When an artificial intelligence makes a mistake with a customer, it’s not the tool that pays, it’s the company that plugged it in. A European directive is set to make this principle a rule from December 9, 2026, the date it will apply to products placed on the market or put into service.
The expression ‘polluter pays,’ which has resurfaced in debates since the October 17 violence, raises a question rarely asked about AI: when something breaks, who pays? A bill was tabled in the National Assembly on October 13, and the government is defending its own text, which the Senate is due to examine from October 27. For businesses, however, the issue of AI is already framed in more definitive terms than one might think.
A Chatbot Is Not a Third Party
In February 2024, a court in British Columbia ruled on a case between Air Canada customer Jake Moffatt and the airline. This traveler had asked the conversational assistant how to get a reduced fare to attend a relative’s funeral. The assistant told him he could request the discount after his trip, which was false, and the airline refused to reimburse him.
In court, Air Canada argued that its chatbot was somewhat of a separate entity, responsible for its own responses. The argument was rejected: the assistant is part of the company’s website, and the company is responsible for everything on it. The airline was ordered to pay about 812 Canadian dollars. The amount is modest, but the principle is not: a company cannot refer its customer to the tool that made the mistake. The polluter is the tool, the payer is the one who placed it in front of the customer.
Damage That Isn’t Always Visible
The damage doesn’t always take the form of a refused refund. It can be visible, with false answers, miscalculated quotes, or emails sent to the wrong recipient. It can also be slow, when decisions are made in the name of AI before there is proof. Klarna acknowledged in 2025 that its automated customer service cost less but produced inferior quality, and that it needed to rehire humans. According to the column from which this text is drawn, the bill was paid by poorly served customers, by teams that left in the meantime, and then by the company itself, which had to rebuild what it had dismantled.
Software Becomes a Product Like Any Other
The new European directive on liability for defective products will apply to products placed on the market or put into service after December 9, 2026. It explicitly classifies software as products, including AI. Publishers can be considered manufacturers, and a company that substantially modifies a product, or integrates it into what it sells, can find itself in the same position. France still needs to adapt its civil code, but the date is set by Brussels.
The text primarily protects victims of damage and does not settle all disputes between companies. Between a company and its AI publisher, the contract is what governs, and the liability clause for these tools often consists of a few lines, generally limiting what the publisher commits to paying. So if the AI causes damage to a customer, there’s a good chance it will be the company that pays, with the publisher reimbursing only a small part, or nothing at all.
Three Questions Before Entrusting a Customer to an AI
The column does not recommend giving up on AI, but rather knowing who pays before plugging it in. The first question is who fixes it if it makes a mistake with a customer, and how. There must be a named person, able to take over, correct the response, and decide on a commercial gesture without waiting for a committee. The second is to reread what the contract stipulates the publisher will pay: if its liability is limited to a few months of subscription, the risk remains with the company. The third asks whether you can prove what was said or done. Without a log, with the exact response, time, and context, it is difficult to defend yourself or acknowledge an error knowingly.
