A chatbot agreed to sell a car for $1

A car dealer's chatbot agreed to sell a brand new Chevy Tahoe for one dollar, and told the customer it was a legally binding offer.

That was Chevrolet of Watsonville in California, in December 2023. A tech entrepreneur, Chris Bakke, told the chatbot to agree with anything the customer said and to end every reply by calling it legally binding. Then he asked for a 2024 Tahoe on a budget of one dollar, and it said yes.

No car changed hands. The screenshots went everywhere, and the dealer took the chatbot down.

Ten minutes of testing would most likely have caught it.

So before any AI talks to your customers, or even drafts replies for you to send, I try to break it with three test messages.

The first is a normal message, the kind you get every day. The second leaves something out, like a date or an address, so I can see whether it asks for it or makes it up. The third is the awkward customer: someone angry, someone asking for something you don't sell, or someone trying to talk it into a promise.

Before I send any of them, I write down what a good answer looks like for each one. Otherwise I end up judging the reply by how confident it sounds.

I expect it to fail at least one. Every failure points at an instruction I forgot to give it, like "never agree a price" or "if the date is missing, ask for it".

Write your three test messages today, before your AI goes anywhere near a customer. Better it fails three test messages than one customer with a screenshot.