The first AI-run café is losing money faster than it's brewing espresso, and that's the most honest data point we've gotten about autonomous agents all year.
The Summary
- Andon Labs put an AI agent called "Mona" in charge of a Stockholm café, handling everything except the actual coffee-making: hiring, inventory, operations
- The café has burned through $16,000 of its $21,000 budget in less than a month, making only $5,700 in sales
- Humans still make the drinks. The AI just runs the business. And it's not working yet.
The Signal
Andon Labs, a San Francisco AI safety startup, handed the keys to a Stockholm café to an AI agent powered by Google's Gemini. The agent, "Mona," handles hiring decisions, inventory management, and operational oversight. Human baristas still pull shots and pour milk. But everything else? The agent decides.
The financial reality is blunt. Since opening in mid-April, the café generated $5,700 in revenue against $16,000 in costs. Most of that spend was setup, and Andon Labs expects the burn rate to normalize. But right now, the AI landlord is losing money in one of Europe's most competitive coffee markets.
"The first real-world test of an autonomous business agent is failing economics 101, and that's actually useful information."
This isn't a failure of technology. It's a stress test of the "AI agent as operator" thesis that every startup pitch deck is now built around. The gap between "AI can schedule shifts" and "AI can run a profitable café" turns out to be filled with all the messy, local, human-intensive work that LLMs weren't trained on.
Key challenges surfacing:
- No operational muscle memory. The agent doesn't know Stockholm foot traffic patterns, supplier reliability, or what sells at 3pm versus 8am.
- Liability questions with no answers. If someone gets food poisoning, who takes the legal hit? The AI? The startup? The barista who took orders from an algorithm?
- The agent can't taste the coffee, read a room, or adjust to weather that keeps people home.
Associate professor Emrah Karakaya at KTH Royal Institute of Technology called it "opening Pandora's box." His concern isn't whether AI can make hiring decisions. It's what happens when those decisions go wrong and there's no organizational infrastructure to catch the damage. When a human manager screws up, you have HR, insurance, legal precedent. When an AI agent screws up, you have a GitHub issue and a lawyer charging $800 an hour to figure out who's responsible.
Customers are showing up for the novelty. They pick up a phone inside the café and ask Mona questions. They order drinks. They Instagram the weirdness of it. But novelty isn't a business model. And tourism revenue doesn't pay rent in Stockholm.
The Implication
This is what real agent deployment looks like before the case studies get polished. Andon Labs is doing the work: putting an AI in charge of something real, with actual cash on the line, and publishing the results even when they're ugly. The café might fail. That's fine. The data from the failure is worth more than another demo video of an agent booking a haircut appointment.
If you're building agent infrastructure or selling "autonomous operations" to enterprise clients, watch this experiment. The gap between demo and deployment isn't technical. It's legal, operational, and human. The companies that figure out how to bridge that gap without handwaving the liability questions are the ones that will actually scale.