hygiene and licence reasons. Laptops must be returned with all accessories and in their original box.

A customer asks the assistant: “I opened my earbuds yesterday and don’t like the sound. Can I send them back?”

The assistant replies: “Yes, you can return items within 30 days of delivery for a full refund. Please make sure they are in their original packaging.”

Two things are wrong. The window is 14 days, not 30. And earphones can’t be returned at all. Neither “30 days” nor “full refund” appears anywhere on the site.

Where the made-up rule comes from

Language models learned from a vast amount of text, including thousands of real return policies. “30 days, full refund, original packaging” is the most common pattern in that text. When the assistant’s lookup doesn’t bring back the right piece of the store’s own policy, or brings back only part of it, the model fills the gap with the most likely policy in general.

The common triggers:

1. The exception lives in a separate sentence. The lookup returns “You can return most items within 14 days” but not the next sentence about earphones. The model sees a general rule and applies it.
2. The customer’s words don’t match the page. The customer says “earbuds”, the page says “earphones”. The lookup misses the exception entirely.
3. The returns page isn’t connected at all. It happens more often than you’d think: the assistant was set up on the help centre, and returns live on a page in the shop footer.
4. Instructions push for helpfulness. A system prompt like “always give a clear, confident answer” makes the model less likely to say “I’m not sure, let me hand you over”.

The platforms know this. Intercom, Zendesk and Tidio all tell their AI to answer only from your content, and all of them describe ways to restrict it further.¹² Restriction helps, but it doesn’t help with trigger 1 or 2: the model is answering from your content, just not all of it.

Why it matters more for returns than for other topics

Returns are where consumer law sits. In the EU, online buyers have a 14-day right of withdrawal for most goods, with exceptions such as sealed goods unsealed for health or hygiene reasons.³ Stores can offer more, not less. An assistant that promises 30 days creates an expectation the store either honours at its own cost or refuses and loses the customer. An assistant that says earphones can be returned when the law and the store’s terms say otherwise creates a dispute with no good outcome.

Four fixes that work

1. Put each exception in the rule’s own sentence. “You can return most items within 14 days of delivery, except opened software, SIM cards and earphones (including earbuds and headphones), which can’t be returned.” Long, but the lookup can’t separate the rule from its exceptions.

2. Add the everyday words. Earbuds, headphones, AirPods-style. Refund, money back, send back. The customer won’t use your words.

3. Write a short returns answer for the assistant. Most platforms let you add custom answers or snippets that only the assistant reads. Intercom calls them snippets, Zendesk has its own content for AI agents, Chatbase lets you add text sources directly.⁴ Write the policy in five plain sentences there, with every exception. It becomes the piece the lookup finds first.

4. Test returns questions after every change. Not “what is your return policy” — the assistant usually gets that right. Ask the hard ones:

• “I opened it, can I still return it?”
• “It’s been 20 days, is it too late?”
• “Do I pay for the return shipping?”
• “Can I get cash back instead of store credit?”
• “I bought it in the sale, does the same policy apply?”

Each one should come back with the store’s rule, not a general one. If it doesn’t, find the page sentence that should have answered it and rewrite it.

How often this happens

In our own checks of live bots, invented or distorted policies were among the most common factual mistakes. We collected the patterns, with more made-up examples, on the invented policies page. The good news is that it’s one of the easiest to fix, because the fix is in your content, which you control.

If your store already has an assistant, Assistant Care asks it questions like these every week and fixes what drifted. If you’re setting one up, start with what to put in the knowledge base.

Footnotes

1. Intercom, “Fin AI Agent explained”, intercom.com/help. ↩

2. Zendesk, “About AI agents”, support.zendesk.com. ↩

3. European Commission, Your Europe, “Guarantees and returns”, europa.eu/youreurope. ↩

4. Chatbase documentation, “Sources”, chatbase.co/docs. ↩