# Who's Liable When Your AI Support Agent Gives a Wrong Answer?

> Two court rulings — Air Canada in 2024 and a German appeals court in May 2026 — have settled the question: a business owns what its support chatbot says, even when the AI makes it up. Here's what that means operationally, and the controls that keep a wrong answer from becoming a binding promise.

*By Mithun · Published June 29, 2026 · 11 min read*

Category: AI customer support

Tags: AI safety, Citations, Human handoff, Hallucinations, Grounded answers, Support automation

{/* Image note: Use the generated editorial hero above the title — a chat answer turning into a signed/stamped document, with a citation chip on the grounded path. It intentionally avoids competitor logos and uses no headline text. This is a trust/accountability evergreen guide, not a comparison post, so no competitor screenshots are required. */}

If your AI support agent quotes a refund policy that doesn't exist, your business has to honor it — and "the AI said it, not us" has now failed in court twice, on two continents. That is the practical reality every team deploying a support chatbot in 2026 is working under, whether or not they've read the rulings.

This is a guide about accountability, not hallucinations specifically. We've written separately about [why support agents make things up and how to prevent it](/blog/ai-customer-support-hallucinations/). This post answers a narrower, more uncomfortable question: when an automated answer is wrong, who owns the consequence — and what do you actually have to put in place so a wrong answer stays a small problem instead of a binding one? (This is general information for operators, not legal advice; talk to a lawyer about your situation.)

## Two rulings that say the same thing: you own the answer

The legal direction of travel is no longer ambiguous. Two cases, two countries, one principle.

**Air Canada (Canada, 2024).** A traveler asked Air Canada's website chatbot about bereavement fares after a death in the family. The bot told him he could claim the discounted rate retroactively. That was wrong — the real policy required applying before travel. When he asked for the difference back, the airline refused and argued to the British Columbia Civil Resolution Tribunal that the chatbot was, in effect, "a separate legal entity that is responsible for its own actions." The tribunal rejected that outright, found the airline had committed a negligent misrepresentation, and ordered it to pay. The decision (*Moffatt v. Air Canada*, 2024 BCCRT 149) landed in February 2024. ([CBC News](https://www.cbc.ca/news/canada/british-columbia/air-canada-chatbot-lawsuit-1.7116416), [McCarthy Tétrault analysis](https://www.mccarthy.ca/en/insights/blogs/techlex/moffatt-v-air-canada-misrepresentation-ai-chatbot))

**OLG Hamm (Germany, 2026).** More than two years later, a German appellate court reached the same conclusion in a different context. A consumer association sued a cosmetic-treatment provider because its AI chatbot described the operators using specialist medical titles that don't exist. In its judgment of 12 May 2026 (Az. 4 UKl 3/25), the Oberlandesgericht Hamm held that the chatbot is not an independent third party but a tool of the operator, and that its statements are attributed directly to the company — *even when the AI hallucinates or misprocesses correct input.* The court's reasoning: the operator defines the bot's scope and has enough control over the system to prevent the false statements, as shown by the fact that the company was able to reprogram it afterward. ([OLG Hamm press release](https://www.olg-hamm.nrw.de/behoerde/presse/Pressemitteilungen/16_26_PE_KI-Chatbot/index.php), [English analysis](https://policy-insider.ai/who-blames-the-bot-the-olg-hamm-ruling-and-the-reality-of-ai-liability-in-professional-services/))

The amounts in these specific cases were small. The principle is not. Strip away the jurisdictions and the facts and you get one rule a buyer can plan around: **a statement your automated agent makes is a statement your company made.** Hallucination doesn't transfer the liability to the model vendor, and "users should have known to double-check" hasn't carried the day.

## "The chatbot did it" is not a defense — and it never was the strong defense anyway

It's tempting to read these as exotic legal edge cases. They aren't. They're the predictable result of consumer-protection law that already existed, applied to a new interface.

In the United States, the framing is converging on the same place from a different direction. Legal and regulatory commentary through 2025–2026 treats statements made by a company's chatbot as representations of the company itself, which means an inaccurate or unsubstantiated one can trigger the same unfair-and-deceptive-practices exposure as any other false claim a business makes — and a growing list of states (California, Colorado, and others) now require disclosure that a customer is talking to AI rather than a human. ([Arnall Golden Gregory, AI chatbot compliance, 2026](https://www.agg.com/news-insights/publications/ai-chatbot-compliance-key-legal-risks-and-regulatory-considerations-for-businesses-in-2026/)) Disclosure laws don't fix wrong answers, but they remove a defense — the customer can't be assumed to have known they were getting machine output if you never told them.

The deeper point is that the "blame the bot" argument was never going to work, because it doesn't match how anyone experiences support. A customer who reads a confident answer on your site, in your brand's voice, in your chat widget, reasonably believes it's your answer. Courts have simply confirmed the obvious. So the useful question isn't "can we disclaim liability for the bot?" It's "what do we put in place so the bot rarely says something we'd be on the hook for, and so we can prove what it said and why when it does?"

