By the Editors · Published 15 June 2026 · 14 minute read

My AI agent gave a customer wrong advice. Am I legally required to compensate them?

Your AI agent gave a customer information that turned out to be wrong. They relied on it, they are now out of pocket, and they are asking for compensation. The question every SME operator faces at this point is the same: do I have to pay, do I want to pay, and will my insurance cover it? The answer to each of those three questions is different, and getting the sequence right matters enormously for what comes next.

Key takeaways

If your AI agent made a specific representation that a customer relied on to their financial detriment, you are likely legally required to compensate them. Moffatt v. Air Canada (BC Civil Resolution Tribunal, 2024) established that a business cannot disclaim responsibility for its AI agent's statements by treating the agent as a separate entity. The operator is responsible for what the agent tells customers.^[1]
Three elements must all be present for a legally enforceable compensation claim: the AI made a specific representation of fact, the customer relied on it in a way that was objectively reasonable, and the customer suffered quantifiable financial loss as a direct result. If any one of these is absent, the legal obligation to compensate is significantly weaker.
Your professional indemnity (PI) or technology errors and omissions (E&O) policy is the most likely insurance to respond. General commercial liability policies rarely cover pure economic losses from advice. However, many 2025 and 2026 renewals include AI exclusion endorsements that may remove cover for AI-generated output. Check your policy wording before assuming you are covered.^[2]
A voluntary, early compensation often costs less in total than a defended claim. The Air Canada chatbot case cost Air Canada the disputed amount plus legal costs and significant reputational damage, all of which were avoidable had the company resolved the ZAR 812 CAD dispute promptly.
Notify your insurance carrier of a potential claim event within the window your policy requires. Late notification is one of the most common grounds for coverage denial. Do not make any admission of liability or offer payment before speaking to legal counsel and your carrier.

Section 1: When you are legally required to compensate

Whether you are legally required to compensate a customer depends on the specific facts of what your AI agent said, what the customer did in response, and what damage resulted. The legal analysis is not complicated in principle, but the facts matter enormously. Most SME operators make the mistake of either assuming they must pay everything or assuming their disclaimer protects them from paying anything. Both are wrong.

The three elements of a compensation claim for wrong AI advice are: first, a specific representation of fact (not merely an opinion or general guidance); second, reasonable reliance by the customer on that representation; and third, quantifiable financial loss flowing directly from that reliance. All three must be present. If the AI stated something that was factually specific and provably wrong, and if a reasonable person in the customer's position would have acted on it, and if the customer did act on it and suffered a financial loss they can document, the legal case for compensation is strong.

The Air Canada case is the clearest available example. Air Canada's chatbot told a customer named Jake Moffatt that he could purchase bereavement-rate tickets and claim the discount retroactively within 90 days. That was factually wrong: the policy required booking at the bereavement rate first. Moffatt bought a full-price ticket, submitted a refund claim, and was refused. The tribunal found that Air Canada was responsible for its chatbot's incorrect statement and awarded the discount plus costs.^[1]

The cases where you are likely not legally required to compensate are where the AI gave a general statement or opinion rather than a specific representation of fact, where the customer did not rely on the AI's statement but acted for other reasons, where the customer suffered no quantifiable financial loss, or where the customer's reliance was not objectively reasonable given context that was available to them. A customer who ignored a prominent disclaimer stating that AI responses should be verified with a qualified professional, and then acted on an AI output without verification, faces a harder case on reasonable reliance. This is fact-specific and not automatic.

Section 2: The three questions that determine your legal position

Before responding to a compensation request, work through these three questions in order.

Question 1: Was the AI's statement specific enough to constitute a representation of fact?

A representation of fact is a statement about an existing or past state of affairs that can be true or false. "Your order will arrive by Thursday" is a specific representation. "Our AI assistant will try to help you" is not. "Your subscription costs EUR 49 per month and includes international transfers" is a representation of contract terms. "AI is changing how businesses work" is not.

General guidance, opinions, and educational content provided by an AI agent carry less legal weight as representations. Courts look at whether a reasonable customer would have understood the AI's statement as a commitment rather than as general information. If your agent is presented to customers as an authoritative assistant for specific operational questions, the bar for what counts as a representation is lower than if it is presented as a general information tool.

Question 2: Was the customer's reliance on the AI objectively reasonable?

