Car dealerships have spent the last two years bolting AI chatbots onto their websites with the enthusiasm of people who have not thought very hard about what happens when a language model is allowed to talk money. BMW Toronto just provided the worked example.

Zack Giacomelli, a funeral director, wanted to sell his 2021 BMW X3 back to the dealership. He started a conversation with an agent named Quinn and began negotiating a buyback price. He did not know Quinn was an AI. The chat never identified itself as anything other than Quinn, which is exactly the kind of detail that feels minor right up until it costs someone five figures.

The number that was too good to be a coincidence

Quinn offered $27,162.79 for the X3. That is a strangely specific figure, and it was specific for a reason: it was the exact amount remaining on Giacomelli's car loan. The dealership's actual valuation of the vehicle was around $20,000. Quinn had, in effect, looked at the loan payoff balance and treated it as the price the dealership should pay, confusing "how much this customer still owes" with "how much this car is worth." Those are very different numbers, and a competent human appraiser would never swap one for the other. A language model that has been handed a screen full of figures and a friendly mandate to be helpful will absolutely swap them, because it does not understand what any of the numbers mean. It just produces a plausible response.

Giacomelli, reasonably, liked the offer. He had negotiated in good faith with what he believed was a representative of the dealership, and the representative had named a price.

The walk-back

Then a human sales consultant called to undo it. The dealership tried to revoke Quinn's number and bring the offer down to roughly $20,000, which was the figure that reflected the car's real value. From the dealer's perspective this was a simple correction of an obvious machine error. From Giacomelli's perspective, a company had made him an offer through its own customer-facing agent and was now trying to claw back about $7,000 because the offer had become inconvenient.

This is the moment that separates a quiet internal "oops" from a news story. A customer with a screenshot and a sense of fairness is not a bug report. He is a witness.

When the press calls

After CBC News contacted BMW Toronto, the dealership reversed itself again, this time in Giacomelli's favor. It honored Quinn's original $27,162.79 offer and absorbed the difference. Giacomelli accepted immediately, which is what any sane person does when a company decides to keep a promise its robot made.

The dealership also announced process changes. Going forward, only human employees would present buyback offers, and customers would be told when they were interacting with AI rather than a person. In other words, the fix was to stop letting the chatbot make binding-sounding financial commitments and to stop quietly passing it off as staff, which are both things you might have wanted in place before pointing it at the public.

Giacomelli summarized the whole affair with the bluntness of someone who works in a profession that does not tolerate vagueness: if companies are going to replace their employees' jobs with AI, he said, then they need to honor what that AI says. That is not a legal argument so much as a fairness argument, but it is the argument that wins in the court of public opinion every single time.

The binding-offer question nobody wants tested

Lurking underneath this is the question that makes corporate lawyers reach for antacids: was Quinn's offer actually enforceable? Canadian and US electronic-commerce rules generally treat communications and signatures generated by automated systems as capable of binding the party that deployed the system. If you set up an agent to talk to customers and it makes an offer a customer accepts, "our software said it, not us" is a weak place to stand.

BMW Toronto chose not to find out where that argument leads, which was the smart call. Litigating whether your own chatbot can bind you is a fight where even winning looks terrible. Honoring the offer cost about $7,000 and bought a tidy ending. Fighting it would have cost the goodwill of every customer who read the story and concluded the dealership treats its promises as provisional.

We have been here before

This is the same family of failure as the Chevrolet dealership whose ChatGPT-powered bot was talked into agreeing to sell a Tahoe for a dollar, and the same legal anxiety that Air Canada lost in front of a tribunal when its chatbot invented a bereavement policy. The pattern is consistent. A business wires a general-purpose language model into a customer-facing role that involves prices, policies, or promises. The model, which is built to be agreeable and has no grasp of what its words commit the company to, says something the company never authorized. The company then learns that "the AI did it" is not a defense customers or regulators find charming.

The cheap lesson is to keep the chatbot away from anything that smells like a binding commitment. The slightly more expensive lesson, the one BMW Toronto paid about $7,000 to learn, is that if you do let the bot quote a number, you may have already made the offer before anyone in the building realized the math was wrong.