AI Chatbot Liability: Who Answers for Your AI Agent?

In most cases, yes: the company that puts an AI agent in front of customers answers for what that agent tells them. In Moffatt v. Air Canada, a Canadian tribunal held the airline responsible for its chatbot’s wrong answer, just as it would be for a wrong web page. A disclaimer does not move that responsibility to the bot or its vendor.

This article explains the cases, rules and practical controls for support leaders. It is general information, not legal advice: for your own market and contracts, ask a lawyer.

In short

• AI chatbot liability starts with an old rule: a business is responsible for the information it gives customers, whichever channel or system gives it.
• In Moffatt v. Air Canada (2024), the tribunal called the idea that a chatbot is responsible for its own actions “a remarkable submission” and ordered the airline to pay.
• An “AI can make mistakes” disclaimer is not a shield: unfair terms rules, product liability rules and regulators all limit what it can do.
• When a customer disputes an AI promise, the side with the transcript, the policy version and the agent version in force that day has the stronger position.
• The safest operating rule is to honor, correct and review: keep reasonable promises, contact everyone who got the same wrong answer and fix the cause before the next conversation.

Is a company liable for what its AI chatbot tells customers?

Yes, in the decided case most people cite. In Moffatt v. Air Canada, British Columbia’s Civil Resolution Tribunal treated the chatbot as part of the airline’s website and found the airline liable for negligent misrepresentation when a customer relied on its wrong answer.

The facts are simple (decision, 2024 BCCRT 149, issued February 14, 2024). In November 2022, after a death in the family, a customer asked Air Canada’s website chatbot about bereavement fares. The chatbot said they could book now and submit the ticket for a reduced rate within 90 days of issue, even after travel. The linked bereavement page said the opposite. An Air Canada representative later admitted in an email that the chatbot had used “misleading words” (paragraph 22).

The tribunal tested the claim against the five parts of negligent misrepresentation (paragraph 25). Each one maps onto a support conversation:

Air Canada argued it could not be held liable for its chatbot. The tribunal answered: “It should be obvious to Air Canada that it is responsible for all the information on its website. It makes no difference whether the information comes from a static page or a chatbot” (paragraph 27). The order came to CA$812.02 with interest and fees (paragraph 44).

Keep the case in proportion. It is a small claims decision in one Canadian province, it binds no court elsewhere, and it did not consider whether the chatbot’s vendor owed anything to the airline. Its value is the reasoning, which applies the ordinary rules every business already lives under. For the operating model around this, start with the AI agent governance guide.

Who is liable when AI goes wrong: the company, the vendor or the AI?

Toward the customer, the company that deploys the AI agent is liable in most situations, because the agent speaks for it. The AI itself is not a legal person. The vendor’s share is decided mostly by your contract, and in the EU by product liability rules that cover personal injury, property damage and lost data, not a wrong refund answer.

Three layers decide who pays:

1. You and the customer. Consumer law, contract and misrepresentation rules apply to what your channels say. The customer does not need to know which vendor built the model.
2. You and the vendor. Warranties, indemnities and liability caps in the vendor agreement decide whether you can recover a loss. Many vendor terms limit or exclude liability for output accuracy, so read them before launch, not after an incident.
3. The EU product liability layer. The revised Product Liability Directive, (EU) 2024/2853, states that “software” is a product (Article 4(1)), whether it runs on a device, in the cloud or as a service (recital 13). It applies to products placed on the market or put into service after December 9, 2026 (Article 2(1)), the same date Member States must have it in national law (Article 22(1)).

The directive matters less for support than its headlines suggest. It covers death, personal injury including damage to psychological health, damage to property and the destruction or corruption of data not used for professional purposes (Article 6). Recital 24 says pure economic loss does not by itself trigger liability under it. A wrong refund or fee promise is pure economic loss, so it stays with consumer and contract law. Where the directive does apply, liability toward the injured person cannot be limited or excluded by a contract term (Article 15).

The difference between a chatbot and an AI agent matters here. A chatbot that answers only makes statements. An agent that issues refunds, cancels orders or changes accounts also creates transactions, so a mistake is no longer only a misleading answer: it is an action your systems carried out. That is why the actions an agent may take alone belong in an autonomy matrix, not only in the prompt.

Does an “AI can make mistakes” disclaimer protect the company?

Not much, and not on its own. A disclaimer can help customers judge an answer and supports transparency duties, but it does not make a misleading answer accurate. Courts and regulators look at what the customer was told and whether relying on it was reasonable, and consumer law limits terms that shift that risk onto the customer.

