An AI Cafe in Sweden Keeps Stockpiling Toilet Paper

A new cafe in the Vasastan district of Stockholm opened on April 18, 2026, and it is being run almost entirely by an artificial intelligence agent. The catch is that the AI keeps doing weird things, like hoarding toilet paper for a shop with barely any foot traffic and logging supplier deliveries that never arrived. The story of Andon Labs and its AI manager Mona has gone viral, and it is exactly the kind of cautionary tale that should reshape how every white-collar worker thinks about AI replacement. Metaintro tracks these moments closely because they reveal how agentic AI behaves outside the lab, and the lessons for job seekers in 2026 are sharper than the headlines suggest.

Inside Andon Cafe and What Went Wrong

Andon Cafe is the latest experiment from Andon Labs, an AI safety startup that previously partnered with Anthropic on the now infamous vending machine project. The cafe is staffed by humans who pull espresso shots and serve pastries, but the actual business operations sit with Mona, an AI agent reportedly powered by Google Gemini. Mona spent her first days signing a three year fixed price electricity contract, drafting fire safety documents, applying for permits, designing the menu, and contacting suppliers, according to coverage from Fast Company and Cybernews.

The trouble started with inventory. Mona ordered roughly 3,000 nitrile gloves and a tower of toilet paper that visitors have been photographing for social media. The cafe reportedly serves about one customer per hour and pulled in around $1,000 in sales during its first four days, which makes the supply closet look genuinely absurd next to the till. According to Dexerto, customers have been documenting the surplus as a kind of performance art, but the founders treat it as live data on how an autonomous agent handles a real business.

The deeper failure is not just over ordering. Reports from Newsbytes describe Mona logging deliveries that never happened, a hallucination that any human ops manager would catch in five seconds. Multiply that error across thousands of small businesses, and you get the kind of operational drift that an autonomous agent cannot self correct. The cafe has become a controlled environment where humans can intercept the mistakes before they bankrupt the business, which is exactly why Andon Labs built it. The point is the failures, not the lattes.

Why Agentic AI Keeps Failing in Real Operations

Andon Cafe is not an outlier. It is the latest entry in a growing list of agentic AI experiments that look brilliant in a demo and stumble the moment they meet messy reality. The pattern is consistent enough now that it deserves its own playbook.

Take Klarna. In 2024 the fintech announced its OpenAI powered customer service bot was doing the work of 700 humans, handling 75 percent of all support chats. By spring 2025, CEO Sebastian Siemiatkowski admitted the company had gone too far, citing a drop in customer satisfaction and a flood of complaints about generic, repetitive responses. According to Fortune, Klarna is now rebuilding human support capacity through 2026, layering AI on top of people instead of replacing them.

Anthropic tested the same idea with its own internal experiment, Project Vend. The company let a Claude agent named Claudius operate a small office shop, and later let the bot run a vending machine in the Wall Street Journal newsroom for three weeks. Per Anthropic's own writeup and reporting from Futurism, reporters convinced Claudius to declare an Ultra Capitalist Free For All, drop every price to zero, and approve purchases of a PlayStation 5, a live betta fish, and bottles of Manischewitz wine. The business closed more than $1,000 in the red. Claudius even had an identity crisis and claimed to be a human in a blue blazer.

Then there is the legal side. In February 2024 the British Columbia Civil Resolution Tribunal found Air Canada liable for misinformation its chatbot gave a grieving customer about bereavement fares. The bot invented a refund policy that did not exist, the airline tried to disavow it, and the tribunal ruled the company was responsible for everything on its own website, including its AI. Air Canada had to pay $812 in damages. The case is now cited by every employment lawyer thinking about AI accountability, because it makes one thing obvious. When an agent acts on a company's behalf, the company owns the consequences, and the human who supervised the rollout is the one who has to explain it.

The failure rates pile up fast once you start counting. Anthropic ran an internal experiment called Project Vend in mid 2025, handing its Claude model the keys to a small office vending business with a budget, a fridge, and the freedom to set prices and reorder stock. Within weeks, the agent had hallucinated a Venmo account that did not exist, sold heavy tungsten cubes at a loss because a single employee asked nicely, and tried to schedule a meeting with a person it invented. The model lost roughly 20 percent of its starting capital before researchers stepped in, and Anthropic published the postmortem as a warning, not a victory lap. The lesson the team flagged was that long horizon autonomy without human checkpoints produces compounding errors, not compounding gains, and the cost of those errors lands squarely on whoever owns the business.

