Who Pays When an AI Agent Gets It Wrong?
A business answers for what its AI agent tells callers. What Moffatt v. Air Canada decided, the EU's new rule on software, and the controls that cut the risk.
Kurze Antwort
Generally, treat the business as responsible for what its AI agent tells a caller. In Moffatt v. Air Canada (2024 BCCRT 149, 14 February 2024), a tribunal rejected the idea that a chatbot is a separate legal entity, found the airline did not take reasonable care to ensure its chatbot was accurate, and ordered it to pay $812.02. This is one tribunal decision, not legal advice.
When an AI agent tells a customer something false, who has to make it right? The vendor that built the model, the business that deployed it, or nobody? One answer comes from a short decision in British Columbia about an airline chatbot. This guide quotes it from the decision itself, adds the limits of the EU's new product liability directive, and then carries both over to the phone channel.
It is general information, not legal advice. DialerBee does not provide legal advice or guarantee regulatory compliance.
What did Moffatt v. Air Canada decide?
The decision is Moffatt v. Air Canada, 2024 BCCRT 149, issued by the Civil Resolution Tribunal of British Columbia on 14 February 2024 (Sources, item 1). The decision is recorded as a small claims dispute, so this is a tribunal decision and not a ruling of a higher court. Its reasoning is persuasive, but it does not bind courts elsewhere.
The facts, as the decision states them:
- In November 2022, after a death in the family, the customer booked a flight with Air Canada and, while researching flights, used a chatbot on the airline's website.
- The chatbot suggested that bereavement fares could be applied for retroactively. The customer later learned from airline employees that Air Canada did not permit retroactive applications.
- When the customer complained, an airline representative admitted the chatbot had provided "misleading words" and said the issue was noted so the chatbot could be updated.
- Air Canada argued that it could not be held liable for information provided by its chatbot.
The tribunal's reasoning has two parts that matter to anyone who deploys an agent:
- Paragraph 27. The tribunal said that, in effect, Air Canada suggested the chatbot was a separate legal entity responsible for its own actions, and called this "a remarkable submission". It added that a chatbot is still just a part of the airline's website, that the airline is responsible for all the information on its website, and that it makes no difference whether the information comes from a static page or a chatbot.
- Paragraph 28. The tribunal found Air Canada did not take reasonable care to ensure its chatbot was accurate. It also rejected the idea that the customer should have double-checked the chatbot against another part of the website, because the airline did not explain why the static page was more trustworthy than its chatbot.
The tribunal applied negligent misrepresentation: a seller that does not exercise reasonable care to make sure its representations are accurate and not misleading. It found the customer reasonably relied on the chatbot (paragraph 29). On the airline's reliance on its tariff, the tribunal found Air Canada had not proven a contractual defence because it did not provide the relevant portions of the contract (paragraph 31).
The order
Paragraph 44 orders Air Canada to pay a total of $812.02, made up of $650.88 in damages, $36.14 in pre-judgment interest and $125 in tribunal fees. The damages are the difference between what the customer paid and what the tribunal found the bereavement fares would have cost. The amount is small. The reasoning is the part that travels.
Why a phone call raises the stakes
The decision is about a website chatbot. Nothing in it limits the idea to text, and a voice agent raises two practical differences. These are our observations, not holdings of the tribunal.
- There is no page to compare against. The tribunal rejected the argument that a customer should cross-check the chatbot against the static page. On a call the caller is even less able to cross-check, because they are listening in real time and often making a decision while on the line.
- The recording is evidence. A recorded call shows exactly what the agent said. That helps a business that did nothing wrong, and it limits the room for dispute when something was said that should not have been. Whether you may record, and what you must tell the caller, depends on local law. See AI call disclosure rules for the disclosure side.
For the wider question of when callers should be told they are speaking to an AI, and when a human should take over, see our AI or human calls buyer's guide.
What changes in the EU?