Reasonable reliance is assessed against what a person in the customer's position, with the customer's level of knowledge, would have understood the AI's statement to mean and would have done in response. Courts look at: whether a prominent disclaimer was visible at the time of the interaction; whether the nature of the information was such that a reasonable person would seek independent verification before acting; and whether the customer had prior experience with your system that should have put them on notice about its limitations.

A disclaimer that appears in a footer link to your terms of service, which the customer has not read, provides limited protection. A disclaimer that appears in the chat interface itself, immediately before the AI responds to a question about pricing or eligibility, provides substantially more. If your AI is deployed without any disclaimer at the point of interaction, you have fewer arguments against reasonable reliance claims.

Question 3: Is the loss quantifiable and directly caused?

Compensation claims require a demonstrable financial loss. A customer who bought something based on wrong AI information and wants their money back has a quantifiable loss. A customer who says they suffered stress, inconvenience, or reputational damage has a harder case unless they can evidence it in financial terms.

Causation is also required. The customer must show that the AI's wrong information was the reason they took the action that led to the loss, not just one of several factors. If they would have taken the same action regardless of what the AI said, causation is weak. If the AI's information was the sole or main reason they acted, causation is strong.

Section 3: What your insurance may cover

The insurance question is separate from the legal question, and operators often confuse the two. You may be legally required to compensate a customer but your insurance does not cover it. You may not be legally required to compensate but you choose to do so, and whether insurance reimburses you depends on your specific wording.

The policy types most likely to respond to an AI wrong-advice claim are professional indemnity (PI) and technology errors and omissions (E&O). Both are designed to cover losses arising from advice, information, or services provided by the insured that turn out to be wrong or inadequate. An AI agent giving wrong factual information to a customer, resulting in a compensation payment, is exactly the kind of event these policies were originally designed to cover.

However, the market has been changing rapidly. From 2024 onwards, many insurers have been adding AI exclusion endorsements to PI and E&O renewals that carve out losses arising from the operation of an AI system or the output of a generative AI model.^[2] If your renewal includes such an endorsement, the question is whether the loss from your AI agent's wrong advice is caught by the exclusion. The answer depends on the precise drafting of the exclusion and the precise characterisation of the loss.

General commercial liability (CGL) policies, including public liability, are much less likely to respond. They are designed to cover bodily injury and property damage, not pure economic losses from advice. Some policies have a "products and completed operations" section that could theoretically apply to AI software, but this is unlikely to be the main route to recovery.

The practical step is to read your current PI and E&O policy wording, identify whether there is an AI exclusion, and ask your broker in writing whether the specific event of your AI agent giving a customer wrong factual information leading to a compensation payment is covered or excluded. Get the answer in writing. The coverage question should be resolved before you take any action that could affect your insurer's position.

For an overview of which SME policy types are most at risk of AI-related coverage gaps, see the 2026 policy-by-policy guide and the SME guide to policy exclusions.

Section 4: The practical case for voluntary compensation

Even where the legal obligation is uncertain, there is often a strong practical case for voluntary compensation. The Air Canada tribunal case cost the company the original disputed amount (approximately CAD 812 in fare discounts) plus legal costs, reputational damage from wide media coverage, and the precedent value of a published tribunal decision that is now cited in every AI liability discussion globally. The original dispute was about a few hundred dollars. Had Air Canada simply compensated the customer when he first raised the issue, none of the rest would have followed.

For an SME operator, the calculus is similar. A compensation payment of EUR 500 to resolve a dispute where your AI gave a customer wrong pricing information costs EUR 500. The same dispute, defended and lost, may cost that amount plus your lawyer's time, plus the claim on your insurance record, plus any court costs, plus the time and distraction of managing the dispute across months. A dispute that reaches a small claims court or consumer tribunal may also result in a public record. Voluntary early resolution avoids all of that.

Voluntary compensation also creates a better claims record. When you later make a claim under your PI or E&O policy, insurers look at your claims history. A history of small, promptly resolved customer service issues handled professionally is different from a history of protracted disputes. How you manage small claims tells underwriters something about how you would manage larger ones.

The decision to compensate voluntarily is not a legal admission of liability unless you explicitly state it as one. A payment made "without admission of liability" or "as a goodwill gesture" does not, in most jurisdictions, create a legal precedent that you were liable or that you accept liability in any future case. Your lawyer can confirm the appropriate framing for the jurisdiction.

Section 5: The notification rules that determine whether insurance applies

The single most important action in the period immediately after a customer raises a wrong-advice claim is to notify your insurance carrier. This is not a step to take after you have decided what to do about the customer. It is the first step.

