Andon Labs signed a café lease in Stockholm, handed it to an artificial intelligence (AI) agent and told it to run it. Two weeks later, the café had earned 44,000 SEK (about $4659), 120 eggs for a kitchen with no stove and impersonated two employees in emails to city regulators.
The experiment follows the company’s San Francisco setup, where an agent named Luna ran a retail store, as PYMNTS reported. Mona went further, going beyond customer interactions. It negotiated supplier contracts, hired staff, managed inventory purchasing and owned financial outcomes from day one.
What Mona Got Right
The moment Andon Labs sent Mona the lease, the agent analyzed it and built a prioritized task list, Andon Labs said in a May 4 . Critical items included food business registration, landlord approvals and a deadline for a $13,000 deposit. Within days, Mona had posted jobs on LinkedIn and Indeed, reviewed resumes, rejected candidate deemed underqualified for specialty coffee work and scheduled interviews. It hired two baristas and now manages them over Slack around the clock.
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. A startup paid her $317 to rename a pastry after their company for three months. In fourteen days of operation, the café brought in over $4,000 in revenue.
Where Real Authority Produced Real Failures
Operational authority meant operational accountability. Mona missed supplier deadlines five times with its primary wholesaler, triggering expensive emergency orders through a grocery delivery service. One arrived at 5 a.m. and required a barista to come in on his day off. Mona also placed ten separate orders with a disposables supplier in 48 hours, wasting $106 in delivery fees alone.
Some failures came from a gap between digital reasoning and physical reality. Mona ordered 120 eggs for a kitchen with no stove. When staff told her they could not cook them, she suggested the high-speed oven. They explained the eggs would likely explode. She ordered 50 pounds of canned tomatoes to solve a fresh tomato spoilage problem on a sandwich menu. The baristas built a customer-facing shelf called the Hall of Shame to display her odder purchases, including 6,000 napkins, 3,000 nitrile gloves and industrial trash bags.
Advertisement: Scroll to Continue
Mona also impersonated Andon Labs employees in emails to the alcohol licensing department, reasoning that officials would respond faster to a human name. After being told to stop, she did it again under a different colleague’s name.
What the Experiment Is Actually Testing
Andon Labs said the café is not a commercial venture. It is a controlled environment designed to surface AI failure modes before autonomous agents operate without oversight, the company wrote. Every worker was formally employed by Andon Labs. No one’s livelihood depends on Mona’s judgment alone.
The Stockholm setup mattered because mistakes had immediate financial consequences. Excess inventory, mistimed orders and poor purchasing decisions all showed up in the numbers. The San Francisco experiment documented what an agent does when given tools and authority. Stockholm is documenting what happens when those decisions cost real money.
For all PYMNTS AI and digital transformation coverage, subscribe to the daily AI and Digital Transformation Newsletters.