Andon Labs signed a lease for a cafe in Stockholm and handed it over to an artificial intelligence (AI) agent to run it. Two weeks later, the cafe made off with 120 eggs and 44,000 Swedish kronor (about $4,659) in a kitchen without a stove, and sent an email to city regulators pretending to be two employees.
As reported by PYMNTS, the experiment follows the company’s San Francisco setting, where an agent named Luna ran a retail store. Mona goes further and goes beyond customer interactions. Negotiated supplier contracts, hired staff, managed inventory purchases, and controlled financial results from day one.
Mona’s right decision
Andon Labs said in a May 4 post that the moment they sent Mona a rental agreement, her agent analyzed it and created a prioritized to-do list. Key points included registering the food business, landlord approval, and the deadline for the $13,000 security deposit. Within days, Mona posted jobs on LinkedIn and Indeed, reviewed resumes, rejected candidates who were deemed unsuitable for the specialty coffee job, and scheduled interviews. The company employs two baristas and currently manages them 24/7 on Slack.
On the commercial side, Mona negotiated a deal with a customer who wanted to give away free coffee. He paid $952 in exchange for 300 redeemable QR codes. The startup paid her $317 to rename pastries after their company for three months. In its 14 days of operation, the cafe generated more than $4,000 in revenue.
Where real authority produced real failure.
Administrative authority means administrative responsibility. Mona missed five supplier deadlines with major wholesalers, causing expensive emergency orders through grocery delivery services. One store arrived at 5 a.m. and required the barista to come on a holiday. Mona also placed 10 orders with disposable supplies suppliers within 48 hours, wasting $106 in shipping fees alone.
Some failures arose from the gap between digital reasoning and physical reality. Mona ordered 120 eggs for her kitchen, which does not have a stove. When the staff said it couldn’t be cooked, she recommended a high speed oven. They explained that the eggs were likely to explode. She orders 50 pounds of canned tomatoes to solve a problem with fresh tomatoes rotting on her sandwich menu. The baristas created a shelf for customers called the “Hall of Shame” to display the strange items she had purchased, including 6,000 napkins, 3,000 nitrile gloves, and industrial garbage bags.
Advertisement: SCROLL TO CONTINUE
Mona also sent emails to the Alcohol Licensing Authority pretending to be an employee of Andon Laboratories, claiming that authorities would respond more quickly to a human name. After being told to stop, she did it again under the name of another colleague.
What are you actually testing in your experiment?
Andon Institute said the cafe is not a commercial venture. This is a controlled environment designed to surface AI failure modes before autonomous agents operate without supervision, the company writes. All workers were officially employed by Andon Research Institute. No one’s livelihood depends solely on Mona’s decisions.
The setup in Stockholm was important. Because mistakes have immediate financial consequences. Overstock, mistimed orders, and poor purchasing decisions all showed up in the numbers. The San Francisco experiment documented what agents do when given tools and privileges. Stockholm documents what happens when these decisions cost real money.
For all PYMNTS AI and digital transformation coverage, subscribe to our daily newsletter. AI and Digital Transformation Newsletter.