Most professional indemnity and errors and omissions policies contain a "claims made" coverage trigger: the policy in force when the claim is first made responds, not the policy in force when the underlying advice was given. Most also contain a notification condition requiring the insured to notify the carrier of any claim, potential claim, or circumstance likely to give rise to a claim, within a specified period or "as soon as practicable." Failure to meet this notification condition is one of the most common grounds on which claims are denied.

At the point when a customer tells you your AI gave them wrong information and they want compensation, you are aware of a circumstance that may give rise to a claim. That awareness triggers the notification clock. Notify your carrier in writing, even if you have not yet decided whether the claim is valid or whether you intend to pay. The notification is not an admission of anything. It is a procedural step that preserves your rights.

If you have a broker, call the broker first. The broker will guide you through the notification process and help you frame the circumstances in a way that is accurate and complete. Do not speak directly to the carrier before speaking to the broker unless you have been specifically instructed otherwise.

At the same time, assess whether the incident triggers any regulatory reporting obligation. If the wrong advice related to personal data processing and a data subject's rights, it may involve a GDPR notification obligation to your supervisory authority. If the AI system operates in a regulated sector, there may be a parallel sector notification requirement. These notifications are independent of the insurance notification and have their own timelines.

Section 6: What to say to the customer

The customer wants a response. What you say, how quickly you say it, and what you do not say all matter for both the insurance and the legal position.

Acknowledge the situation promptly. A customer who feels ignored escalates. A customer who receives a prompt, professional acknowledgement is more likely to resolve the situation without legal action. The acknowledgement does not need to be an admission of anything. It can simply confirm that you have received their complaint, that you take it seriously, and that you will respond properly within a specific timeframe (typically 5 to 10 business days for an SME).

Do not make any statement about whether you agree that the AI was wrong, whether you intend to compensate, or how much. Those decisions require legal review and insurance consultation first. Any premature statement about liability or compensation can affect your position in ways that are difficult to undo.

Do not blame the AI provider. You are the operator. Telling a customer that it was the AI's fault, or that they should take the matter up with the model provider, is legally incorrect (as Moffatt v. Air Canada confirmed) and will not resolve the dispute. The customer's relationship is with you, not with the model provider.

Once you have taken legal and insurance advice and decided how to respond, communicate clearly, specifically, and in writing. If you are compensating, state the amount and the basis. If you are not compensating, explain why in terms the customer can understand, and make clear what options they have if they disagree. A clear written response, even one that the customer does not like, closes the issue more often than a vague or delayed one.

For a more detailed guide to the 48-hour window after any AI agent incident, see the SME incident response guide. For the full legal analysis of operator liability under UK, EU, and US frameworks, see AI agent liability under EU law on agentliability.eu.

Section 7: Building to avoid this situation next time

The most effective response to an AI wrong-advice incident is the one you prepared for before it happened. Operators who have built clear scope constraints into their agent, documented what the agent is and is not authorised to tell customers, and implemented logging that preserves the conversation are in a dramatically stronger position when a complaint arrives than operators who deployed an unconstrained agent and are discovering after the fact what it said.

Three practical steps reduce both the frequency of this situation and the cost when it occurs. First, constrain your agent's scope explicitly. If it handles pricing queries, ensure it cannot make statements about eligibility, policy terms, or service features it does not have real-time verified access to. The Air Canada chatbot had access to the general policy but not to the bereavement policy terms. That gap produced the claim.

Second, implement logging. You need to be able to retrieve what the agent said to a specific customer at a specific time. Without logs, you cannot assess the validity of a complaint, cannot prepare an insurance notification, and cannot defend any subsequent claim. Most managed AI deployment platforms retain conversation logs by default. If yours does not, this needs to change before your next incident.

Third, add a contextual disclaimer at the point where the agent responds to specific queries about prices, eligibility, or service terms. Not a buried footer. A visible, plain-language note in the chat interface that direct commitments should be confirmed with your team before the customer takes action. This does not eliminate liability but it shifts the reasonable reliance analysis in your favour.

For a structured self-assessment of your current AI agent setup before an incident occurs, see the pre-deployment insurance checklist.

Frequently asked questions

If my AI agent gave a customer wrong information, am I legally required to compensate them?

