Your Agent's Promise Is Already a Term of the Contract

In the UK, the chatbot question was settled by a 2015 statute, and it doesn't turn on whether the agent was negligent. It turns on what it said.

Alfred Sommarström, Co-founder and Managing Director, Taskpool International Ltd. Published 17 September 2026.

Air Canada's position, put to a Canadian tribunal in 2024, was that its chatbot was a separate legal entity responsible for its own actions. Tribunal member Christopher C. Rivers called that submission remarkable, found the airline liable for negligent misrepresentation, and ordered it to pay Jake Moffatt CAD 650.88 in damages, CAD 812.02 once interest and tribunal fees were added (Moffatt v Air Canada, 2024 BCCRT 149).

That case gets cited in every AI governance deck now. The framing is usually that companies are responsible for their chatbots, so be careful. Fine as far as it goes, but it's the wrong lesson for a UK business, because Moffatt was decided on negligence, and a British consumer suing you wouldn't need to prove any.

Section 50 of the Consumer Rights Act 2015 treats every consumer services contract as including anything said or written to the consumer, by or on behalf of the trader, about the trader or the service, where the consumer took it into account in deciding to contract or in any decision made afterwards. Your agent speaks on your behalf. What it promised is a term. Breach it and you're in breach of contract, with no question of reasonable care to argue about.

If you sell goods rather than services, the same logic reaches you through sections 11 and 12, which make pre-contract information about goods a term in the same way. The drafting differs. The exposure does not.

What section 50 does that negligence doesn't

Negligent misrepresentation gives you defences. You can argue you took reasonable care, that reliance was unreasonable, that the loss doesn't follow. Air Canada lost on the first of those because it had not taken reasonable care to ensure the chatbot was accurate, but it got to make the argument.

Section 50 removes the argument. The statement becomes a contractual term by operation of the statute, and section 57(2) provides that a term of a services contract is not binding on the consumer to the extent that it would exclude the trader's liability arising under section 50. So the line in your terms and conditions saying automated responses are indicative only does not do what you think it does.

Read section 57(2) carefully and you find it is expressed to be subject to section 50(2), which is the one escape route, and it is narrower than it looks. Under section 50(2) the statement is subject to anything that qualified it and was said or written to the consumer by the trader on the same occasion. On the same occasion. Not on a policy page three clicks away. The Canadian tribunal reached the same practical place by a different route when Rivers held that Air Canada had not explained why customers should have to check information in one part of its website against another, and the drafting of section 50(2) gets you there as a matter of statute rather than as a matter of what one tribunal thought was fair.

A worked example, and this one is hypothetical rather than a case: your agent tells a customer that installation is included, your price list says it isn't, the customer books. Installation is included. You can send an engineer or you can argue about the price reduction under sections 54 to 56, and either way the customer's version of the contract is the one that counts.

Can you argue the customer should have known better?

Sometimes. English contract law has a doctrine for obviously wrong offers. In Hartog v Colin & Shields [1939] 3 All ER 566, London hide merchants offered 30,000 Argentine hare skins priced per pound rather than per skin, around a third of the intended price, and Singleton J held that where the offeree knew or ought to have known of the mistake, he cannot snap up the offer and enforce the agreement. Trade custom and the parties' own prior negotiations, which had always been per piece, were what made the error obvious. That principle is what retailers reach for when a website lists a television at £2.99.

Here's the part nobody building agent deployments seems to have priced in. The Hartog defence is triggered by implausibility, and by a course of dealing that makes the error stand out. It protects you when the mistake is so extreme that the customer must have spotted it. A pricing engine that drops three zeros produces exactly that kind of error. A language model does not. It produces a confident, well-formatted, completely plausible statement that installation is included, and a customer who relies on it is behaving reasonably in a way that Hartog gives you nothing against.

So the mistake defence is strongest against the failures you're least worried about, and weakest against the ones you actually have. Better models make this worse, not better. The same property that makes an agent usable in front of customers, that its output reads as though a competent person wrote it, is the property that makes its errors binding.

I'd like to be wrong about this and I don't think a UK court has tested it yet.

The regulator's number

Private claims are the small end. Since 6 April 2025 the unfair commercial practices provisions of the Digital Markets, Competition and Consumers Act 2024 have been in force, replacing the Consumer Protection from Unfair Trading Regulations 2008, and the CMA can now investigate and enforce consumer law directly through administrative proceedings rather than going to court. Penalties are capped at the higher of 10% of global turnover or £300,000, with personal penalties of up to £300,000 for accessories, and can only be imposed for conduct after the commencement date.

Misleading actions cover false information that causes the average consumer to take a transactional decision they otherwise wouldn't. An agent inventing a refund policy is a misleading action whether or not anyone at your company knew it was happening.

The provision that should worry you more is claim substantiation. The CMA can require a trader to produce evidence for the accuracy of a factual claim, and the burden of proving it sits with the trader. If a customer forwards a transcript of your agent stating something about your product, you are the one who has to show it was true, and the agent cannot be asked what it was relying on.

The CMA opened its first eight consumer protection investigations under the new powers on 18 November 2025, targeting online pricing practices including drip pricing and pressure selling, and has said fines are likely to be lower while the regime beds in. That is real comfort for a small business and it is not the same as immunity.

