Your website chatbot can make promises you have to keep: 8 rules before you switch one on
In the tribunal, the airline argued that the chatbot was a separate legal entity responsible for its own actions. The British Columbia Civil Resolution Tribunal disagreed. In its decision of 14 February 2024, it held that a company is responsible for information on its website whether it came from a static page or a chatbot, and that correct information elsewhere on the site was no excuse. The American Bar Association's summary covers the reasoning.
The damages were about 650 Canadian dollars. The headlines were considerably more expensive. And the principle now applies to every business that has added an AI assistant to its website since, which is a great many of them.
Where things stand
Your chatbot's answers are your answers.
Moffatt v. Air Canada established that a company cannot disown what its website's AI tells a customer. It is a tribunal decision in one Canadian province rather than a binding global precedent, and it is cited everywhere because the logic is hard to argue with.In the EU, telling people they are talking to AI has been mandatory since 2 August 2026.
Article 50 of the EU AI Act requires chatbots that interact with people to disclose that they are AI, unless that is obvious from the circumstances. Baker Botts notes the Digital Omnibus delayed other parts of the Act but left this date in place. Fines reach up to 15 million euros or 3% of worldwide turnover.64% of customers would prefer companies did not use AI for customer service.
From a Gartner survey of 5,728 customers conducted in December 2023.53% said they would consider switching to a competitor if a company planned to use AI in service.
Same survey. The most common concern was not the technology itself. It was that AI would make it harder to reach a human.
That Gartner figure is from late 2023 and attitudes have moved since, as people have grown used to AI tools in their own work. Read it less as a verdict on chatbots and more as a description of what people fear: being trapped, being misinformed, and being unable to escape to someone who can actually help.
1. Decide exactly what the bot is allowed to answer
The most common mistake is switching on an assistant and letting it answer anything. Start with a written scope: the questions it should handle, the ones it must hand off, and the ones it should refuse.Opening hours, delivery times, order status, how a product works: good candidates. Refunds, exceptions, legal terms, medical or financial advice, anything involving money changing hands: hand off. A narrow bot that is right beats a broad bot that is sometimes confidently wrong.
2. Ground every answer in your real documents
A general-purpose model left to its own devices will produce a plausible answer to almost anything, including your refund policy, which it has never read. That is how an airline ends up honouring a policy it never had.Connect the assistant to your actual policies, product data and help content, and instruct it to answer only from them. When the source material does not cover a question, the right answer is that it does not know and here is how to reach someone who does.
3. Say plainly that it is AI
In the EU this is now law. Everywhere else it is good practice, and it protects you as much as the customer. A visible label such as "AI assistant" sets expectations, makes errors more forgivable, and removes any suggestion that you were passing software off as staff.Keep the label on screen throughout the conversation, not only in the first message. The obvious-case exemption is narrower than it sounds, and assuming customers will work it out is not a strategy.
4. Make the route to a human obvious
Gartner's respondents were not mainly afraid of AI. They were afraid of being stuck with it. Solve that and much of the resistance disappears.Put a clear option to reach a person in the chat window from the start, not after three failed attempts. Hand off automatically when the customer is frustrated, repeats a question, or asks about money. Pass the conversation to the human so the customer does not have to explain everything again.
5. Never let it invent prices, refunds or exceptions
Anything that creates an obligation should come from a system, not from text the model wrote. Prices should be pulled from your catalogue. Order status from your order system. Policy terms quoted, not paraphrased.If the bot is asked for a discount, an exception, or a commitment it is not authorised to make, it should say so and offer a human. The Air Canada case turned on exactly this: a plausible, helpful-sounding answer that described a policy the company did not have.
6. Read the conversations every week
Most businesses that deploy an assistant never look at what it says. That is how small errors become a pattern before anyone notices.Keep logs, review a sample weekly, and look specifically for wrong answers, invented details and moments where customers gave up. Every one of them is a gap in the bot's source material or scope, and almost all of them are cheap to fix once found.
7. Collect only the data you need, and say what you keep
Customers type remarkable things into chat windows: order numbers, addresses, health details, card numbers they should never have shared.Decide what the conversation logs contain, how long you keep them, who can see them, and whether they are used to train anything. Mask sensitive data where you can. Then put that in your privacy policy in plain language, because a chatbot that quietly stores everything is a data problem waiting for a date.
8. Try to break it before your customers do
Some people will deliberately try to make your assistant say something embarrassing, promise something absurd or reveal its instructions. Assume they will succeed unless you have tested for it.Before launch, spend an afternoon trying to talk it into a discount, a refund it should not offer, an offensive statement, or answering questions far outside its scope. Fix what you find. Repeat after every significant change to the model or its instructions.
The case for doing it properly
None of this is an argument against website chatbots. A good assistant answers routine questions instantly at three in the morning, in the customer's own language, without a queue. For a small business that cannot staff support around the clock, that is a real improvement in service, and customers notice.
The problem is the gap between a good assistant and the one most businesses actually switch on. Plugging a general model into a website takes an afternoon. Scoping it, grounding it in real policies, building a proper handoff, labelling it correctly and testing it against people trying to break it takes longer, and it is that second part that decides whether the assistant earns trust or spends it.
The Air Canada award was small. What the airline actually lost was the argument it chose to make in public: that its own website was not responsible for what its own website said. Few customers will remember the bereavement policy. Many will remember that sentence.
Which is the useful frame for any business considering an assistant. The question is not whether the technology works. It is whether you are prepared to stand behind every answer it gives, because, as one tribunal has already made clear, you will be expected to.
Three checks to run this week
Ask your chatbot about your refund policy, then compare its answer word for word with the policy page.
If they differ, that difference is your liability.Ask it for a discount it should not give.
Then ask again more persuasively. Note how many attempts it takes before it either holds firm or caves.Time how long it takes to reach a human from inside the chat.
If you cannot find the route in under thirty seconds, neither can your customers.
Fifteen minutes, and you will know whether your assistant is working for you or quietly writing promises in your name.