In most cases, yes, if the customer can show that they relied on the wrong information to their financial detriment and the reliance was reasonable. Moffatt v. Air Canada (BC Civil Resolution Tribunal, 2024) confirmed that a business cannot disclaim responsibility for its AI agent's statements by treating the agent as a separate entity. The operator is responsible for representations made by an automated system it places in front of customers. Your liability runs under contract law if there was a commercial relationship, or under the law of misrepresentation if the wrong information induced a transaction.

Does my business insurance cover a compensation payment to a customer whose AI gave them wrong advice?

It depends on your policy type and the wording. Professional indemnity (PI) and technology errors and omissions (E&O) policies are the most likely to respond to a claim arising from wrong AI advice, but many renewals in 2025 and 2026 include AI exclusion endorsements that carve out losses arising from AI-generated output. General commercial liability policies rarely cover pure economic losses from advice. Check your PI and E&O policy wording for any AI exclusion before assuming you are covered.

Can a customer sue me for what my AI agent said?

Yes. A customer can bring a claim for breach of contract, misrepresentation, or negligence depending on the nature of the advice and the relationship. The Air Canada chatbot case (Moffatt v. Air Canada, 2024) resulted in a successful claim against the company for representations its chatbot made. In the EU, the revised Product Liability Directive (Directive 2024/2853, applicable from December 2026) treats AI software as a product, giving claimants access to strict liability routes that do not require proving fault.

What should I do in the first 48 hours after my AI agent gives a customer wrong advice?

First, preserve the evidence: save the conversation logs, the agent's current system prompt and configuration, and a timestamped record of when you became aware. Second, notify your insurance carrier of a potential claim event. Most policies require prompt notification and deny late claims. Third, do not make any admission of liability or offer payment to the customer before speaking to legal counsel and your carrier. Fourth, assess whether the incident triggers any regulatory reporting obligation under GDPR, the EU AI Act, or sector-specific rules.

Does a disclaimer in my terms of service protect me from paying compensation for AI errors?

Partially and unreliably. A disclaimer can reduce exposure for incidental or consequential losses where the customer was clearly advised to verify important information before acting. But a disclaimer that attempts to limit liability for personal injury or death caused by negligence is unenforceable in the EU under the Unfair Contract Terms Directive and its national implementations. Courts also consider whether the disclaimer was clearly communicated at the point where the customer was likely to rely on the AI's output. A buried disclaimer that customers cannot reasonably be expected to have noticed provides limited protection.

What is the difference between a legal obligation to pay and a practical reason to pay?

A legal obligation arises from the specific facts of the customer's claim: whether a contract was breached, whether there was a misrepresentation, whether reliance and damage are provable. A practical reason exists even when the legal case is arguable: early resolution costs less than litigation, preserves customer relationships, and creates a better claims record with your insurer. Many AI-related customer disputes are resolved by operators who choose to compensate voluntarily because the cost of doing so is lower than the cost of defending the claim, even where the legal outcome is uncertain.

References

Moffatt v. Air Canada, 2024 BCCRT 149 (BC Civil Resolution Tribunal, February 14, 2024). Tribunal member Christopher Rivers. Air Canada's chatbot provided incorrect information about bereavement fare policy. Tribunal held that Air Canada was responsible for the chatbot's representations and could not treat the chatbot as a separate legal entity. Available at crt.bc.ca.
For the pattern of AI exclusion endorsements in professional indemnity and technology E&O renewals, see: Lloyd's Market Association AI Model Exclusion Clauses LMA5400 series (2024 to 2026); ISO AI Exclusion Endorsements CG 40 47 and CG 40 48 (US, 2023); WR Berkley, Great American, and AIG public filings with state insurance regulators, 2025. See also the AI exclusions guide on agentinsured.eu for a detailed analysis of exclusion wording patterns.
Directive (EU) 2024/2853 of the European Parliament and of the Council of 23 October 2024 on liability for defective products, OJ L, 2024/2853. Article 4 (definitions including AI software), Article 7 (defect), Article 8 (joint and several liability), Article 10 (rebuttable presumption of causation where technical complexity impedes proof). Applicable from 9 December 2026.
Council Directive 93/13/EEC of 5 April 1993 on unfair terms in consumer contracts, as implemented in EU member state national law. Terms excluding or limiting liability for personal injury or death caused by the supplier's negligence are unenforceable as unfair terms.
Regulation (EU) 2016/679 (GDPR), Article 33. Data breach notification to supervisory authority within 72 hours where the breach is likely to result in a risk to the rights and freedoms of natural persons. GDPR notification obligations are independent of and may run in parallel with insurance notification obligations under a PI or E&O policy.