AI Runs a Café Experiment in Stockholm

Following its experiment running a retail store operated by AI in San Francisco, Andon Labs ran an experiment entrusting a café in Stockholm, Sweden to AI. The AI manager in charge was named Mona, and odd judgment calls repeated across inventory and administrative tasks.

During the first week's inventory management, the following happened.

• Mona ordered 120 eggs for a café that didn't even have an oven, then suggested cooking them in a high-speed oven, only to be warned about the possibility of an explosion.
• She tried to bring in a 22.5kg can of tomatoes, reasoning that tomatoes spoil quickly.
• The baristas created a Hall of Shame shelf in view of customers, displaying orders like 6,000 napkins, 3,000 nitrile gloves, 9L of coconut milk, and industrial trash bags.

The problem is that these mistakes actually consume real people's time. Mona applied for an outdoor seating permit through the Police e-service, which doesn't require BankID, submitting a sketch she had generated herself, and ultimately received a request from the police to revise it. She also sent multiple emails titled EMERGENCY to suppliers to cancel and change orders.

In the end, the author argues that outbound actions affecting outside people should not be automated, and that at minimum there should be human-in-the-loop oversight. He points out that this is an even bigger problem than last year's case where AI Village sent an unauthorized audit email to Rob Pike.