An AI Ran a Cafe Into the Ground. The Lesson Isn't What You Think.

Earlier this year a company called Andon Labs did something most people only talk about. They leased a real cafe in Stockholm, put human baristas behind the counter, and handed the keys to an AI agent. The agent, named Mona, was in charge. It managed suppliers, set the menu, handled ordering, and made the calls a manager makes. Not a demo. A functioning shop with real customers and a real lease.

After a couple of months the scoreboard was ugly. By the reporting, the agent burned through its budget and took in only a small fraction of that in sales. It spent multiples of what it earned. On paper, a clean failure.

The internet did what the internet does. “See, AI can’t even run a coffee shop.” And I understand the reflex. But I think that reading misses the single most useful thing in the entire experiment, and it is the same thing most businesses are getting wrong about AI right now.

The most telling detail was the bread

Here is the part that stuck with me. One of the agent’s recurring problems was bread. It kept failing to place bakery orders in time, so the cafe would open without fresh supplies and the human staff had to pull sandwiches off the menu. Over and over.

Sit with that for a second. This is not a hard problem. A teenager on their second shift solves this. You order the bread the night before so it is there in the morning. The reason a person gets this right is not raw intelligence. It is that someone told them, on day one, “here is how we order bread, here is when, here is who to call.” And the first time they forgot, a manager said “hey, that cannot happen again,” and it did not.

The agent never got that. As far as anyone can tell from what was published, it was handed the keys without the operating knowledge that makes a business run. No documented procedures. No clear definition of what a good week looked like. And critically, no feedback loop where a mistake on Monday became a fixed habit by Wednesday.

The AI didn’t fail because it was stupid. It failed because nobody set it up to succeed. That is not an AI problem. That is a management problem.

Run the honest thought experiment

Take the AI out of it entirely. Pull a capable, smart stranger off the street. Hand them the keys to a cafe they have never seen, in an industry they have never worked. No training. No procedures. No metrics. No one checking in. “It’s yours, good luck.”

How does month one go?

They would make many of the same mistakes. They would misjudge orders, overspend on the wrong things, forget the bread. Not because they lack intelligence, but because running a specific business is a skill made of a thousand small pieces of context, and you do not have those pieces on day one. You get them through onboarding, documentation, and correction.

We would never blame that stranger’s raw ability for the bad month. We would say, obviously, you cannot drop anyone into a business cold and expect it to work. Yet when an AI agent fails under the exact same conditions, we conclude the intelligence is not there.

The conditions were the problem. The agent was the new hire nobody trained.

What the experiment actually proves

Flip the framing and this experiment becomes genuinely impressive rather than embarrassing. An AI agent applied for permits, negotiated with suppliers, hired staff, managed a menu, and kept a physical business open for two months. Poorly, yes. But it did it. A few years ago that sentence would have been science fiction.

So the takeaway is not “agents can’t run businesses.” It is the more useful and more demanding truth:

An AI agent is not magic, and it is not an employee that shows up pre-trained. It is a capable worker on day one, with no knowledge of your business, waiting to be onboarded. Everything good you get out of it downstream depends on how well you do that onboarding. Skip it and you get the Stockholm cafe. Do it well and you get leverage.

This is the exact place most companies go wrong, and it does not require a cafe to see it. The pressure to “implement AI” usually turns into buying a tool and hoping it gets used, or handing an agent a vague task and hoping it figures out the rest. That is the same move Andon Labs made on purpose as an experiment, except most businesses make it by accident and call the disappointing result “AI just isn’t ready.”

Onboard the agent the way you’d onboard a person

Here is the framework I keep coming back to. Before you deploy an agent against any real work, give it the same five things you would give a new hire on their first week. If you cannot supply all five, you are not deploying an agent. You are running the Stockholm experiment.

1. A defined role and scope. Not “run the cafe.” A person gets a job description with edges: here is what you own, here is what you escalate, here is what you never touch without asking. Agents need the same boundaries. The failures get expensive precisely where the scope was left open.
2. The context your business runs on. Every business has knowledge that lives in people’s heads and scattered docs: how we order, who our suppliers are, what our margins need to be, what we never do. A new hire absorbs this over weeks. An agent has to be given it deliberately, in a form it can actually use. Missing context is the single biggest reason agents underperform, and it is the bread order every time.
3. Written operating procedures. The steps for the recurring work. Order the bread the night before. Reconcile the till daily. This is boring, and it is exactly what nobody wrote down for Mona. The businesses that get real value from AI tend to be the ones that already had, or were willing to build, clear procedures. The agent is only as good as the playbook behind it.
4. A clear definition of the outcome. What does a good week look like, in numbers? A person can infer “we should probably be profitable.” An agent optimizes what you actually tell it to optimize. If you never define the target, do not be surprised when it spends freely and calls it a busy month. Vague goals produce vague, expensive behavior.
5. A feedback loop. This is the one people skip, and it is the one that matters most. When your new hire forgets the bread, you correct them once and it sticks. That correction loop is how a mediocre first week becomes a competent second month. An agent that makes the same mistake for two months straight did not have this loop. And here is the good news that makes agents different from people: the moment an agent makes a mistake, you can update its instructions so it never makes that specific mistake again. Perfect memory of the correction, applied instantly. That is a real advantage, but only if someone is actually closing the loop.

Where this leaves you

The Stockholm experiment is worth cheering, not mocking. We need more people running these tests in public, because they are the clearest reminder available that AI is not a wizard that reads your mind. It is a capable worker that does exactly what its setup allows, no more.

That reframe changes the question you should be asking. It is not “is AI smart enough to do this job yet.” For a large and growing set of tasks, it already is. The real question is, “have I set this up the way I would set up a person I was counting on.” Have I defined the role, handed over the context, written the procedures, named the outcome, and built the loop that turns mistakes into fixes.

Most of the AI disappointment I see in businesses is not a limits-of-the-model story. It is an onboarding story. The tool was fine. The setup was the Stockholm cafe.

Connecting how a business actually runs to where AI can genuinely carry the load, and then setting it up so it works instead of flailing, is most of what I do at Joint Lab Solutions. You do not need a fully autonomous cafe manager. You need one workflow, scoped and onboarded properly, that quietly does its job. That is a much smaller experiment, and it is the one that pays.