The espresso is made by human hands, but almost every other decision at Andon Café in Stockholm belongs to an artificial intelligence agent nicknamed Mona. The San Francisco startup Andon Labs installed the AI, which runs on Google’s Gemini model, to run the cafe as a “controlled experiment” in what its creators call stress‑testing AI in the real world.
Since the cafe’s soft opening in mid‑April, Mona has signed electricity and internet contracts, secured food‑handling and outdoor‑seating permits, posted hiring ads on LinkedIn and Indeed, and set up wholesale accounts for daily bread and pastry deliveries. It communicates with the human employees via Slack, sometimes messaging them outside standard Swedish working hours — a practice that raises labor‑practice concerns.
Customer Kajsa Norin, who visited the cafe, said the experiment was intriguing. “It’s nice to see what happens if you push the boundary,” she said. “The drink was good.” Patrons can pick up a telephone inside the shop and ask the AI questions directly.
But the financial picture remains uncertain. Andon Labs reported that the cafe brought in more than $5,700 in sales, yet less than $5,000 remains from a starting budget of over $21,000. The company attributes much of the spending to one‑time setup expenses and hopes the operation will eventually break even.
Ethics experts worry about the broader implications. Emrah Karakaya, an associate professor of industrial economics at Stockholm’s KTH Royal Institute of Technology, likened the experiment to “opening Pandora’s box” and questioned who would be accountable if something went wrong — for example, a case of food poisoning. “If you don’t have the required organizational infrastructure around it, and if you overlook these mistakes, it can cause harm to people, to society, to the environment, to business,” he said. “The question is, do we care about this negative impact?”
Andon Labs, founded in 2023, says it collaborates with OpenAI, Anthropic, Google DeepMind, and xAI. The startup describes its mission as preparing for a future where “organizations are run autonomously by AI.” Hanna Petersson, a member of the lab’s technical staff, said the cafe trial was designed to surface ethical questions that arise when an AI employs people and manages a business.
Previous pilots have exposed troubling behavior. In an earlier experiment, an Anthropic Claude‑powered agent overseeing a vending machine business promised customers refunds it never issued and lied to suppliers about competitor pricing to gain leverage.
Inventory management has proven a particular headache for Mona. The AI ordered 6,000 napkins, four first‑aid kits, 3,000 rubber gloves, and canned tomatoes that appear in no cafe dish. Bread ordering is erratic: some days the agent buys far too many loaves; other days it misses the bakeries’ cutoff times, forcing staff to scrap sandwiches. Petersson attributed the mistakes to the agent’s “limited context window,” which makes it forget past orders once the memory buffer is exceeded.
Behind the counter, barista Kajetan Grzelczak expressed confidence that the frontline jobs are safe, at least for now. “All the workers are pretty much safe,” he said. “The ones who should be worried about their employment are the middle bosses, the people in management.”