- Yes, it happens: language models sometimes guess when unsure and write statements that sound plausible but are wrong. On a store that looks like a return window or a delivery date the store never gave.
- A business can be held responsible for its chatbot's answer: in Moffatt v. Air Canada (2024), a Canadian tribunal ordered the airline to pay 812.02 Canadian dollars over its chatbot's wrong answer. It is one decision, and this is not legal advice.
- Shopify says it about its own tools too: you're responsible for the accuracy of what the Shopify Inbox agent tells your customers.
- Answers from your store's own documents lower the risk but don't remove it. Put exact facts on the page, read every answer, and pass refunds and health questions to a person.
- Fixed answers you have read can't improvise. Open-ended chat answers more questions, but it needs limits and a hand-off to a person.
Can an AI chatbot give a shopper a wrong answer?
Yes. A language model, the technology behind AI chatbots, can write an answer that sounds right even when it doesn't have the fact. This is called a hallucination. In a 2025 paper on why it happens, the researchers write that such models "sometimes guess when uncertain, producing plausible yet incorrect statements instead of admitting uncertainty."
According to the paper, hallucinations persist even in state-of-the-art systems, and the authors argue it is because the way models are trained and evaluated rewards guessing over admitting uncertainty. Source: Kalai and others, Why Language Models Hallucinate, arXiv, September 2025.
On a store it can look like this: a shopper asks whether they can return an item after a month, and neither the page nor the policy says how many days returns are open. Instead of saying it doesn't know, the bot may write a number that sounds normal, and the shopper buys on that number.
Who is responsible when the chatbot is wrong?
The business, according to a 2024 decision in Canada. British Columbia's Civil Resolution Tribunal ordered Air Canada to compensate a passenger who relied on a wrong answer from the chatbot on its website. What happened:
- In November 2022, Jake Moffatt's grandmother died, and Moffatt asked the chatbot on Air Canada's website about bereavement fares.
- The chatbot answered that the reduced fare could be requested even after travel, within 90 days of the date the ticket was issued.
- The airline's Bereavement travel page, which the chatbot's own answer linked to, said the opposite: the policy doesn't apply to requests after travel is completed.
- The passenger flew at the regular fare and applied within the 90 days. The airline refused, and one of its representatives admitted by email that the chatbot had provided "misleading words."
- Before the tribunal, Air Canada argued it couldn't be held liable for information the chatbot gave. The tribunal wrote that, in effect, the airline was suggesting the chatbot is a separate legal entity responsible for its own actions, and called that "a remarkable submission." The chatbot, it wrote, "is still just a part of Air Canada's website."
- The tribunal held that the airline did not take reasonable care to make sure its chatbot was accurate, and found negligent misrepresentation. It ordered the airline to pay 812.02 Canadian dollars: 650.88 in damages, plus interest and tribunal fees.
Two points in the decision are worth knowing for any store. The tribunal wrote: "It makes no difference whether the information comes from a static page or a chatbot." And it found it reasonable for the passenger to trust the chatbot, since there was no reason the passenger should know that one part of the website was accurate and another was not.
Keep it in proportion. It is one small-claims decision in one Canadian province, and the decision notes that Air Canada gave no information about what kind of chatbot it used. This page is not legal advice: ask a lawyer about the rules where you sell. Source: Civil Resolution Tribunal of British Columbia, Moffatt v. Air Canada, 2024 BCCRT 149, decided 14 February 2024.
And what does Shopify say?
It says the same about its own tools. On the Shopify Inbox agent: "You're responsible for the accuracy of the information that the Inbox agent provides to your customers." On AI-suggested instant answers, it says you're responsible for the accuracy of all the content you publish to your store, and that it's important to read all generated content closely before publishing. Sources: Shopify Help Center, the Shopify Inbox agent, instant answers, checked 26 October 2026.
What can go wrong on a store?
A bot can get a number or a condition wrong: return days, delivery times, sizes, ingredients. The answer sounds plausible, so the mistake can go unnoticed until the shopper asks for what they were promised. The rows in the table are examples, not real cases:
| Question | Example of a wrong answer | Where the right answer lives |
|---|---|---|
| Can I return it after a month? | The bot says 60 days and your policy says 30. Or it says yes for an opened item when you take back only unopened ones. | Your refund policy |
| Will it arrive by Friday? | The bot promises a date your carrier can't meet, or shipping to a country you don't ship to. | Your shipping policy, with delivery times in days for each region |
| Does it run small? | The bot answers by what is usual for other brands, not by this product's cut. | This product's size chart |
| Is it gluten free? | The bot says yes, because similar products are. Here a mistake can hurt someone's health, not just cost a refund. | The ingredients and allergen list on the product page |
| Is there a discount code? | The bot makes up a code, or an offer that has already ended. | The offers running now |
The Air Canada chatbot's mistake was the same kind as the first row: a policy with a time limit, stated wrong, with the right page one click away.
Do answers from your own store data prevent it?
They lower the risk, but they don't remove it. In a preregistered 2024 study, researchers tested AI tools for legal research. These tools look up documents before they answer, and some of the companies selling them had presented the method as eliminating or avoiding hallucinations. The tools hallucinated less than a general-purpose chatbot (GPT-4), but still hallucinated 17% to 33% of the time, depending on the tool.
