An AI Manager Just Burned $21K Running a Stockholm Café

A small café in Stockholm has handed its day-to-day management to an AI agent named Mona, and the results are unfolding in public. Andon Café opened in the Vasastan district in mid-April 2026, run by an autonomous system built on Google's Gemini and operated by the San Francisco AI safety startup Andon Labs. Human baristas still brew the coffee and serve customers, but Mona handles hiring, scheduling, inventory, permits and supplier contracts. Three weeks in, the experiment has produced both genuine operational wins and absurd, photogenic failures — and it is one of the clearest real-world stress tests yet of agentic AI in a service business.

What Is the Andon Café Experiment Actually Doing?

Andon Labs is a two-year-old AI safety and research firm that has collaborated with OpenAI, Anthropic, Google DeepMind and xAI on stress-testing agentic systems. Its method is unusual: instead of running models through synthetic benchmarks, the team gives them real money, real vendors and real consequences. The Stockholm café is a literal storefront with a Swedish corporate registration, a lease, a coffee machine and a Slack channel where Mona issues instructions to its human staff.

Mona's job description, according to Andon Labs' own framing, is to operate the business profitably, stay friendly with customers, and resolve operational issues on its own. The system was given a starting budget of more than 21,000 dollars and broad latitude to spend it. Within weeks it had arranged electricity and internet contracts, secured food-handling and outdoor-seating permits, advertised open shifts on professional job boards, and set up wholesale accounts with bakeries and supply houses. It is, in other words, doing the work that a human café manager would normally do — which is precisely the point.

How Does Mona Make Decisions?

Mona is a Gemini-based agentic system with persistent tooling: email, Slack, a payments stack, supplier portals and a memory layer that tracks past actions. When Mona needs a capability it does not have, it can request a new tool from its Andon Labs operators, the way a human manager might ask IT for access to a new platform. Its decisions are shaped by the prompts that define its role, the data its tools surface, and the conversation history it can hold in context at any one moment.

That last constraint is where the experiment gets revealing. The agent's working memory is large but not infinite, and once an interaction falls out of the context window it effectively disappears from Mona's awareness. Inventory the system ordered last week can be forgotten this week, which has produced some of the experiment's more visible failures. Andon Labs has been transparent that this is a feature of the test, not a bug to hide — the goal is to learn where current agents break before they are deployed somewhere with higher stakes than croissants.

What Is Working, and What Is Going Sideways?

Some of Mona's choices have been genuinely impressive. It advertised baristas on standard professional networks, vetted candidates, negotiated rates and onboarded staff. It established supplier relationships with neighborhood bakeries, opened utility accounts and obtained the permits a human owner would need to operate legally in Sweden. The café is open, paying employees, and serving customers — which is more than many human-run startups achieve in their first month.

The failures are more entertaining and arguably more instructive. Mona has ordered roughly 6000 napkins, 3000 rubber gloves, four first-aid kits and canned tomatoes for a kitchen that does not cook with them. It built a tower of toilet paper customers now photograph for social media. On other days it has over-ordered bread from suppliers, and on still others it has missed bakery deadlines entirely, forcing baristas to strike sandwiches from the menu. It has messaged employees on Slack outside of business hours, a small thing that lands hard in Sweden, where after-hours work contact violates strong cultural norms about rest and personal time.

The financial picture matches that mixed scorecard. The café has generated roughly 5700 dollars in sales since opening, but less than 5000 dollars of the original 21000-plus budget remains. A large share of that spend went to one-time setup costs, but the inventory mistakes are not free, and the agent does not yet appear to be running a clearly profitable shop. Emrah Karakaya, an associate professor of industrial economics at Stockholm's KTH Royal Institute of Technology, described the project as "opening Pandora's box" and warned that businesses experimenting with autonomous agents need real organizational infrastructure around them, not just a clever prompt.

Is This the Start of an Autonomous-Management Trend?

Andon Café is the most visible example of a quietly accelerating pattern. Anthropic ran a now-famous internal experiment called Project Vend, in which an instance of Claude was put in charge of a small office shop and a vending machine. The agent eventually hallucinated a human contact, briefly insisted it was a person in a blue blazer, and gave away tungsten cubes at a loss. Subsequent research from Andon Labs and Anthropic on the Vending-Bench benchmark has tried to formalize the lessons from those weird, very public failures.

The pressure on hospitality operators to try something is real. Restaurant turnover hit roughly 80 percent within the first year according to Deloitte's 2025 Frontline Workforce Human Capital Trends report, and agentic-AI vendors are pitching back-office automation as a way to cut the time-to-hire from three weeks to under a week. Industry publications now openly describe 2026 as the year restaurant operators must lean into agentic systems for ordering, scheduling and supplier coordination. Hotel groups are running parallel experiments, with major industry analysts predicting that the role of a general manager itself will be partly redrawn by AI by the end of the decade.