Four limits show up again and again:

• Reasonable reliance survives a link. The Air Canada chatbot’s answer linked to the correct policy page. The tribunal still asked why customers “should have to double-check information found in one part of its website on another part” (paragraph 28).
• Terms have to be proven and fair. Air Canada pointed to its tariff but did not file the relevant text, so it “has not proven a contractual defence” (paragraph 31). In the UK, an unfair term in a consumer contract “is not binding on the consumer” (Consumer Rights Act 2015, section 62).
• Some liability cannot be excluded at all. Under the EU Product Liability Directive, a contract term cannot limit or exclude liability toward the injured person (Article 15).
• Regulators look past labels. When the FTC announced its Operation AI Comply sweep on September 25, 2024, its then Chair, Lina Khan, said “there is no AI exemption from the laws on the books.”

A disclaimer that tells customers they are talking to an AI is a different thing, and in the EU it is now required. Keep it, and keep it honest. Just do not treat it as the control that keeps you out of trouble. The control is an agent that does not make the wrong promise.

What do US, EU and UK rules expect from AI agents in customer service?

All three expect the same core behavior: tell customers the truth, do not mislead them, and stay accountable for the systems that talk to them. The EU adds an explicit AI disclosure duty, the UK’s FCA applies its Consumer Duty to financial firms using AI, and US regulators enforce existing consumer protection law against AI claims and chatbot failures.

The UK expectations are the most concrete for support teams in regulated firms. In its AI Update of April 2024, the FCA wrote that AI chatbots can help customers understand products, but “firms should consider their obligations under the Consumer Duty” (paragraph 3.25). It points to the consumer understanding rules on meeting customers’ information needs (paragraph 3.36), and says that in banks, insurers and enhanced SM&CR firms any use of AI in a business area falls within a Senior Manager’s responsibilities (paragraph 3.40). In practice: one named owner per agent, outcome data on its conversations, and a record that the board looked at it.

The CFPB spotlight states that it is not an interpretation of any regulation. For the EU detail, including risk classes and dates, see the EU AI Act guide for customer service.

Which commitments may an AI agent make to customers?

An AI agent may explain policy in the words of the approved policy text, confirm facts that a system has just returned, and commit to an outcome only when a tool has carried it out within its limits. It should never grant exceptions, promise compensation, give legal or financial advice, or quote a date or amount it did not get from a system.

The table works because it turns “be accurate” into rules a reviewer can check against a transcript. The full set of limits, with numbers per action, is in the guide to AI agent guardrails, and the types of invented answers to look for are in how to catch AI agent hallucinations.

What evidence do you need when a customer disputes an AI promise?

You need to show what the agent said, word for word and with a timestamp, which policy and content it should have used that day, which version of the agent ran, what any tool actually did, and how the conversation was checked. Without that record, the customer’s screenshot is the only evidence, and it usually wins.

The Air Canada decision shows what an empty file costs. The customer filed a screenshot of the chatbot’s answer. The airline “did not provide any information about the nature of its chatbot” (paragraph 14), did not file the tariff it relied on (paragraph 31), and offered no evidence of the real bereavement fare, so the tribunal drew an adverse inference and used the customer’s figure (paragraph 36).

Keep these seven records for every AI agent conversation, retained as long as your complaint and legal teams say:

1. The full transcript, unedited, with timestamps and the channel.
2. The agent version: model, prompt or instructions, and configuration in force.
3. The content version: which help articles and policy texts the agent could use that day.
4. Tool calls and results: what the agent asked a system to do, and what the system returned.
5. The AI disclosure the customer saw, and whether a human handover was offered.
6. The review record: score, reviewer or automated check, and any flag raised.
7. The fix log: what changed after a flag, when, and who approved it.

The hardest record to rebuild later is the link between dates and versions: “which return window did the agent apply on March 3?” Write it down at every release. AI agent policy management covers keeping policy versions and agent content in step.

How should you handle a wrong promise from your AI agent?

Honor, correct, review. Honor the promise when the customer reasonably relied on it and the cost is proportionate. Correct it with every other customer who got the same answer, before they act. Review the cause and confirm the fix on new conversations. Fighting a single small claim usually costs more than it saves.

Honor. Set a written rule in advance, owned by the Head of CX with legal: for example, honor any AI agent promise under a set amount where the customer acted on it, and send larger ones to a named approver within one working day. Air Canada offered a $200 coupon, which the customer did not accept, and the tribunal refused to offset it (paragraph 41). A goodwill gesture is not a settlement unless the customer agrees to it.