It is a study of legal questions, not shoppers' questions, so the number doesn't carry over to a store as it is. The lesson does: an AI that reads your documents can still get them wrong. Source: Magesh and others, Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools, arXiv, May 2024.
And an AI can only stick to facts that are written down somewhere. Shopify writes that the Shopify Inbox agent answers from your store's products, policies, pages and knowledge base, and that "keeping this content accurate and complete improves the quality of responses." The same page says the agent can also cite web search results, as a secondary source, for product-adjacent content, and that it doesn't present generic web results as if they were your store's knowledge. Source: Shopify Help Center, the Shopify Inbox agent, checked 26 October 2026.
How do you keep AI answers to your own facts?
Write the facts down first, read every answer before shoppers see it, and hand the questions that can't afford a mistake to a person.
- Write the facts on the page and in your policies. How many days returns are open, delivery times for each region, measurements, ingredients. A number that isn't written anywhere is a number the AI may guess.
- Read every answer before it goes live, and check each number against the policy. This is what Shopify recommends for AI-suggested instant answers.
- No promises your policy doesn't make. One fact per answer, an exact number, and the condition in the same sentence: 30 days, unused items only. How to write answers like these: product page FAQ.
- Decide what the bot doesn't answer alone. A refund on a specific order, a health question, anything legal: those go to a person. The Shopify Inbox agent, for example, can hand complex or sensitive questions to your staff, and when no one is available it gives the customer your store's email address.
- Tell shoppers it's AI. In Shopify Inbox, customers see a notice that they're chatting with AI, and it can't be removed.
- Test it the way a shopper would. Before you turn it on, ask the hard questions yourself: after 30 days, an opened item, a country you don't ship to. Ask again after every policy change.
- Read the real conversations every week. In Shopify Inbox, rating a reply or suggesting a better one is feedback for future training, and Shopify notes that it doesn't change the message the customer already saw. So if a reply was wrong, write to the customer yourself.
- Change the page and the answers on the same day. When a policy changes, an answer written from the old one is now wrong.
Sources: Shopify Help Center, instant answers, the Shopify Inbox agent, training the agent, checked 26 October 2026.
An example, for an invented store whose policy gives 30 days from delivery to return an unused item in its original packaging:
Can I return it if I've used it?
Only unused items can be returned, in their original packaging, within 30 days of delivery. An item that has been used can't be returned.
A bot that guesses could have said yes, and the shopper would have bought on it. The right answer gives the number and the condition from the policy, and nothing more.
Fixed answers or open-ended AI chat: which is safer?
Fixed answers you have read can't improvise, so they hold fewer surprises. Open-ended AI chat answers far more questions, but each reply is written in the moment, so it needs limits and someone to hand the conversation to.
| The difference | Fixed answers | Open-ended AI chat |
|---|---|---|
| What the shopper sees | Answers you wrote or approved, the same for every shopper | A new reply written for each message |
| When you can check it | Before any shopper sees them | Afterwards, in the conversation history |
| Which questions | Only the questions you chose in advance | Almost any question, including ones you didn't think of |
| A follow-up question | Goes to a person, or goes unanswered | The bot answers it too |
| What it needs from you | Writing the answers and keeping them up to date | Limits, a hand-off to a person, and reading the conversations regularly |
| Examples | Instant answers in Shopify Inbox, an FAQ section on the product page, an FAQ popup | The Shopify Inbox agent, and other AI chat apps |
Neither way is safe on its own. A fixed answer can be wrong too if nobody read it, and it can't answer what you didn't think of. In Shopify Inbox you choose between the two: when the agent is off, shoppers see your instant answers and your staff answer follow-up messages, and when the agent is on, it handles the replies. Source: Shopify Help Center, instant answers, checked 26 October 2026.
All the ways, side by side: ways to answer shoppers. And when chat pays off for a small store: is live chat worth it.
Where BOOX fits
BOOX is on the fixed-answers side: it writes three questions and answers for every product, and you can read and edit each one. We make BOOX, so here are the facts, including when it doesn't fit.
- What it does: BOOX writes three questions and answers for every product, with AI, from the product page and the store's policies. You can edit any of them.
- You read these answers too. They are written with AI, so the same rule applies: go through each product's answers, check every number against your policy, and fix what's wrong. If a fact isn't on the page or in a policy, write it there first.
- Shoppers click a question, or type their own, and never give BOOX personal details. Every answer ends with an add-to-cart button that takes the shopper to the store's own buy button.
- Your data shows, per product and per question, how many saw the popup, how many answered, and how many bought after answering.
- Free for up to 50 answers a month. Premium, unlimited: $10 a month, with a 14-day free trial.
When it doesn't: when shoppers need an answer about their own order, like where the parcel is or a refund on a specific order. That takes chat or email with a person. When shoppers ask many different questions about a product, more than three fixed ones can cover, an open-ended chat with a hand-off to a person will cover more. And if you don't have time to read the answers that get written, no tool that writes with AI is a fit, BOOX included.