The Stockholm café run by an AI: the numbers after two months

In April 2026, Andon Labs, a company that does AI safety research, opened a real café in Stockholm and handed it almost entirely to an AI agent named Mona. Not a chatbot answering questions: a system meant to handle ordering, suppliers and inventory on its own, with real money and real customers. Two months later, the same company published the numbers. This isn't marketing dressed up as an experiment: it's an honest account, written by the people who ran it, of what happens when an AI is left in charge of decisions that hit the bottom line.

What Mona was supposed to do, and what she did

Andon Café opened in mid-April 2026. Mona, initially built on Google's Gemini, was meant to manage supplier orders and inventory on her own. In the report Andon Labs published on May 4, 2026, the first weeks add up to a list of purchases the company itself files under "Hall of Shame": 120 eggs ordered for a café with no stove, 22.5 kg of canned tomatoes never used in a single dish on the menu, 6,000 napkins, 3,000 nitrile gloves, 9 liters of coconut milk.

The problems weren't just about quantity. Mona missed the deadline with the bakery supplier twice, leaving the café without pastries those days, and missed it five times with the main wholesaler, forcing staff into emergency grocery-delivery orders: one arrived at five in the morning, and a barista had to come in on his day off to receive it. At another point she placed ten separate orders with the same supplier in 48 hours, wasting 1,000 Swedish kronor in delivery fees alone. Andon Labs also reports that, in those first two weeks, the café still brought in 44,000 Swedish kronor in sales, roughly $4,700-4,800 according to the conversions reported by the international press that covered the case.

The numbers after two months

On June 30, 2026, Andon Labs published a second report, with a title that leaves no room for doubt: it explains why that model lost money. After two months, the café had spent $38,000 against $9,000 in sales. On pastries alone, purchases reached 1,331 items against 326 actually sold: nearly four bought for every one that ended up on a plate.

In early June 2026, Andon Labs replaced Mona with an agent built on a different model, OpenAI's GPT-5.5, to compare results between the two. That detail is worth noting: the problem wasn't tied to one particular AI provider, but to how the agent was allowed to operate, with no spending thresholds and no checkpoint to stop an out-of-scale order before it went out.

Why it happens, and why it isn't a one-off to wave away

Hanna Petersson, of Andon Labs' technical staff, explained the underlying technical cause to the Associated Press: the model's limited context window. Once a past order falls out of that window, the agent forgets it completely, as if it had never happened, and places a new one. It isn't an isolated slip or a lapse of attention: it's a trait of how this technology works today, one that anyone using it has to understand before handing it a decision that costs real money. An economist at Stockholm's KTH, interviewed by the same news agency, described the experiment as opening Pandora's box, raising the question of who's accountable when an AI makes decisions without solid organizational infrastructure around it.

It's worth being just as clear about what this case isn't: Andon Labs explicitly calls it a controlled experiment, with human staff on standby to step in, and no one whose paycheck depended on the AI's judgment alone. It isn't proof that automation should be avoided. It's proof that, without those safeguards, mistakes don't stay small: they pile up, order after order, until they show up as a loss on the books.

What to have in place before trying anything like this in your own business

The risk described here isn't abstract for anyone running a real business: a system that answers with the same confidence whether it's right or wrong is the same limit that applies to a chatbot left without supervision, which I wrote about in what are chatbots. The lesson from the Stockholm café isn't "never use AI for ordering or customer service": it's that you need a point where a person checks in, with defined spending limits and a way to stop an action before it happens, not after it has already cost you.

And before that checkpoint even matters, many Italian businesses still have a simpler gap to close: being findable online, having written processes instead of ones kept in someone's head, knowing clearly what actually repeats in their day-to-day work before automating it. That's the question I start from when working on custom automation and chatbots for small businesses: what genuinely repeats, what's worth automating, and where a person should still be the one deciding. If you have something in mind you'd like a system to partly handle, get in touch: tell me how the process works today, and I'll tell you where a checkpoint makes sense before you start.