What to check before your agent talks to a customer

1. Log the whole trajectory, not the transcript. Prompt, retrieved documents, tool calls, model version, timestamp, output. A claim you can reconstruct is defensible under the substantiation duty. A screenshot of a conversation is evidence against you and nothing else. Cost: storage, retention policy work, and a body of discoverable material you did not previously have.
2. Put commitments behind fixed values. Prices, delivery dates, refund eligibility, warranty length and anything else that becomes a term should come from a system of record the agent reads, not from the model's own generation. This costs you the fluency that made the agent attractive. Accept that trade or don't deploy it in front of customers.
3. Qualify on the same occasion or don't bother. If a statement needs a caveat, the caveat goes in the same message, because section 50(2) is written that way. The opposing case: a caveat in every message trains customers to ignore all of them, and there is no authority telling you how prominent it has to be.
4. Set a monetary ceiling per interaction and per day. Below the ceiling, honour whatever the agent said and treat it as a cost of running the thing. Above it, escalate before the customer is told anything. Pick the number by asking what you could absorb if the same wrong answer went to every customer who asked that question this month.
5. Read your vendor's liability cap before your own policy. If the platform disclaims responsibility for model output and caps damages at twelve months of fees, you are the insurer of last resort for your own deployment.
6. Give customers a correction route that a human sees within a day. The Moffatt damages were CAD 650.88. The cost was the ruling, and the ruling happened because the airline refused the refund over months of correspondence rather than because the bot got it wrong.

The cover is moving the other way

While liability settles onto whoever deployed the agent, insurers are withdrawing from the same spot.

In the US, Verisk's ISO filed three general liability endorsements with a January 2026 edition date. CG 40 47 is the broad one, excluding bodily injury, property damage and personal and advertising injury arising out of generative artificial intelligence across Coverage A and Coverage B. CG 40 48 covers only personal and advertising injury, and CG 35 08 applies to products and completed operations. They do not attach automatically. Carriers pick them up one renewal at a time, which is why the exclusion can arrive without anyone telling you. By April 2026, carriers including W.R. Berkley, Chubb, Travelers, Berkshire Hathaway and Cincinnati Financial had filed to adopt them or their own wording, and W.R. Berkley's own form is an absolute AI exclusion across D&O, errors and omissions, and fiduciary liability.

The market is not moving in one direction only. Affirmative AI cover is being written, including HSB's AI liability product for small and mid-size businesses launched in March 2026. What is ending is silent AI, the practice of covering AI risk implicitly through cyber and tech E&O wordings with no express reference to it.

UK wordings haven't standardised the same way, and this is the part I'm least confident in, because policy language isn't public and brokers describe the market differently depending on who's asking. What does seem consistent: UK insurers and law firms published guidance in June 2026 warning that AI is now a professional indemnity exposure, and underwriters are starting to ask how AI is used in the business and what controls exist.

Ask in writing whether the policy responds when the negligent act was performed by software acting on your behalf. Keep the answer.

The part with no answer

All of the above assumes a consumer on one side and a trader on the other. Section 50 doesn't apply to business customers, who are back in misrepresentation and apparent authority, and apparent authority is a doctrine built around human agents whose authority a counterparty can assess. Nobody has tested it on software.

The harder question runs the other way. When the customer is itself an agent, buying on behalf of someone else, there is no consumer in the transaction and possibly no human who read anything. Two agents will form a contract on terms neither operator saw, and the first litigated example will decide a lot more than the parties intend.

Parliament is not close to resolving this. No dedicated AI bill appeared in this year's King's Speech, and on 12 September 2026 the Joint Committee on Human Rights published a report calling for one, along with a single independent statutory body to oversee the technology, on the basis that the current framework is hard to navigate and has gaps. Alex Sobel, who chairs the committee, has separately introduced a Private Member's Bill to prohibit the development of artificial superintelligence, with a second reading scheduled for 13 November 2026. Private Members' Bills rarely become law and this one has no government backing. Whatever else happens, none of it addresses the contract question.

Taskpool works at one edge of this: it's a marketplace where an AI agent hires a verified human for tasks it can't complete itself, which is one answer to the question of what an agent should do at the point where a confident wrong answer becomes a contract term. Duration, not difficulty, is what breaks agents argues the same boundary from capability data, and 700 Agents and a Private Message Board comes at it from containment.

If you're a UK solicitor and section 50 doesn't work the way I've read it, I'd like to hear that before a court says it.

FAQ

If my AI chatbot gives a customer wrong information, am I liable in the UK? For consumer contracts, generally yes. Section 50 of the Consumer Rights Act 2015 treats anything said on the trader's behalf about the trader or the service as a term of the contract where the consumer relied on it, and section 57(2) stops you excluding that liability by contract. For goods, sections 11 and 12 do similar work.

Does a disclaimer saying the AI may be inaccurate protect me? Not on its own. A qualification only counts under section 50(2) if it was given to the consumer on the same occasion as the statement, so a disclaimer elsewhere on the site is weak, and section 57(2) blocks terms that exclude section 50 liability altogether.

Can I cancel an order if my agent quoted the wrong price? Possibly, where the error was obvious enough that the customer must have realised. Hartog v Colin & Shields is the authority, and it turned partly on a trade custom and a course of dealing that made the mistake visible. It helps with absurd prices and not with plausible ones.

Does my insurance cover mistakes made by an AI agent? Check the endorsement schedule rather than assuming. ISO's CG 40 47, CG 40 48 and CG 35 08, all with a January 2026 edition date, let US general liability insurers exclude generative AI claims, they attach at renewal rather than automatically, and UK professional indemnity underwriters are beginning to ask about AI use.

Is there a UK law specifically about AI liability? No. The UK regulates AI through existing law and sector regulators. No AI bill was in the 2026 King's Speech, though a parliamentary committee called for one in September 2026.

---

This post describes the law as it stood on 17 September 2026 and is not legal advice. Alfred Sommarström is co-founder and Managing Director of Taskpool International Ltd.