When the chatbot makes an offer and the dealership says it didn't count

The incident

A customer contacted a Toronto BMW dealership's website chatbot to ask about trading in their car. The chatbot offered a buyback price. The customer accepted. The dealership later withdrew the offer, saying the chatbot had generated it autonomously and the offer wasn't binding.^[1]

The incident was logged in the AI Incident Database as case 1772. Details are sparse, but the basic pattern is familiar: customer gets answer, dealership says it doesn't count, and the only record is whatever the customer happened to keep.^[1]

It's a bit like a bouncer at a club who waves you in, then the manager appears and says the bouncer wasn't authorised to let anyone through. Who's right depends on whether anyone wrote down what the bouncer said.

What Article 50 requires

Since 2 August 2026, Article 50 of the AI Act has required providers to design AI systems that interact with people so that people are informed they're talking to an AI, unless it's obvious from the circumstances.^[2] The rule applies whether the system gives advice, answers questions or takes instructions.

The disclosure has to happen, but the regulation doesn't say the AI must keep a record of what it said. That's a separate question, and often a harder one.

When a chatbot makes a commercial statement—a price, a promise, a warranty—somebody has to be able to show what was said. If the only record is the customer's screenshot, the customer has the evidence and the business has a problem.

The accountability gap

The Toronto case raises a question businesses running customer-facing AI will recognise: if the chatbot made an offer the company doesn't want to honour, can the company show what the chatbot actually said?

Often, the answer is no. Customer service systems log that a conversation happened, sometimes the topic, rarely the full exchange. The customer, meanwhile, has a screenshot.

In a dispute, the side without its own record starts on the back foot. A court or an ombudsman will want to see what was said. If one party has a screenshot and the other has nothing, the screenshot tends to win.

The dealership's claim—that the chatbot acted autonomously—doesn't make the problem go away. A business is responsible for the systems it puts on its website.^[3] Whether the chatbot was following instructions, hallucinating or improvising, it was still the dealership's chatbot talking to the dealership's customer on the dealership's site.

What this means for you

If you run a chatbot that talks to customers, three questions are worth asking now:

1. Can you show what your chatbot said to a specific person on a specific day?
2. If a customer turns up with a screenshot six months later, do you have your own record to compare it with?
3. Who in your organisation decides what counts as a binding statement, and does the chatbot know that?

The first question is about logging. The second is about whether those logs can be trusted. The third is about governance, and it's harder than it sounds: when a chatbot offers a price or a service, is that an invitation to treat, or is it an offer the company is bound by? If it's the latter, someone had better be able to show what the offer was.

Many businesses assume their chat platform logs everything. It's worth checking. Some platforms keep conversations for a few weeks, some don't keep the content at all, and almost none protect the logs against later editing by someone with administrator rights.

How a digital witness helps here

When a chatbot answers a customer, the thing to seal is the answer itself: what the chatbot said, when it said it and, ideally, the notice that the customer was talking to an AI.

Sealing takes a fingerprint of the answer, not the answer itself, so nothing leaves your systems. If the conversation is questioned later—by the customer, by an ombudsman or in court—anyone you share the record with can check that it's the one written at the time. Change a single character in the answer and the fingerprint won't match.

It doesn't settle whether the company is bound by what the chatbot said. That's a question of contract law and authority. What it does is give you the same quality of record the customer has, so the dispute isn't decided by whoever happened to take a screenshot.

You seal answers as they're sent, through WitniumChain, using the REST API, the SDK or the MCP server. The content stays with you. To check, you calculate the fingerprint again and compare it with the sealed one. The first 2,000 witnesses a month are free. Read the docs for how to start.

How it would work

1. Customer asks a question. Chatbot answers.
2. Your system seals the answer (question, response, timestamp and AI disclosure) before it's sent.
3. Six months later, the customer disputes what was said. You retrieve the sealed record, calculate the fingerprint and show it matches. Nobody has to take your word for it.

What to watch

Article 50's transparency requirements are live, but enforcement is still ramping up. The AI Office has published guidance on how to comply, and national authorities are watching how businesses implement the disclosure.^[2]

For high-risk AI systems, logging requirements under Article 12 come into force on 2 December 2027 for systems in Annex III.^[4] Those logs will need to be protected, and the question of how to prove a log hasn't been edited will move from theoretical to urgent.

The new Product Liability Directive treats software as a product from 9 December 2026, and courts can order defendants to hand over evidence.^[5] If your chatbot makes statements that could lead to liability—and most customer-facing ones do—being able to show what it said becomes part of your risk management.

Screenshots are evidence. Decide now who keeps yours.