The Air Canada Chatbot Ruling: Who Is Liable When Your AI Gets It Wrong?

In 2024, the British Columbia Civil Resolution Tribunal ordered Air Canada to pay C$812.02 because its website chatbot gave a grieving customer a refund policy that did not exist. The airline argued the chatbot was a "separate legal entity" responsible for its own words. The tribunal flatly rejected that defense. Here's the part every business owner needs to sit with: the company was held liable for what its software said, and that precedent now shapes how courts everywhere think about AI mistakes.

This article breaks down what actually happened in Moffatt v. Air Canada, why "the chatbot did it" is not a legal defense, and the concrete steps you can take to deploy an AI agent without inheriting a lawsuit.

What happened in the Air Canada chatbot case?

Air Canada was held legally liable for negligent misrepresentation after its website chatbot told a customer he could apply for a bereavement discount after booking, when the airline's real policy required the request before travel. The customer, Jake Moffatt, booked a flight to attend his grandmother's funeral, followed the chatbot's advice, and was later denied the refund. He took the airline to the BC Civil Resolution Tribunal and won (Moffatt v. Air Canada, BC Civil Resolution Tribunal, 2024).

The chatbot had essentially invented a policy. It blended a real bereavement-fare program with a fabricated process, and presented the whole thing with the confidence of an official source. Moffatt relied on it, made a decision, and lost money. That chain of events — false statement, reasonable reliance, financial harm — is the textbook shape of negligent misrepresentation.

The dollar figure was tiny: C$812.02 plus tribunal fees. But the ruling traveled far beyond its size, because it answered a question businesses had been quietly avoiding. When a machine on your website says something wrong and a customer gets hurt, who pays? The tribunal's answer was unambiguous. You do.

Is my business liable if the chatbot gives wrong information?

Yes. If your AI agent makes a false statement that a customer reasonably relies on to their detriment, your business is on the hook, exactly as if a human employee had said it. The Air Canada tribunal made this explicit: a company is responsible for all the information on its website, "whether the information comes from a static page or a chatbot."

This matters because a lot of vendors and buyers have quietly assumed the opposite. The instinct is that AI is a third-party tool, almost like weather data or a stock ticker, and that if it glitches, the fault lies somewhere upstream. The Air Canada ruling kills that thinking. Your AI agent is not a contractor with its own legal standing. It is a representative of your business, and its words are your words.

The practical translation is simple. Treat every sentence your AI agent can produce as if your most junior employee said it to a customer in writing, on the record, with no manager checking first. If that sentence would create a problem coming from a person, it creates the same problem coming from your software. The technology changed; the liability did not.

This is not a uniquely Canadian quirk, either. The negligent-misrepresentation logic the tribunal applied — false statement, reasonable reliance, resulting harm — exists in some form across most common-law jurisdictions and has close cousins in consumer-protection regimes worldwide. Air Canada is simply the first widely cited case to apply it to a chatbot, which is why it gets quoted in boardrooms far from Vancouver. If you operate anywhere your customers can take you to a small-claims court or file a consumer complaint, assume the same principle reaches you. The cost of getting this wrong is rarely the C$812 award; it is the time, the public story, and the erosion of the trust you were automating to build.

Why can't I just disown the chatbot legally?

You cannot disown the chatbot because the law treats it as part of your business, not as an independent agent acting on its own. Air Canada's "separate legal entity" argument failed precisely because a chatbot has no independent existence — it does not own assets, sign contracts, or bear responsibility. It is software you deployed, configured, and pointed at your customers.

There is a deeper reason this defense will keep failing. Liability follows control and benefit. You control what the AI agent says by choosing the platform, feeding it your content, and setting its behavior. You benefit when it deflects support tickets and captures leads at 2 a.m. The legal system is consistent here: if you reap the upside of an automation, you carry the downside when it errs. A disclaimer buried in your terms of service does not transfer that risk, especially when a customer never saw it before relying on a confident, specific answer.

This is also why "the AI hallucinated" is not a mitigating excuse — it is the exact risk you are responsible for managing. Which leads to the uncomfortable technical reality underneath all of this.

Why does AI sound so confident when it's wrong?