## This is an operations problem before it's a legal one

The legal exposure is the headline, but it's the rarest outcome. The everyday cost of an unaccountable AI agent shows up long before a tribunal does.

- **A wrong answer becomes a promise you have to negotiate out of.** Once the agent states a discount, a policy, or an eligibility rule, your team is stuck choosing between honoring a fabricated commitment or telling a customer the AI misled them. Both are bad, and both land on a human after the fact.
- **Reputational damage compounds faster than human error.** A confidently wrong machine answer tends to be screenshotted and shared in a way a flustered human reply isn't — and audiences are measurably less forgiving of AI mistakes than equivalent human ones.
- **The damage often hides inside your "deflection" numbers.** A bot that gives a plausible-but-wrong procedure looks like a resolved conversation on the dashboard, then resurfaces days later as a new ticket that never mentions the bot. Deflection-rate guides that account for re-contacts consistently find real self-service resolution well below the headline deflection figure — the gap is where the silent wrong answers live. ([eesel AI, deflection rate explained](https://www.eesel.ai/blog/deflection-rate-what-is-it-and-how-to-improve-it))

None of these require a lawsuit to hurt you. They're the routine tax on deploying an agent that can say anything, with no record of why it said it.

## The accountability gap most teams can't close

Here's a test. Pick a real answer your AI agent gave a customer last week and ask three questions:

1. **Which source did this answer come from?** Not "the knowledge base" — the specific document or passage.
2. **Was the agent allowed to answer this at all,** or was it a topic that should have gone to a human?
3. **If a customer disputes it next month, can we reconstruct exactly what was said and what it was based on?**

If you can't answer all three quickly, you don't have an AI support agent — you have an unmonitored spokesperson. Closing that gap is what "accountability" means in practice, and it comes down to an operating model, not a model upgrade. The same controls that make an agent accurate are the ones that make it *defensible*: grounding, citations, scope, handoff, disclosure, and a transcript you can produce.

## Five controls that keep a wrong answer from becoming a liability

Think of these as the answer-accountability stack. Each one does double duty — it reduces the chance of a wrong answer, and it limits the blast radius when one slips through.

### 1. Ground every answer in sources you control

An accountable agent doesn't answer from the model's training data. It retrieves the relevant passages from *your* content — help center, policy PDFs, product docs — and is instructed to answer only from those. This is the difference between "what does a refund policy usually look like?" and "what does *this company's* refund policy say?" Grounding is the single biggest lever on both accuracy and accountability, because an answer that traces to a source you own is an answer you can stand behind — or correct at the source. ([why grounded answers matter](/blog/why-grounded-answers-matter/))

### 2. Make every answer cite its source

Citations are what turn "the bot said something" into "the bot said this, from here." When each public answer carries a clickable reference to the exact passage it used, the customer can verify it, your team can debug a bad answer in seconds, and — critically for the accountability question — you have a record of the basis for every statement. A citation is the audit trail the OLG Hamm and Air Canada cases would have wanted: proof of what the answer was grounded in. If the agent can't cite a source, that's the signal it shouldn't have answered.

### 3. Define scope, and let the agent refuse outside it

Most liability-grade mistakes happen on a short list of high-stakes topics: pricing, refund eligibility, contractual terms, anything regulated, anything account-specific. Decide explicitly which topics the agent may answer and which it must hand off, and configure it to say "I don't know, let me get someone who can help" when a question falls outside its sources or its scope. A clean refusal is a success. An agent tuned to be maximally helpful will treat "be helpful" as "always produce an answer" — which is exactly how the Air Canada bot invented a policy.

For the few answers where a wrong response is unacceptable, don't rely on retrieval at all — author the exact wording and pin it, so those questions return a vetted response verbatim instead of a generated one.

### 4. Hand off before the agent improvises

The handoff is the safety valve for everything grounding and scope don't cover. It should trigger *before* the model is tempted to fill a gap — on low confidence, on out-of-scope topics, on anything touching money or legal commitments — and it should pass the full conversation context to a human so the customer doesn't start over. A handoff that carries context turns a potential wrong answer into a routed one. A dead-end "please email support" is barely better than the bot guessing. ([AI support handoff playbook](/blog/ai-support-handoff/))

### 5. Disclose the AI, and keep the transcript

Two housekeeping controls that matter more than they look. **Disclosure** — telling the customer they're talking to an AI — is now a legal requirement in a growing number of jurisdictions, and it's free to implement. **Logging** — keeping a durable, reviewable record of every conversation, the answer given, and the source it cited — is what lets you answer that third accountability question when a dispute arrives. "We sample and review our agent's answers weekly, and here are the records" is a far stronger position than "we assumed the vendor handled it." Persistent, reviewable session history is the difference between an incident and a crisis.

## An answer-accountability checklist

Before you let an AI agent speak to customers in your name, you should be able to check every box. If you're evaluating vendors, ask these on your own content during a trial — not from the marketing site.

