A Hallon customer asks the operator’s chatbot for the customer service phone number. The bot answers: “We are a digital operator, which means we don’t offer customer service by phone.”
Hallon does have a phone number. Its customers have it in their contracts.
In April 2026, Sweden’s Patent and Market Court ruled that the answer was misleading and banned it, under a conditional fine of SEK 500,000. A month later, a German appeals court went further: a company answers for its chatbot’s false statements even if it gave the bot only correct data.
Courts now treat what your chatbot or AI agent tells customers as your own words. Under EU consumer law, which Norway applies through the EEA, a bot’s answer to a consumer is the company’s own commercial practice. If it misleads, it can be judged like any misleading ad. In Germany, correct training data was no defence. Don’t count on a disclaimer or a vendor contract to be one either.
Sweden: one answer, banned
The Consumer Ombudsman (KO) took Hallon to court over two things. Its customer service number was hard to find on the website, and its chatbot, Berrys, told customers there was no phone service at all.
The court split the case. Hiding the number wasn’t illegal, since existing customers have it in their contracts. But the chatbot’s answer was misleading. Hallon is now barred from giving that answer, “or a similar answer”, when a consumer asks for its phone number. Breaking the ban can cost SEK 500,000. Hallon has changed the answer.
The case isn’t final. KO has appealed, and the Patent and Market Court of Appeal has yet to decide whether to hear it. Even so, the ruling shows how a Swedish court reads a bot’s answer. It treated it as Hallon’s own marketing.
Germany: the chatbot is not a third party
In Germany, a cosmetic clinic’s website chatbot told visitors that the two doctors running it were specialists in plastic and aesthetic surgery. They held no such titles. A consumer organisation sued.
On 12 May 2026, the Higher Regional Court of Hamm ruled against the clinic (case 4 UKl 3/25). The clinic argued that the bot’s wrong answers weren’t its own act. The court rejected that. The chatbot is not a “third party”, so the false statements are the company’s own misleading practice. The court added that the clinic would be liable even if it had programmed the bot only with correct data.
Because the question is new, the court allowed an appeal to Germany’s Federal Court of Justice. Until that court rules, Hamm is the clearest judgment on chatbot answers we know of from an EU country.
A Canadian tribunal took the same view in 2024. It held Air Canada to a refund its chatbot had promised and rejected the claim that the bot was “responsible for its own actions”.
Norway: same rule, no case yet
No Norwegian court has ruled on a chatbot yet. If one does, it will start from the same rule. Sweden, Germany and Norway all build on the EU’s Unfair Commercial Practices Directive, which Norway applies through the EEA. In Norway it is section 7 of the Marketing Control Act (markedsføringsloven), enforced by Forbrukertilsynet.
Forbrukertilsynet’s guidance on misleading practices, updated in April 2026, names how a company handles customer enquiries, such as complaints and cancellations, as part of its commercial practice. An AI agent that answers those enquiries is part of that practice.
Why AI agents raise the stakes
A human agent who gets the cancellation rules wrong does it on a few calls, and a team leader may catch it in a review. An AI agent with the same wrong belief gives it to every customer who asks, in the same words, until someone notices.
Hamm adds a second point. “We gave it the right documents” is no defence. What counts is what the customer was told.
Both cases were about website chatbots, but nothing in either ruling depends on the answer being typed. The same reasoning is likely to apply to any AI that speaks for you. That includes the bot in your app, the voice agent that answers your line after hours, and the agent your outsourcing partner runs for you.
The questions to test first
Both cases turned on ordinary questions. The risky ones are those where a wrong answer changes what the customer does next.
| The customer asks | A wrong answer looks like |
|---|---|
| “What’s your phone number?” or “Can I talk to a person?” | “We don’t offer that.” Hallon’s case |
| “How do I cancel?” | A route that doesn’t exist, or a notice period that’s wrong |
| “Can I get a refund?” or “Can I return this?” | A right denied, or a refund promised that policy doesn’t allow. Air Canada’s case |
| “What does it cost?” | A price, fee or discount that isn’t yours |
| “Who will I be dealing with?” | A title, licence or qualification nobody holds. The Hamm case |
| “Am I talking to a person?” | Anything but a clear “No, I’m an AI”. Since August 2026 the AI Act requires the disclosure |
A checklist for your AI agents
- Ask your own bot. Put each question in the table to your chatbot and voice agent, in every language you serve, and write down the answers. Ask follow-ups too. Hallon’s bot answered a simple question with a confident no.
- Write the right answer down. For each question, agree the correct answer with legal and customer service. That’s what you test against.
- Test after every change. A new prompt, model or knowledge base can change answers that were right last month. Run the same questions again.
- Score real conversations, not just tests. Customers ask in ways you didn’t script. Review a sample of real AI conversations against the same answers, and look at every conversation where a customer asked about cancelling, refunds or reaching a person.
- Hold AI and people to one bar. Score your AI agent on the same criteria as your human agents, and the vendor or partner who runs it too. We’ve written about one quality bar for human and AI agents.
- Keep the record. When a regulator or a customer asks what your agent said, you need the conversation, not a guess.
Where Conversational Intelligence fits
Conversational Intelligence scores every call against your scorecard, whether a person or one of our AI voice agents handled it. Add criteria such as “gave the correct cancellation route” or “offered a person when asked”. Every score links to the sentence, timestamp and speaker behind it, so you find a wrong answer before a customer or a regulator does.
Your legal and customer service teams decide what the right answer is. We help you check that customers get it on every call.
