Here's a deceptively simple question that AI travel platforms keep getting wrong: what does "non-refundable" mean? The customer thinks it means non-refundable. The terms say it means non-refundable. And then the chatbot, asked directly, decides it means whatever sounds most helpful in the moment. That's the trap a Priceline customer walked into, and it cost them a hotel stay, nearly six months of their life, and a small lesson in who actually pays when an AI assistant makes things up.

The booking, the bot, and the promise

The details come from a case taken up by Ruth Emery, The Telegraph's consumer champion, in a column published in May 2026 (the full text is also carried in a Yahoo News UK reprint). In August 2025, a reader identified as AB from north London booked three nights at the SIS Kata Resort in Thailand through Priceline, for early November, at a cost of GBP386.91. It was a non-refundable rate. AB knew that.

By late October, AB's plans had changed and they had another place to stay. On October 28, 2025, they tried to reach Priceline to ask about a partial refund and weigh their options. Crucially, there was no obvious way to contact Priceline except through its AI assistant, branded "Penny," which the company markets as a customer's "personal hotel expert, destination scout, booking support, and help desk - all in one."

AB asked Penny whether the booking could be cancelled. Penny's answer, quoted from the customer's screenshot, was that although the reservation was "originally listed as [a] non-refundable booking," they were "in fact eligible for a full refund" and that the money would land in their bank account within 10 working days. Surprised but reassured, and sensibly having it in writing, AB confirmed. Priceline emailed back a cancellation confirmation. So far, so good, if you ignore that the bot had just overruled the one rule the booking was built on.

The refund that wasn't, and the buck nobody held

The 10 working days came and went. No refund. When AB chased it, the AI's status response now said the refund was "being handled by the hotel." So AB called the hotel directly, and the story fell apart entirely: not only would there be no refund, since it violated the hotel's policy, but the hotel said it had never even been informed of any cancellation. They pointed AB to Agoda, Priceline's sister brand under the Booking Holdings umbrella, which they said had confirmed the booking on their end.

What followed was the part anyone who has fought a billing error will recognize with a wince. AB called Agoda, which first declined to investigate, then said that because the booking went through Priceline, it wasn't Agoda's problem. AB eventually unearthed a Priceline phone number "in the depths of the internet," only to be sent back to Agoda, which repeated its position. Around in circles for about a month, with no one inside this corporate family willing to own the mistake their own software had made.

AB summed up the trap precisely: "I wouldn't have cancelled the hotel if I hadn't been led to believe I was eligible for a refund." That's the whole case in one sentence. The harm wasn't an abstract bad answer; it was a real decision, made in reliance on a written promise from the company's only available support channel, that destroyed a booking the customer would otherwise have kept.

It took a journalist to make "misinformation" mean something

The runaround only ended when The Telegraph's consumer column got involved and pressed Priceline to look again. A couple of weeks later the company agreed to refund the money, first offering a cheque in the post, which is its own small comedy given Penny had specifically promised a bank deposit. After the column pushed for a bank transfer instead (and flagged the risk of a US-dollar cheque eaten by conversion), Priceline paid GBP363.11, a bit short of the original GBP386.91 after foreign-currency charges, nearly six months after the booking.

Tellingly, Priceline emailed AB acknowledging that the chatbot had given "misinformation," because the reservation was supposed to be non-refundable. That admission is the entire problem stated plainly. The bot didn't surface a real, if obscure, refund entitlement. It invented one, in writing, and the company only honored the invented promise once a national newspaper was watching.

The pattern, and why Priceline's version is its own beast

Regular readers will feel a sense of déjà vu, and they should. The Vibe Graveyard already houses the canonical version of this failure: Air Canada was held liable by a Canadian tribunal after its chatbot fabricated a bereavement-refund policy, with the tribunal flatly rejecting the idea that a company can disown its own bot's statements. There's the Chevrolet dealership bot manipulated into a one-dollar car, and a UK retailer's chatbot that conjured an 80 percent discount. The throughline is consistent: regulators and courts increasingly treat what a chatbot says as a statement by the business itself, which means "the AI got it wrong" is a confession, not a defense.

But the Priceline case has a couple of features that make it more than a reskin of Air Canada. First, the channel problem. AB found no human-reachable support at the point of need; the bot wasn't one option among several, it was the entire front door. When the only way to ask a question is an AI assistant prone to making things up, the company has effectively outsourced its legal exposure to a system it admits can produce "misinformation," and left customers with no human to sanity-check the answer before they act on it.

Second, the structural fog of a multi-brand booking stack. The booking touched Priceline, Agoda, and the hotel, all connected through Booking Holdings, and that very connectedness became the excuse for inaction; each party could plausibly gesture at another. The chatbot sat on top of this machinery confidently issuing promises it had no authority or mechanism to keep. It told AB the refund was "being handled by the hotel" while the hotel had never heard of the cancellation. That's not a hotel-policy dispute; that's an AI narrating a fictional back-office process to a paying customer.

What it actually costs

The dollar figures here are small, a few hundred pounds. The instructive damage is the design it exposes. Priceline, like its sister brands, has been pushing Penny as an increasingly capable assistant; one independent test (by Travel AI Playbook) describes Priceline adding Anthropic's Claude to Penny and pitching it as more "agentic," while noting the checkout still falls back to a conventional web flow. The more these assistants are positioned as authoritative, the more customers will reasonably treat their statements as binding, and the more often a confident fabrication turns into a real-world action with a real-world cost.

The lesson is the same one companies keep relearning and then re-forgetting. If you deploy a chatbot as your sales and support voice, the things it says are things you've said. "Non-refundable" can't quietly become refundable because the bot wanted to be agreeable, and a refund can't be real because the bot phrased it confidently. Either the assistant is constrained to tell the truth about your own policies, or you're going to keep paying out on promises a machine made in your name, eventually, grudgingly, once a reporter calls. AB got their money. Most people don't have a newspaper on speed dial.