It is tempting to read Directive (EU) 2024/2853 of 23 October 2024 on liability for defective products as making AI vendors liable for anything their software says. The text is more specific, and the details matter for a business deploying an agent (Sources, item 2).
| Bestimmung | Was sie besagt | Why it matters here |
|---|---|---|
| Article 4(1) | A product includes software | Software is inside the regime |
| Recital 13 | Software includes AI systems, whether stored on a device, accessed through a network or cloud, or supplied as a service; information, such as the content of digital files, is not a product | An AI agent is covered as software, but content as such is not |
| Article 2(1) | Applies to products placed on the market or put into service after 9 December 2026 | It looks forward from that date |
| Article 22 | Member States bring the necessary laws into force by 9 December 2026 | Check how each Member State you sell into has transposed it |
| Artikel 6 | Compensation covers only death or personal injury (including medically recognised damage to psychological health), damage to property (with exceptions, including property used exclusively for professional purposes) and destruction or corruption of data not used for professional purposes | A wrong price or a refund promise is not on that list |
| Article 2(4)(b) | Does not affect rights under national rules on contractual liability or on non-contractual liability on other grounds | Contract and misrepresentation claims remain where they were |
The summary for a business: the directive treats software, including AI systems, as a product for the purposes of no-fault liability for defective products, but the harm it compensates is injury, property damage and loss of private data. A customer who relied on a wrong price, a wrong eligibility statement or a refund promise is more likely to rely on contract, consumer protection or misrepresentation rules, which Article 2(4)(b) leaves in place. Whether any specific incident is covered is for EU counsel to say.
The directive is also not a reason to relax. Moffatt shows that, outside product liability, a wrong statement can already create a claim under ordinary rules.
Which answers carry the most risk?
Not every wrong answer costs the same. The ones to protect first are statements a customer will hold you to.
| Type of answer | Why it is risky | Kontrolle |
|---|---|---|
| Prices and fees | A number a caller will repeat back to you | Catalogue-only prices through a price-truth guard |
| Refunds and credits | A promise of money; the Moffatt fact pattern | The agent never promises; hand over to a person |
| Teilnahmeberechtigung | "You qualify" is a representation | Answer from one source document; hand over edge cases |
| Policy exceptions | The agent invents flexibility the policy lacks | Exceptions are a human decision |
| Appointment promises | A booked slot or a guaranteed time that does not exist | Book only after a free-busy check on the calendar |
| Dates, deadlines and notice periods | The Moffatt problem: a time window the customer relied on | Source document reviewed and versioned |
Five controls that reduce the risk of a wrong promise
- Catalogue-only prices. If the agent can speak only the prices in your catalogue, it cannot invent a price. DialerBee's price-truth guard does this on the KI-Sprachagent.
- One knowledge base as the source of truth. Answers come from documents you upload and own, retrieved by semantic search. When a policy changes, you change the document, and there is no second copy of the policy hiding in a script. See the AI voice agent knowledge base. The tribunal's point about a static page and a chatbot disagreeing is the failure this prevents.
- A person for exceptions. Refunds, exceptions and disputes go to a human by design. Our human handoff guide explains when.
- A versioned agent with rollback. In Agent Studio every publish is a version, you promote one deliberately, and you can roll back if a change made things worse.
- A recording, transcript and outcome for every call. Every AI call produces a result with these. Add Anrufaufzeichnung retention that fits your legal advice, and review calls with QA on AI calls.
Guardrails sit around the conversation, and test calls let you hear the agent before customers do. None of these is a guarantee. They reduce the chance of a wrong promise and give you the evidence to deal with one. For the wrong-answer failure mode itself, see what is AI hallucination.
What to do the day an AI agent makes a wrong promise
- Pull the recording and the transcript. Establish exactly what was said, to whom and when.
- Stop the spread. Use the kill switch at the narrowest scope that works: the agent, the campaign, or all AI calling if you cannot tell where the wrong answer came from.
- Find the source. Was it a document that was wrong, a document that was missing, or a model that went beyond its documents? Fix the document or the guardrail, then publish a new version.
- Decide on the customer. With legal and finance owners, decide whether to honour the promise. The tribunal in Moffatt was not persuaded by the argument that the agent is a separate entity.
- Look for other affected calls. Search recordings and transcripts for the same statement from the same period.
- Log it and re-test. Record the incident, the fix and the version. Place test calls on the phone before the agent goes back live.
Praxisbeispiel (zur Veranschaulichung)
A clinic's agent tells a caller that a consultation is free. The price list says it costs a fee. The caller books on that basis. The clinic pulls the recording, pauses the campaign, finds the knowledge base has an old promotion page, replaces it, and publishes a corrected version. It then decides, with its owner, whether to honour the free consultation for the callers who heard it. The numbers and the clinic are invented to show the sequence.
Häufig gestellte Fragen
Is a company liable for what its AI chatbot or voice agent says?
