Does a disclaimer saying the chatbot may be inaccurate protect us?

It helps and it does not cure. A prominent, pre-conversation disclaimer supports an argument that reliance on a specific answer was unreasonable, which matters for misrepresentation claims. It is weaker where the chatbot's statement contradicted nothing else the customer could see, where the customer acted on it immediately, and where the disclaimer was buried in terms rather than presented in the conversation itself.

What if the chatbot quotes a price lower than our real price?

Whether that forms a contract depends on whether the quote was an offer capable of acceptance or an invitation to treat, and on whether the error was obvious enough that the customer could not reasonably have believed it. Businesses that route every quote through a checkout that recomputes price from the system of record convert this from a contract question into a customer-service one, which is the outcome you want.

Can we blame the chatbot vendor?

Not as against the customer. The customer's claim runs against the business whose brand and site the conversation appeared on. A vendor indemnity may let you recover afterwards, but most AI vendor agreements cap liability at fees paid and exclude consequential damages, so treat vendor recourse as partial rather than as a transfer of the risk.

Do we have to keep transcripts of AI chat conversations?

For practical defence, yes. Without the transcript you cannot show what the bot actually said, and the customer's screenshot becomes the only record in evidence. Retention also intersects with litigation-hold obligations and, where the conversation captures personal data, with privacy-law retention limits — so set a defined period rather than either deleting immediately or keeping everything forever.

Is Your AI Chatbot's Promise Binding? Contract Formation Risk 2026

Q: Is a promise made by a company's AI chatbot legally binding?

It can be. In Moffatt v. Air Canada, a Canadian tribunal held the airline responsible for a bereavement-fare policy its chatbot described but that did not exist, rejecting the argument that the chatbot was a separate entity responsible for its own statements. The general principle is that a business is answerable for representations made on a channel it operates and holds out to customers, whether the speaker is a human agent or software.

The Defence That Failed

The case everyone cites involved an airline chatbot that described a bereavement-fare refund process the airline did not in fact offer in that form. The passenger followed what the bot said, booked, and was refused. The airline argued, in substance, that the chatbot was a separate entity responsible for its own statements.

The tribunal rejected that outright. The chatbot was part of the airline's website; the airline was responsible for all the information on that website, whether it came from a static page or from a conversational component. That the information was generated rather than authored changed nothing about who was answerable for it.

The award was small. The principle is not, and it generalises past aviation: if you deploy the agent under your brand, on your surface, to your customers, its statements are your statements for the purposes of everything that follows.

Four Doctrines, Not One

"Is it binding" collapses four separate legal questions that have different answers and different defences. Separating them is what makes the risk tractable.

Contract Formation

Did the exchange create an offer and an acceptance with definite terms? A bot confirming a specific price, quantity and date on a transaction the customer then completes is the strongest case. A bot describing a policy in the abstract is weaker — that is usually a representation, not an offer.

Apparent Authority

Even where a human employee would have lacked actual authority to grant a concession, a business can be bound by what it holds out. A support agent presented as authoritative on refunds looks authoritative on refunds, and the customer does not see your internal approval matrix.

Negligent Misrepresentation

A false statement of fact, made where the business ought to have known it would be relied on, causing loss. This is the theory the airline case actually turned on, and it does not require a contract at all — only reliance and damage.

Consumer Protection / UDAP

A deceptive statement to a consumer can be actionable by a regulator or under a state statute regardless of contract analysis, and disclaimers are weakest here. Systematic misstatement across many conversations is what turns an individual complaint into an enforcement matter.

Note which of these a disclaimer touches. It bears on reasonable reliance, which matters for misrepresentation. It does very little against apparent authority, and less against a regulator applying a deception standard.

The Statements That Actually Cost Money

In practice the exposure clusters into a small set of answer types. Everything else a support bot says is low-stakes, which is why bounding these specific categories is more effective than trying to make the whole system accurate.