AI agents are most dangerous not because they are dumb, but because they can be confidently wrong — fluent, authoritative, and completely fabricated. Peer-reviewed research from the Technion, Oxford, and Hebrew University found that large language models "can hallucinate with high certainty even when they have the correct knowledge" (Simhi et al., 2025). In plain terms, the model often sounds most sure exactly when it is wrong.

The numbers on how often this happens are sobering once you leave demo conditions. On Vectara's grounded-summarization benchmark, the best models hold hallucination rates between 0.7% and roughly 1.5%, but on harder, real-world content even flagship reasoning models exceeded 10% (Vectara HHEM Leaderboard, 2025–2026). A 10% fabrication rate on consequential answers is not a rounding error — it is a Jake Moffatt waiting to happen. McKinsey reports that inaccuracy is the most commonly cited AI risk, with nearly a third of organizations reporting negative consequences from it (McKinsey, 2025).

Here is why this should change how you deploy. A chatbot that says "I'm not certain, let me connect you to someone" is safe. A chatbot that invents a bereavement policy in a smooth, helpful tone is the one that ends up in front of a tribunal. The failure mode that creates legal liability is not the obvious error — it is the believable one.

It helps to understand why this happens mechanically. A large language model is, at its core, a system that predicts plausible-sounding text. When it lacks the right fact, it does not return an error message the way a database would — it generates the most statistically likely continuation, which is often a coherent, well-formatted answer that happens to be false. Confidence and correctness are not linked inside the model the way they are inside a knowledgeable human. That is the precise trap Air Canada fell into: the chatbot was not malfunctioning by its own logic; it was doing exactly what an ungrounded language model does, which is fill the gap with something that reads like a policy.

That distinction is the whole reason guardrails work. You cannot make the underlying model honest about its own uncertainty by asking nicely. You change the deployment so the model is not the source of truth in the first place — it retrieves from documents you control and is instructed to stop when those documents run out. The model stops being a confident author and becomes a careful reader. That architectural choice, not a smarter model, is what separates a safe deployment from a lawsuit.

How can an SME limit its liability when deploying an AI agent?

You limit liability by engineering the AI agent so it can only answer from your verified content, abstains when unsure, escalates cleanly to a human, and tells customers it is AI. The Air Canada disaster was not inevitable — it was a chatbot allowed to make things up with no guardrails. Every safeguard below directly attacks the chain of events that created legal exposure in that case.

The good news for an SME is that none of this requires a legal department or a custom AI team. It requires choosing a platform that bakes these controls in, and configuring them deliberately before you go live. Trust is engineered, not assumed.

Here is the operator-level checklist:

1. Ground every answer in your own content (RAG). Force the AI agent to answer from your verified documents, not its training memory. This is the single biggest defense against a fabricated policy, because the model is no longer free to invent.
2. Curate your knowledge base. Stale and conflicting articles are a top cause of grounded-but-wrong answers; cleaning them up cuts these errors by roughly 20–30% (IrisAgent/Zendesk). Garbage in, confident garbage out.
3. Make "I don't know" the default. Instruct the agent to abstain and escalate when context is missing, rather than guess. A non-answer is never a lawsuit; a wrong answer can be.
4. Set confidence thresholds. A common production pattern: above ~85% confidence the AI proceeds; 70–85% it proceeds but flags for review; below ~70% it escalates to a human. High confidence should never authorize an irreversible action unsupervised.
5. Design the human handoff. Escalate immediately on an explicit request ("talk to a person"), after the second or third failed attempt, on detected frustration, and on high-risk intents like refunds, billing, and policy. Use warm transfers that carry the full transcript so customers never repeat themselves.
6. Cite sources. Showing the customer the document the AI used builds trust — one study found source citations lifted CSAT 8–12% even with no change to underlying accuracy.
7. Be transparent that it's AI. 95% of consumers expect a clear explanation when AI makes decisions affecting them (Zendesk CX Trends 2026). Disclosure is a trust-builder, not an admission of weakness.
8. Track accuracy as a first-class metric. Measure hallucination and accuracy alongside CSAT and resolution, so a drift toward confident-but-wrong answers shows up before a customer finds it.

This is the philosophy behind tools like Omago, an AI agent platform that helps SMEs automate customer conversations across WhatsApp, Telegram, and web chat. A messaging or web agent grounded in your own knowledge base — with honest "I don't know" behavior, clear escalation, and source citations — is far safer than an ungrounded bot tuned to always have an answer.