Correct. Search or score every conversation on the same topic since the error began, list the customers who got the wrong answer, and contact those who have not yet acted. The AI agent resolution audit explains how to check whether “resolved” conversations were really resolved.

Review. Find the cause (a stale article, a missing rule, a prompt change), fix it, re-test, then check the next conversations on that topic to confirm the fix held.

A worked example shows why the correct step pays. The numbers are an illustration, not data from a real company.

• An AI agent handles 300 cancellation conversations a day. A content update wrongly gives a 60-day full refund window; the policy is 30 days.
• The error runs for 12 days: 3,600 cancellation conversations. In 15% of them, the customer is between day 31 and day 60, and the agent says a full refund is still possible: 540 customers.
• If 40% of them act before anyone notices, 216 refund claims arrive. At an average of $90, honoring them costs $19,440.
• The other 324 customers have not acted yet. Contacting them the day the error is found, with the correct policy and an apology, stops most of the remaining cost and keeps them from reasonably relying on the wrong answer.

A 2% random sample would have reviewed about 72 of those 3,600 conversations, and about 11 would contain the error. That can spot the problem. It cannot give you the list of 540 names, and the list is what the correct step needs.

The AI agent liability-risk checklist

Use the checklist below to review an AI agent before launch and once a quarter after. Each row names a risk, a support example, the control that prevents it and the evidence you should be able to produce if a customer, auditor or regulator asks. Delete rows that do not apply to your agent, and give every row an owner.

Rows 1, 2, 3, 11 and 12 are where most support liability starts: a wrong answer about money, repeated, with no record to explain it.

When this checklist is not for you

If your AI agent only answers from public help articles, takes no actions and hands every money, account or complaint question to people, rows 1, 4, 5 and 11 cover most of your exposure. Add the rest the day the agent gets a refund, credit or cancellation tool, or starts serving a regulated product.

The checklist also does not replace legal review. Consumer law, financial rules and sector regulators differ by country, and a lawyer who knows your market should read your disclosure, your terms and your vendor contract.

Where Kaizo fits

Kaizo checks every support conversation, human or AI, against your own rubric, and shows you what to fix. It covers the evidence for rows 2 and 12: AutoPilot scores every resolved conversation against your own scorecard, and every score links to the lines in the conversation. See QA for AI agents for how it works, or book a demo to see it on your own conversations.

How to start this week

You can start this week without a new tool: five steps show where your AI agent can create liability today and which records you could not produce if a customer disputed an answer tomorrow.

1. Pull 50 conversations from last week on refunds, fees, cancellations and complaints, and mark every statement that is a commitment.
2. Check each commitment against the explain, confirm, commit, exception and advice table above.
3. Run a retrieval test: pick one conversation from a month ago and try to produce the transcript, agent version and policy version within an hour.
4. Write the honor rule: the amount up to which a wrong AI promise is honored without debate, and who approves above it.
5. Copy the 14 rows, give each an owner, and bring the gaps to your next AI agent governance review.

Frequently asked questions

What is accountability in artificial intelligence?

Accountability in AI means a named person or organization answers for what an AI system does, can explain it and must put it right when it goes wrong. In customer service, that is the company deploying the agent, with one owner per agent, records of what it said and did, and a process to correct mistakes.

Are AI agents treated differently from chatbots for liability?

The principle is the same: the company answers for both. The exposure differs. A chatbot that only answers makes statements a customer may rely on. An AI agent that refunds, cancels or changes accounts also takes actions, so a mistake can move money or data directly. That is why actions need limits enforced in the tools themselves.

Can we recover the cost of an AI agent’s mistake from the vendor?

Only if your contract allows it. Look at the vendor’s warranties, indemnities and liability cap before launch, because many terms limit liability for output accuracy. In the EU, the revised Product Liability Directive adds no-fault claims against manufacturers of defective software for injury, property and data damage, not for a wrong refund answer.

Does reviewing AI conversations reduce our liability?

Review does not transfer responsibility, but it changes two things. It finds wrong answers early, so fewer customers rely on them. And it produces evidence of reasonable care: in the Air Canada case, the lack of care to ensure the chatbot was accurate was part of the finding against the airline.

Do we have to tell customers they are talking to an AI?

In the EU, yes: since August 2, 2026, people must be told from the first interaction, unless it is obvious. Elsewhere it is good practice, and it supports the honesty duties consumer and financial regulators already apply. A disclosure informs the customer. It does not excuse a wrong answer.