- [ ] **Every public answer cites the specific source it came from.** No citation means no audit trail.
- [ ] **The agent answers only from content we control,** not the model's general knowledge.
- [ ] **High-stakes topics are scoped out or pinned** to vetted, exact answers — pricing, refunds, contracts, compliance.
- [ ] **The agent refuses cleanly** when the answer isn't in its sources, instead of guessing.
- [ ] **Handoff triggers before improvisation** and carries full context to the human.
- [ ] **Customers are told they're talking to an AI,** at or before the start of the conversation.
- [ ] **Every conversation is logged and reviewable,** so we can reconstruct what was said and why.
- [ ] **Someone owns the review loop** — a named person samples answers on a schedule and fixes the source when one is wrong.

A vendor that can support all eight is selling accountability. One that answers "our model is very accurate" is selling hope — and hope is what Air Canada argued in front of the tribunal.

## Where Owlish fits

Owlish is our product, so read this as a vendor being specific rather than a neutral referee.

Owlish is built on the premise that a confident wrong answer is the worst thing a support agent can do — so the controls above are the product, not add-ons. The agent answers only from the websites, documents, and PDFs you ingest, so replies stay inside content you control. ([knowledge base overview](/docs/knowledge-base/overview)) Every public answer can carry a [citation](/docs/knowledge-base/citations) back to the exact source passage, which is both a trust signal and the audit trail these rulings reward. When an answer isn't in your knowledge, the agent is configured to say so and [hand off to a human](/docs/helpdesk/human-handoff) with the full conversation attached, rather than improvise. For the handful of questions where a wrong answer is unacceptable, you can pin an exact response. Conversations are stored as persistent, reviewable sessions, so reconstructing what was said and what it was based on is a lookup, not an investigation. You deploy it as a [web widget](/docs/deploy/widget) or through chat channels like Slack and Microsoft Teams, and pricing is flat and session-based — **Starter is $49/mo monthly or $39/mo billed annually (save ~20%)**, with a free tier so you can test grounding and citations on your own content before paying anything.

That shape suits small and growing teams that want to automate repetitive questions without betting their reputation on the model's restraint. It's a weaker fit if you need a full contact-center suite with telephony and CRM-grade ticket routing, or if nearly every conversation is judgment-heavy enough to need a human anyway — in that case a larger platform or a human-first setup is the better call, and you should choose one. Owlish is the grounded, accountable answering-and-handoff layer, not the entire helpdesk.

## FAQ

### Can a business be held liable for what its AI chatbot says?

Yes. In *Moffatt v. Air Canada* (British Columbia Civil Resolution Tribunal, February 2024), the airline was ordered to compensate a customer after its chatbot described a policy that didn't exist, and the tribunal rejected the argument that the bot was a separate entity. In May 2026, Germany's Oberlandesgericht Hamm reached the same conclusion, holding that a chatbot is a tool of the operator and its statements are attributed to the company even when the AI hallucinates. The consistent takeaway: treat anything your agent can say as something your company is saying.

### Does "the AI made a mistake" reduce our responsibility?

No. Both rulings turned on the idea that the company controls the system and therefore owns its output. The OLG Hamm court was explicit that liability applies even when the AI hallucinates or mishandles correct input, because the operator defines the bot's scope and can change it. The model being imperfect is treated as a known property of a tool you chose to deploy, not an external actor you can point to.

### Do we have to tell customers they're talking to an AI?

Increasingly, yes — several jurisdictions now require disclosure that a user is interacting with AI rather than a human, and the list is growing. Beyond the legal requirement, disclosure removes a defense customers might otherwise lose: they can't be assumed to have known to double-check machine output you never labeled. It's a low-cost control with outsized benefit, so implement it regardless of where you operate. Check current requirements for your specific markets with counsel.

### What's the single most important control for reducing chatbot liability?

Grounding answers in sources you control, with a visible citation on every answer. Grounding cuts the rate of wrong answers; the citation gives you a record of what each answer was based on. Together they convert an opaque "the bot said something" into a verifiable, correctable, defensible statement — which is exactly what the courts have been asking companies to be able to produce.

### Is this just a problem for big companies like airlines?

No — arguably the opposite. Large companies have legal teams, escalation processes, and the cash to absorb a bad ruling. A small business that lets an ungrounded bot promise a refund or misstate a policy can take the same reputational and financial hit with far less cushion. The controls in this post are deliberately achievable for small teams; the cost of skipping them scales down poorly.

---

*Sources were reviewed in June 2026. Court decisions and AI-disclosure requirements change; verify the current state of any legal or regulatory point with qualified counsel for your jurisdiction before relying on it. This article is general information, not legal advice. Company and product names referenced are trademarks of their respective owners; Owlish is not affiliated with or endorsed by them.*

Want an agent that answers only from your content, cites its sources, and hands off cleanly when it should? [Build your first Owlish agent](/docs/quick-start/build-your-first-agent) on your own knowledge base and see what grounded, accountable support looks like before you commit.

---

Source: https://owlish.bot/blog/ai-chatbot-liability/