Go back

Article

Jun 10, 2026

Your Chatbot's Word Is Legally Your Word: Air Canada and the Containment Stack

A tribunal said the quiet part out loud: your chatbot's answer is your answer. Here's the engineering stack that keeps that promise survivable

Single thin white line bisecting black void broken at center glowing orange

TL;DR

A Canadian tribunal ruled in February 2024 that Air Canada owns whatever its chatbot tells a customer.
Hallucination drove 22% of enterprise AI agent rollbacks in the Sinch survey of 2,500+ respondents.
The legal exposure is not the model. It is the absence of a containment stack around the model.
Three layers contain it: grounded retrieval on an approved corpus, refusal tuning, and a human gate on high-stakes answers.
One source of truth between policy pages and bot answers is the unglamorous control that prevents most lawsuits.

The short answer operators came here for

If you run customer service with an LLM in the loop, ai chatbot hallucination liability is no longer a hypothetical. In Moffatt v. Air Canada (February 2024), the BC Civil Resolution Tribunal held the airline responsible for incorrect bereavement-fare information its chatbot gave a grieving customer, ruling the company owns all information on its website "whether it comes from a static page or a chatbot". The award was modest, roughly $650 CAD plus interest and fees. The precedent is not. The fix is also not a better model. It is a containment stack: grounded retrieval, tuned refusal, and a human gate on the answers that are expensive to get wrong.

2. Why "the chatbot said it, not us" failed as a defense

Air Canada's argument relied on a fiction every operator has been quietly tempted by: that an AI surface is somehow downstream of the company that deployed it. The tribunal's reasoning closed that loop in plain language. The chatbot was on the airline's website. The airline chose to deploy it. The airline benefited from the deflection of human agent time. Therefore, the airline owns the output.

For anyone designing chatbot wrong information legal liability controls, three operational implications follow.

3. The hallucination math: when a wrong answer costs more than a human one

Here is the part most legal coverage skips. Hallucination is not a rare event. In a Sinch survey of more than 2,500 enterprise respondents, hallucination was named as the cause of 22% of enterprise AI agent rollbacks. Roughly one in five companies that pulled back an agent did so because it confidently said things that were not true.

Which means the operator question is not "can we eliminate hallucination?" The honest answer is no, not at acceptable cost, not in 2026. The question is "which answers can we afford to let the model attempt unsupervised?"

4. Containment layer 1: ground every answer in an approved corpus

The first control is the one most teams skip because it is operationally annoying. The model does not get to answer from its pretraining. It answers only from a retrieval layer that pulls from a corpus you have approved.

In practice this means three things. A defined corpus, policy pages, product documentation, pricing sheets, support articles, versioned and owned by a named human. A retrieval step that runs before generation and returns the source chunks the answer must be built from. And a system prompt that instructs the model to say "I don't have that information" when retrieval comes back empty or low-confidence, rather than improvising.

This is the engineering pattern behind our ai automation work, and it is the single highest-impact control against ai hallucination customer service failures. It does not eliminate hallucination. It bounds it. The model can still phrase the retrieved chunk badly. It cannot invent a bereavement-fare policy that does not exist in the corpus.

5. Containment layer 2: refusal and escalation tuning

The second layer is teaching the bot when to say nothing.

Ungrounded LLMs are trained to be helpful, which is another way of saying they are trained to produce an answer even when they shouldn't. A containment-tuned bot inverts that bias for specific categories. When retrieval confidence is below a threshold, when the question matches a pattern flagged as high-stakes, or when the user explicitly asks for a commitment ("will you refund me if…"), the bot is tuned to refuse and route, not to attempt.

We have written more on the broader pattern in human-in-the-loop automation patterns.

6. Containment layer 3: high-stakes answers get a human gate

The third layer is the one Air Canada did not have: a hard rule that certain categories never get an autonomous answer.

High-stakes categories are defined by expected cost, not by topic. In a typical services business, the list runs about five entries: refund and credit decisions, pricing commitments outside the standard sheet, eligibility for regulated benefits (bereavement, medical, legal status), contractual language interpretation, and anything where the customer is in distress.

This is not a sophistication ceiling. It is a deliberate decision-rights design, the same logic we lay out in our ai agent governance framework. The bot has decision rights on the low-cost questions. The human retains decision rights on the expensive ones. The contract is written down before the bot ships, not after a tribunal writes it for you.

7. One source of truth: keeping policy pages and bot answers in sync

The last control is the most boring and the most load-bearing.

Moffatt's case had a quiet detail: the correct bereavement policy existed on the Air Canada website. It just didn't match what the bot said. Two surfaces, two answers, one company, one liability.

The fix is structural. Policy lives in one place. The website renders from it. The bot retrieves from it. When the policy changes, both update from the same edit. There is no "bot knowledge base" maintained by the AI team and a "policy page" maintained by legal. There is one document, one owner, one version history.

FAQ

Does the Air Canada ruling apply outside Canada?

Not directly. The Moffatt decision is a BC Civil Resolution Tribunal ruling and is binding only in that jurisdiction. The reasoning, though, tracks established negligent misrepresentation doctrine in most common-law systems. US, UK, and Australian courts considering similar facts would likely reach a comparable conclusion: a company owns the statements its deployed software makes to customers.

Can a terms-of-service disclaimer protect us from chatbot liability?

Probably less than your legal team hopes. Air Canada had standard terms in place. The tribunal still ruled against them. Disclaimers can shape damages and narrow scope, but they rarely override a clear negligent-misrepresentation claim when a customer reasonably relied on a bot's answer. Treat disclaimers as one layer, not the layer.

What is retrieval-augmented generation in plain terms?

It's a pattern where the model retrieves relevant documents from your approved corpus before answering, then writes the answer using only those documents as source material. Instead of asking the model what it knows, you ask it in short what your policy pages say. It bounds hallucination by removing the model's license to improvise.

How do I decide which questions need a human gate?

List the question categories your support team handles. For each, estimate the cost of a wrong answer, refund exposure, regulatory risk, reputational damage. Any category where the cost of one wrong answer exceeds a month of automation savings goes behind a human gate. In our client work, that is typically refunds, pricing exceptions, and any eligibility question tied to a regulated benefit.

How often should we audit bot answers against policy?

In practice, weekly sampling of 30 to 50 transcripts is the floor for any bot handling more than a few hundred conversations a day. Every answer should be traceable to a specific source document. Anything that cannot be traced is a hallucination by definition, and the audit is how you catch the pattern before a customer does.

How should a small team prioritize ai-chatbot?

Start with the workflow that already has a baseline: hours, leads, errors, or budget waste.

What should be measured before investing in ai-chatbot?

Measure cycle time, volume, handoffs, error rate, and the current owner.

When should ai chatbot hallucination liability stay manual instead of automated?

Keep it manual when judgment, approval, brand nuance, or customer trust is on the line.

How does can ai agents make outbound calls change the budget for ai-chatbot?

can ai agents make outbound calls usually adds integration, QA, and monitoring work.

What is the first project to launch from this ai chatbot hallucination liability playbook?

Launch the narrowest workflow with a visible result.

Ship it this week

Pull 50 recent chatbot transcripts on Monday. For each substantive answer, find the source document. The ones you cannot trace are your containment backlog. Start there.

If you want a second pair of eyes on the stack, send us the transcripts.