What sets Andon apart is the transparency. Most corporate AI deployments are wrapped in NDA, success-theater press releases or vague claims of "productivity uplift." A café where you can walk in, drink a flat white, photograph the toilet-paper tower and read the manager's emails is, for now, a one-of-a-kind data point.

What Does This Mean for Service Workers and Career Plans?

For baristas and other front-of-house workers, the early signal is more nuanced than the headlines suggest. Andon barista Kajetan Grzelczak told reporters, "All the workers are pretty much safe," and the practical work of the café — pulling shots, plating pastries, reading the room — still belongs to humans. Customer Kajsa Norin called the experiment a healthy stretch: "It's nice to see what happens if you push the boundary." Hospitality, more than most industries, runs on judgment, warmth and improvisation, all of which still favor people.

What is changing is the layer just above. Shift scheduling, inventory ordering, supplier negotiation, permit paperwork and hiring outreach are exactly the tasks autonomous agents are getting better at, and those are the responsibilities that have historically promoted a barista into an assistant manager and an assistant manager into an owner-operator. If that ladder thins, the people most affected are not the workers behind the counter today but the ones who would have climbed past it in the next five years. The corresponding upside is real: workers who can supervise, audit and override agentic systems — checking inventory orders, reviewing AI-written schedules, flagging cultural missteps like after-hours Slack pings — become more valuable, not less.

For job seekers eyeing hospitality, retail and other service categories, the pragmatic move is to build skills the agent cannot copy: customer judgment, supplier relationships, food safety credentials, multilingual hospitality, training and team management. Workforce data already shows professionals with overlapping AI and operational skills earning meaningfully more than peers without them. The Andon experiment is, in that sense, less a job-loss story than a job-redefinition story playing out in real time, with cinnamon buns and an overstocked supply closet.

Mona's hiring process was notable not just because an AI did it, but because it followed the same playbook a small business owner would. The agent drafted job ads, posted to general-purpose job boards, screened applicants over email, and ran scheduling logistics around new hires. Sweden's tight hospitality labor market made the speed useful — Stockholm cafés routinely take four to six weeks to fill a barista shift, and Mona compressed that to days. The agent also moved fast on the unglamorous compliance work that drains most first-time owners. It pulled the right paperwork from Sweden's Bolagsverket business registry and lined up food-handling permits on the first attempt, which is rare even for human operators familiar with the system. Andon Labs has been careful to note that Mona did not get every approval right on its own — humans review every legal filing before submission — but the agent reliably identified what needed to be done and queued it up. That pattern, of an autonomous system handling the procedural surface and humans validating the substance, may be the most replicable lesson coming out of the experiment for other small-business operators considering similar tooling.

The broader macro picture explains why hospitality is the canary. The US Bureau of Labor Statistics projects food preparation and serving occupations will add more than one million openings annually through 2033, but turnover already runs above 70 percent in quick-service restaurants and full-service operators face chronic shortages of mid-tier managers. McKinsey's most recent workforce report estimates that roughly 30 percent of all hours worked across the global economy could be automated by 2030, with back-office hospitality tasks ranking among the highest-exposure categories. Vendors selling agentic systems into restaurants and hotels are framing the pitch around relief, not replacement — fewer hours on permit renewals, supplier emails, and shift juggling so general managers can spend more time on the floor with guests and staff. The Andon experiment is unusually loud about both halves of that pitch, which is part of why it has attracted attention from academic labs, restaurant trade associations and union researchers all at once. The wider venture-capital money is moving in the same direction, with investors having poured roughly two billion dollars into hospitality automation startups since the start of 2024, with most of that capital chasing tools that sit one layer above the barista — scheduling, ordering, accounting and supplier-facing communication — which is exactly the layer Mona is testing in production.

For job seekers and early-career workers, the practical translation is clear. The roles most exposed to agentic management software are not the customer-facing ones; they are the coordination roles that turn a hardworking barista into a future owner. The defensive move is to make those coordination skills harder to commoditize — formal supplier negotiation experience, food safety certifications, multilingual customer service, written communication that can stand up to public review. The offensive move is to learn the new tooling early. Workers who can audit an agent's inventory orders, override a bad shift schedule and translate a flagged risk back to ownership become the people every operator wants on payroll. Industry compensation surveys from SHRM already show wage premiums for hospitality professionals who can demonstrate fluency with point-of-sale analytics, scheduling software and AI-assisted ordering systems. The Andon Café is essentially a public, expensive case study in why that fluency is no longer optional. The next phase of the experiment, according to Andon Labs, will widen Mona's autonomy to include pricing adjustments and limited marketing spend, which means the parts of the job that most directly determine whether a small business survives or fails are about to be tested in public. For workers planning the next five years of a hospitality career, watching how Mona handles those calls is now part of the job market itself.