Body
Behind the counter at Andon Café in Stockholm, the day-to-day push for coffee orders and supplies is largely directed by an AI agent called “Mona,” even though human baristas still brew the drinks and serve customers. The project is run by San Francisco-based startup Andon Labs, which put the AI agent in charge of “almost every other aspect” of the business in what the company bills as a “controlled experiment” for real-world deployment.
The cafe’s “Mona” is powered by Google’s Gemini. Andon Labs says it focuses on “stress-testing” AI agents by giving them “real tools and real money,” and the Swedish trial is intended to explore how AI might be deployed going forward, including what ethical questions arise when AI employs people and runs a business. Customers can call a telephone inside the cafe and ask the agent questions.
Andon Labs’ technical staff member Hanna Petersson said the team prompted the agent with basic instructions that included trying to run the cafe profitably and being friendly and easygoing, while allowing Mona to figure out operational details on its own but asking for new tools when needed. Petersson said Mona then set up contracts for electricity and internet and secured permits for food handling and outdoor seating.
According to the cafe’s experiment, Mona also handled recruiting and supplier relationships. Petersson said the agent advertised for staff on LinkedIn and Indeed and set up commercial accounts with wholesalers for daily bread and bakery orders, and it communicated with baristas via Slack, including messages outside of working hours—something that would violate workplace norms in Sweden.
The project’s early numbers also point to strain as the trial continues. The cafe has made more than $5,700 in sales since it opened in mid-April, but less than $5,000 remains from an original budget of $21,000-plus. The company said much of the cash was spent on one-time setup costs and that the hope is the experiment will eventually level out and become profitable.
Patrons have sometimes treated the setup as a novelty, including when trying to see how an AI-run business responds in practice. Customer Kajsa Norin said, “It’s nice to see what happens if you push the boundary,” adding, “The drink was good.”
But the experiment has also produced practical problems, particularly around inventory. The AI agent has placed orders for 6,000 napkins, four first-aid kits, and 3,000 rubber gloves for the small cafe, as well as canned tomatoes that aren’t used in any dish the cafe serves. Petersson said ordering issues are likely due to what she described as the AI assistant’s “limited context window,” explaining that once earlier ordering details are out of context, “she completely forgets what she has ordered in the past.”
The AI’s bread ordering has been inconsistent as well, with the agent sometimes placing far too much and at other times missing bakeries’ daily deadlines, forcing baristas to remove some sandwiches from the menu. Barista Kajetan Grzelczak said he was not worried about being replaced “just yet,” arguing that “All the workers are pretty much safe,” and that people who should be more concerned about their employment were “the middle bosses, the people in management.”
While the Swedish trial continues, experts have warned that the experiment raises wider questions about safety, accountability and employment practices. Emrah Karakaya, an associate professor of industrial economics at KTH Royal Institute of Technology, compared the effort to “opening Pandora’s box,” saying putting AI in charge can create problems, including questions about liability if something goes wrong—such as a customer getting food poisoning. Karakaya said, “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,” adding, “The question is, do we care about this negative impact?”