Treat it as yes. In Moffatt v. Air Canada (2024 BCCRT 149) the tribunal called the argument that a chatbot is a separate legal entity a remarkable submission, held that Air Canada did not take reasonable care to ensure its chatbot was accurate, and ordered it to pay. It is one tribunal decision and not a binding rule everywhere, so this is not legal advice.
What did Moffatt v. Air Canada decide?
A customer relied on an airline chatbot that suggested bereavement fares could be applied for retroactively, which the airline did not permit. The Civil Resolution Tribunal of British Columbia found negligent misrepresentation and ordered Air Canada to pay a total of $812.02: $650.88 in damages, $36.14 in pre-judgment interest and $125 in tribunal fees. It was issued on 14 February 2024.
Does the new EU product liability directive make AI vendors liable for wrong answers?
Not for a wrong price or a wrong refund promise, on our reading. Directive (EU) 2024/2853 treats software as a product, but its right to compensation covers only death or personal injury, damage to property and destruction or corruption of non-professional data (Article 6). Article 2(4)(b) leaves national contractual and other liability rules untouched, which is where a wrong promise to a customer sits.
When does the EU directive apply?
Article 2(1) says it applies to products placed on the market or put into service after 9 December 2026, and Article 22 requires Member States to bring the necessary laws into force by 9 December 2026. Which of your products and which Member States are affected is a question for EU counsel.
Does it matter that the wrong answer was spoken on a phone call?
It raises the stakes in two ways. A caller has no screen to compare against the written policy, and a recording is evidence of exactly what was said. That makes the content of the agent's answers, and your records of them, more important than on a web page.
What should we do the day an AI agent makes a wrong promise?
Pull the recording, stop the affected agent version or campaign, correct the source document, decide with your legal and finance owners whether to honour the promise, and log the incident and the fix. Then re-test the agent on the phone before it goes live again.
Haftungsausschluss
This article is general information, not legal advice. DialerBee does not provide legal advice or guarantee regulatory compliance. It summarises one tribunal decision and one EU directive as published, checked on 2026-11-19, and the law may differ in your country or change. Confirm your position with qualified counsel. The AI voice agent has run on live calls in Arabic and English, so for any other language, test it on a demo first.
Quellen
- Moffatt v. Air Canada, 2024 BCCRT 149, Civil Resolution Tribunal of British Columbia, issued 14 February 2024, paragraphs 2, 4, 22, 24 to 29, 31 and 44. decisions.civilresolutionbc.ca, 2024 BCCRT 149 (Civil Resolution Tribunal, checked 2026-11-19)
- Directive (EU) 2024/2853 of the European Parliament and of the Council of 23 October 2024 on liability for defective products, Official Journal L series, 18 November 2024, Articles 2, 4, 6 and 22 and Recital 13. eur-lex.europa.eu, Directive (EU) 2024/2853 (EUR-Lex, checked 2026-11-19)
Weiterführende Lektüre
- KI-Sprachagent
- AI Agent Studio
- What Is AI Hallucination?
- AI Voice Agent Knowledge Base
- AI Call Disclosure Rules
- How to QA AI Voice Agent Calls
- Anrufaufzeichnung
- AI or Human Calls: A Buyer's Guide
Weiterführende Lektüre: AI Voice Agent Security: Can Callers Trick It? · Legen Anrufer auf, wenn eine KI sagt, dass sie eine KI ist?
Verwandte Artikel
France's Caller ID Rules for Automated Calls
9 Minuten Lesezeit
EinhaltungIst die Aufzeichnung von Telefonaten in Deutschland erlaubt?
9 Minuten Lesezeit
TechnologieAI Voice Agent Security: Can Callers Trick It?
11 Minuten Lesezeit
TechnologieArabic Sentiment Analysis on Calls: What Works
9 Minuten Lesezeit
Bereit, DialerBee in Aktion zu sehen?
Fordern Sie eine 15-minütige Live-Demo an oder starten Sie eine 14-tägige kostenlose Testphase – ohne Folien, ohne Verpflichtung. Das Telefonieren beginnt, sobald Ihr eigener SIP-Trunk oder Ihre TK-Anlage verbunden und freigegeben ist.
14-tägige kostenlose Testphase · kostenloses KI-Guthaben · keine Kreditkarte · 11 Sprachen · BYOC · Kontrollen zur Unterstützung der Compliance