AI customer support bots: liability, disclosure and a checklist

An AI support bot can answer a large share of customer chats at a fraction of the cost of a human team, but the company that deploys it owns every answer the bot gives. Three public sources show how that plays out: Klarna’s announcement of its AI assistant’s first-month numbers, a Canadian tribunal ruling that an airline was responsible for what its chatbot told a passenger, and a U.S. regulator’s report on the frustrations customers report with bank chatbots. This guide pulls out the practical lessons, explains the disclosure rule in the EU AI Act, and ends with a ten-point checklist you can use before launch. It is based on public documents and does not replace legal advice.

The upside is real

In February 2024, Klarna said its AI assistant, built with OpenAI, had handled 2.3 million conversations in its first month, equal to two-thirds of the company’s customer service chats. The company claimed the work equaled that of 700 full-time agents, that customer satisfaction was on par with human agents, that repeat inquiries fell 25%, and that errands were resolved in under two minutes compared with eleven before. It also said the assistant ran in 23 markets and more than 35 languages, around the clock. These are the company’s own figures from a press release, not an independent audit, but they show why support is one of the first places companies deploy generative AI. Our overview of how AI is transforming customer service covers the tool landscape.

The Air Canada lesson: you own the bot’s answers

In 2022, an Air Canada chatbot told passenger Jake Moffatt that he could book a full-fare flight for his grandmother’s funeral and apply for a bereavement fare afterward. When he applied, the airline said the policy actually required a request before travel, and argued that the chatbot was a “separate legal entity that is responsible for its own actions.” As the BBC reported in February 2024, the British Columbia Civil Resolution Tribunal rejected that argument and ordered Air Canada to pay $812.02 in damages and tribunal fees. Tribunal member Christopher Rivers wrote that “it makes no difference whether the information comes from a static page or a chatbot.”

The practical reading is straightforward. If your bot states a price, refund rule or deadline, a customer may rely on it, and “the bot was wrong” is not a defense that this tribunal accepted. It is one tribunal decision in one jurisdiction, so it is not binding everywhere, but the reasoning is easy for other courts to follow.

What regulators have found about chatbot experiences

The U.S. Consumer Financial Protection Bureau studied chatbots in banking and published its findings in 2023. It found that each of the ten largest commercial banks had deployed a chatbot and estimated that about 37% of the U.S. population interacted with a bank chatbot in 2022. Its analysis suggests chatbots can resolve basic questions but become less effective as problems get complex, and that complaints describe wasted time, feeling stuck, inaccurate information and extra fees, especially when people cannot reach a human. The report is about banking, but the failure patterns apply to any sector: loops that never escalate, bots that cannot see account context, and answers that sound confident but are wrong.

Disclosure: the EU rule

Article 50 of the EU AI Act says providers must design AI systems that interact directly with people so that those people are informed they are dealing with an AI system, unless that is obvious to a reasonably well-informed person. The practical effect for support teams selling to or serving EU customers is to label the bot clearly at the start of the conversation, not hide it behind a human-sounding name. Our EU AI Act explainer places this duty alongside the Act’s other requirements. Check the Act’s current application dates and any guidance with your legal counsel, because timelines have shifted during implementation.

A ten-point rollout checklist

1. Define scope. List the topics the bot may answer and those it must refuse, such as legal threats, medical claims or account closures.
2. Ground answers in your own documents. Retrieve from a controlled, versioned knowledge base instead of letting the model improvise policy.
3. Quote and link sources. Show which help-center page an answer came from, so customers and agents can verify it.
4. Escalate early. Offer a human at any time and hand over the full transcript so customers do not repeat themselves.
5. Disclose clearly. State at the start that the user is talking to an AI system.
6. Block commitments you cannot honor. Keep refunds, discounts and exceptions behind approved workflows rather than free-text promises.
7. Test with real tickets. Replay past conversations, including angry and ambiguous ones, before launch, and score accuracy by topic.
8. Protect data. Limit what personal data reaches the model, set retention periods and review your vendor’s data terms.
9. Monitor and sample. Review a weekly sample of conversations, track escalation and repeat-contact rates, and keep an incident process for wrong answers.
10. Keep a rollback. Be able to switch the bot off or restrict it to a safe subset within minutes.

For a look at one commercial product and how we assess such tools, see our Intercom Fin review.

What to watch next

• Further rulings on whether companies are liable for chatbot statements, in Canada and elsewhere.
• How EU supervisors and companies interpret the AI Act’s disclosure duty for chatbots.
• Independent measurements of resolution quality that go beyond vendor-reported figures.
• Whether regulators in finance, travel and telecom issue sector-specific guidance on bots.

Key takeaways

• Companies are responsible for what their support bots say; a Canadian tribunal rejected Air Canada’s claim that its chatbot was a separate entity.
• Vendor figures such as Klarna’s two-thirds-of-chats claim are promising but self-reported.
• The CFPB found chatbots handle basic inquiries but struggle with complex ones, and customers get frustrated when they cannot reach a human.
• The EU AI Act requires that people be told they are interacting with an AI system, unless it is obvious.
• A scoped, grounded, well-monitored bot with an easy path to a human avoids most known failures.

Frequently asked questions

Is a company legally responsible for what its chatbot says?

In the Air Canada case, a Canadian tribunal held that the airline was responsible for its chatbot’s wrong advice and ordered it to compensate the passenger. Other jurisdictions may differ, but the reasoning that a chatbot is part of the company’s website is persuasive.

Do I have to tell customers they are talking to an AI?

In the EU, Article 50 of the AI Act requires systems that interact with people to inform them that they are dealing with an AI system, unless it is obvious. Labeling the bot is good practice elsewhere too.

How much support work can a bot handle?

Klarna reported that its assistant handled two-thirds of chats in its first month, but that is a company claim. Results depend on your product, your documentation and how many requests need account actions or judgment.