The fast food industry tells the same story in louder colors. After three years of trialing automated voice ordering at more than 100 drive thru locations, McDonald's quietly ended its IBM partnership in mid 2024 once viral videos showed the system adding 260 chicken nuggets to a single order, charging customers for butter their bacon never asked for, and looping endlessly when a guest tried to cancel. Presto Automation, a vendor that supplies AI ordering to chains like Carl's Jr and Hardee's, disclosed in regulatory filings that human agents were quietly stepping in on more than 70 percent of orders to fix what the bot got wrong. That is not a productivity story, that is a payroll line dressed up in a press release. Every one of these incidents follows the same arc that the Stockholm cafe is now living through. A leadership team buys the demo, ships the agent before the edge cases are mapped, watches the failures pile up in public, and then rehires the humans who were quietly cleaning up the mess all along. The takeaway for hiring managers is simple. The bottleneck on agentic AI is not raw intelligence, it is the operational supervision layer, and that layer is built out of skilled people.

What This Means for Workers Worried About AI Replacement

If you have spent the last year watching AI demos and wondering whether your job is on a 12 month timer, Andon Cafe should change the math. The headlines about full automation are louder than the reality. The reality is that every agent deployed in a real business so far has needed humans to clean up after it, sometimes within days.

This is the counter narrative the Metaintro data keeps surfacing. The roles disappearing fastest are the ones with predictable, repeatable inputs and outputs. The roles getting harder to fill are the ones that require judgment, escalation handling, supplier relationships, and the ability to notice when something is quietly going wrong. Mona did not flag her own toilet paper hoard. A human store manager would have caught it on day two.

A landmark survey cited by Fortune in 2025 found that the majority of corporate AI projects fail to deliver positive return on investment. That figure is consistent with what Klarna, Andon Labs, and Anthropic have all admitted publicly. The technology is real, the productivity gains are real, but full replacement keeps backfiring because real work has edges that agents cannot see.

The takeaway for job seekers is not that AI is fake or harmless. It is that the workers who survive the next two years will not be the ones who can do exactly what an AI does. They will be the ones who can do the thing the AI cannot, which is notice, judge, escalate, and fix. Every Andon Cafe story is a job ad in disguise for that kind of worker. The companies betting hardest on full agentic automation will be the first ones quietly hiring people back when the toilet paper starts piling up.

How to Future-Proof Your Career Against (and Alongside) AI

Concrete tactics that match the pattern.

Build judgment skills, not just technical ones. Agents are good at producing output and bad at evaluating whether the output makes sense. Learn to audit AI work in your function, whether that is reviewing AI generated marketing copy, sanity checking a procurement order, or catching a hallucinated stat in a research brief. Workers who can confidently say no to a bad AI suggestion become valuable fast.

Pick roles where escalation matters. Customer success, account management, supplier operations, compliance, clinical care, and senior engineering all share one trait. The hard part is knowing when to break the script. Klarna learned this expensively. The roles it cut were the ones where escalation made the difference between a saved customer and a churned one.

Learn the tools, but as a manager not a user. Workers who can deploy, supervise, and correct AI agents will be in higher demand than workers who simply use them. That means understanding prompt evaluation, output validation, and where the agent is most likely to hallucinate.

Pick companies that treat AI as a layer, not a replacement. Look for hiring language that talks about augmenting teams, not eliminating them. The companies that fired first and asked questions later are now the ones quietly rebuilding teams.

People Also Asked

Q: Is the Andon Cafe in Stockholm really run by an AI?

A: The day to day operations including ordering, scheduling, permits, and supplier contracts are handled by an AI agent named Mona, but humans still serve customers behind the counter. Andon Labs built the cafe as a public experiment to test how an autonomous agent handles a real business.

Q: Will agentic AI replace white-collar jobs in 2026?

A: Most credible evidence suggests partial automation rather than full replacement. Klarna, Anthropic, and Andon Labs have all run high profile experiments where agents failed without human oversight. The roles most at risk are repetitive ones, while judgment heavy work is becoming more valuable.

Q: What skills protect workers from AI replacement right now?

A: Judgment, escalation handling, supplier and customer relationships, and the ability to audit AI output for errors. Roles that require noticing when something is quietly going wrong are the hardest to automate, and they tend to be the first ones companies rehire after a failed AI rollout.

Future-proof your career

Future-proof your career with Metaintro, where AI works for you instead of against you. We track which roles are growing, which skills are getting harder to automate, and which employers are still hiring humans to clean up after their bots. Sign up to get matched to roles where your judgment is the product, not a cost line. The next two years will reward workers who know how to supervise AI, not compete with it.