Price and discount quotes — the clearest path to a formation argument, especially where the customer transacts in the same session
Refund, cancellation and return eligibility — the airline fact pattern, and the most common real-world complaint
Coverage, eligibility and qualification statements — in insurance, lending, benefits and healthcare these carry sector-specific regulatory weight on top of the contract question
Delivery, availability and timeline commitments — a stated date the customer relies on to their detriment is a representation even where the sale terms say estimates only
Compliance, safety and legal assurances — a bot telling a customer a product is certified, compliant or safe makes a claim the business must be able to substantiate
Anything the bot invents to be helpful — a policy that does not exist, quoted confidently, is indistinguishable to a customer from one that does

Guardrails That Work, Ranked

1. Architectural — Don't Let the Model Be the Source

☐Serve price, availability, order status and entitlement from the system of record through a tool call, never from model knowledge or a stale prompt
☐Recompute every quoted figure at checkout from the same source, so a wrong number in chat cannot become a wrong number in a transaction
☐Restrict the bot to retrieving and quoting published policy text rather than paraphrasing it, for the answer types listed above

2. Escalation Boundaries

☐Define a refusal-and-handoff list: concessions, exceptions, legal or compliance questions, and anything involving an individual's eligibility
☐Make the bot decline in-scope-only language rather than improvising — an honest 'I can't confirm that, here's a human' is worth more than a fluent guess
☐Prohibit the bot from committing to a remedy; let it open a ticket instead, which converts a promise into a request

3. Disclosure and Disclaimer, Placed Where They Count

☐Disclose that the customer is talking to an AI agent at the start of the conversation — increasingly a legal requirement, not only a trust measure
☐Put the accuracy disclaimer in the conversation surface itself, not only in linked terms, and repeat it where the bot answers a high-stakes category
☐State plainly where the authoritative answer lives, so a customer who relies on the bot instead had an alternative in front of them

4. Evidence and Monitoring

☐Retain transcripts for a defined period — you cannot defend a conversation you did not keep, and the customer's screenshot will be the only record if you did not
☐Sample transcripts for the six high-stakes categories and count how often the bot answered rather than escalated; that rate is the real exposure metric
☐Route customer complaints about bot statements into one place so a systematic misstatement is visible as a pattern before a regulator sees it as one

Frequently Asked Questions

Our bot only answers from our help centre articles. Are we still exposed?

Less so, and the residual risk moves to a different place: stale articles and confident paraphrase. A retrieval-grounded bot that quotes an out-of-date policy is making an accurate report of a wrong source. Content freshness becomes a liability control once retrieval is your architecture.

If we honour the bot's mistake once, have we set a precedent?

Honouring a single low-value mistake is usually cheaper than litigating it and does not create a general obligation, but a documented pattern of honouring the same misstatement can support an argument about what customers were reasonably led to expect. Fix the underlying answer rather than repeatedly paying for it.

Does an arbitration clause in our terms cover chatbot disputes?

Often yes, if the clause is validly formed and broad enough to reach pre-contractual representations, which is a drafting question worth checking. It changes the forum rather than the merits, and it does nothing against a regulator.

We use an autonomous agent that can issue refunds directly. Different analysis?

Materially. Once the agent can execute rather than only describe, actual authority is delegated and the formation question largely disappears — the transaction happened. Controls shift to hard limits on value and frequency, a reversal path, and logging sufficient to reconstruct why the agent acted.

Is 'the AI made an error' ever a defence?

Not on its own. It concedes the statement was made on your surface and offers no reason the loss should fall on the customer rather than the deployer. Where businesses do better is on reasonable reliance — showing the correct information was prominent and the bot's answer was contradicted by what the customer could see.

Count the Answers It Should Not Have Given

Pull a week of transcripts and count how many times the bot answered a price, refund, eligibility or timeline question outright instead of retrieving the figure or handing off. That count is your exposure, expressed in the only unit that matters — statements already made to real customers.

Every one of those is a representation you would have to defend. Most of them can be removed with a tool call and a refusal rule, neither of which requires a better model.

Your Chatbot Made an Offer. A Tribunal Called It a Term.