What's the difference between a safe AI deployment and a risky one?

A safe deployment constrains what the AI can say and routes uncertainty to a human; a risky one lets the AI answer anything in a confident voice with no escape hatch. The Air Canada chatbot sat firmly in the second category. The distinction is not about how advanced the model is — it is about how the deployment is designed.

The table below maps the two approaches against the factors that actually determine legal exposure.

Notice that none of the "safer" column requires the customer to trust the AI blindly. It assumes the AI will sometimes be wrong and builds the business's protection around that assumption. That is the entire shift the Air Canada ruling should produce in how you think.

The reassuring part: this is an engineering and design problem with well-understood mitigations, not an unsolvable legal trap. You are not choosing between deploying AI and avoiding liability. You are choosing between deploying it carelessly and deploying it with guardrails.

Do customers even trust AI customer service after rulings like this?

Customers are skeptical of AI accuracy, but they will accept AI for fast, simple transactions — and transparency plus reliability win them over. The skepticism is real and worth respecting: 84% of consumers believe human agents are more accurate than AI, just 8% prefer AI over humans, and 61% feel humans better understand their needs (SurveyMonkey, 2025). After a story like Air Canada's makes the rounds, those numbers are not surprising.

But skepticism is not refusal. The same body of research shows trust is winnable when the experience is honest and competent. Customers readily accept an AI agent for quick, low-stakes tasks — checking an order, booking a slot, answering a documented question — where speed beats nuance. The friction shows up when an AI overreaches into high-stakes, emotional, or judgment-heavy territory, which is exactly where your escalation rules should be sending the conversation to a human anyway.

So the liability lesson and the trust lesson converge on the same design. Disclose that it's AI. Keep it inside well-defined, well-grounded intents. Hand off cleanly when stakes rise. Do those three things and you simultaneously lower your legal exposure and earn more customer trust than the business running an ungoverned bot. The companies that get burned through 2028 will not be the ones that automated — they will be the ones that automated without guardrails.

If you want to go deeper on the underlying reliability question, our guide on whether you can trust AI customer service and the guardrails that make it safe covers the technical side, and for the broader regulatory picture, see our overview of AI governance for small business.

Frequently Asked Questions

Was Air Canada actually held liable for what its chatbot said?

Yes. In Moffatt v. Air Canada (2024), the BC Civil Resolution Tribunal found the airline liable for negligent misrepresentation after its chatbot gave incorrect bereavement-fare guidance, and ordered it to pay C$812.02. The tribunal rejected the argument that the chatbot was a separate legal entity responsible for its own statements.

Can I avoid liability with a disclaimer in my terms of service?

A disclaimer offers limited protection at best. The Air Canada ruling suggests courts will treat your AI agent's statements as your business's statements, and a customer who relies on a confident, specific answer they never saw contradicted is unlikely to be bound by buried terms. The reliable protection is preventing wrong answers in the first place through grounding and escalation, not disclaiming them after the fact.

How often do AI chatbots actually give wrong answers?

It depends heavily on how they're built. The best grounded models hold hallucination rates around 0.7–1.5% on benchmark tests, but on harder real-world content even leading reasoning models exceeded 10% (Vectara HHEM Leaderboard, 2025–2026). An ungrounded bot answering from memory is far riskier than one constrained to your verified knowledge base.

What's the single most important safeguard against a chatbot lawsuit?

Grounding answers in your own verified content using retrieval-augmented generation (RAG), combined with making the AI escalate to a human when it's unsure. This directly prevents the failure that hurt Air Canada — a chatbot inventing a policy from memory — by forcing the agent to answer only from documents you control, or hand off when it can't.

Does telling customers they're talking to AI increase my liability?

No — it does the opposite. Transparency is a trust-builder, and 95% of consumers expect a clear explanation when AI makes decisions affecting them (Zendesk CX Trends 2026). Disclosing that a customer is talking to an AI agent sets honest expectations and is increasingly an emerging legal expectation, not a confession of weakness.

Sources: Moffatt v. Air Canada, BC Civil Resolution Tribunal (2024); Simhi et al., Technion/Oxford/Hebrew University (2025); Vectara HHEM Leaderboard (2025–2026); McKinsey (2025); SurveyMonkey (2025); Zendesk CX Trends (2026); IrisAgent/Zendesk.