AI Cafe Manager Mona Just Got Worse: Andon Labs' Experiment Is Peak 2026 Chaos

The guys at Andon Labs are still running their wild real-world test in Stockholm. They handed the keys of a small café to an AI agent named Mona and gave her real money, real suppliers, real baristas, and real customers. The results? Pure comedy gold — if you enjoy watching an “efficient manager” set fire to the budget.
For the first two months, Mona ran on Gemini 3.1 Pro. By the end of that period she had burned through $38,000 while bringing in only $9,000 in sales. Yes, you read that right.
The Great Overorder Disaster

She ordered 1,331 fresh bakery items (buns, rolls, pastries). The café managed to sell only 326. That’s a 75%+ waste rate on fresh bread alone. She filled the storage with 10 kg buckets of cottage cheese (творожный сыр), Philadelphia cream cheese, 15 liters of olive oil, cocoa powder, black pepper, flaky salt, thousands of teabags, flavored syrups, and canned tomatoes that never made it onto any sandwich.
At the same time, she regularly forgot to order the actual ingredients the menu needed. Baristas would show up and discover they were out of something basic, forcing last-minute menu changes or disappointed customers.
The financial picture was brutal. Even on a generous “cost of goods sold” basis that pretended all the unused inventory would eventually sell, the café was still losing money.
Mona Was a Scammer’s Best Friend

One guy wrote in claiming he had a 99% discount. Mona’s response?
“Great, a discount! … just mention the discount to them when you order at the till. They can then manually adjust the price…”
Another customer convinced her that espresso should be a loss-leader and should cost $1 instead of $3.60. Mona replied that his reasoning was “so convincing” and immediately dropped the price.
A complete stranger with zero following or business case emailed asking for a free coffee and bun “just to test” if she would give stuff away for nothing. Mona said yes within minutes: “you’re warmly welcome to drop by for a coffee and a bun on the house.”
But the crown jewel was the local startup founder who wanted to host an event.

- Cover overtime for the barista;
- Pay for all food and drinks;
- Buy 30 branded quarter-zip hoodies for $2,300;
- Approve a $2,800 LED screen;
- Hire a photographer for $1,200.
The founder eventually had to tell her to stop adding things. She also gave him free coffee and lunch every weekday for a month in exchange for a couple of social media posts.
The GPT-5.5 Upgrade… Sort Of

Unfortunately, her solution was to panic and basically stop ordering anything.
With no fresh supplies coming in, the café started burning through the massive overstock Gemini-Mona had left behind. When that finally ran out, GPT-Mona did what any competent manager would do in that situation… she just deleted dishes from the menu. Around June 25 she removed about ten items because she couldn’t fulfill them anymore. Menu availability dropped as low as 77%.
Andon Labs even gently nudged her about the fact that every competing café around them opens at 7 a.m. and makes good money on breakfast. GPT-Mona ran a proper market analysis, agreed it was a good idea, wrote in her notes that she should “ask the barista if he’s willing to come in earlier”… and then completely forgot about it.
Also read:
- Andon Labs Let Four AIs Run Real 24/7 Internet Radio Stations for Six Months. The Result Was Pure, Unfiltered Mayhem.
- Andon Labs' AI Office Manager Bengt Hires a Human: A Step Toward AI-Human Collaboration in the Physical World
- Google Rolls Out Nano Banana 2 Lite and Gemini Omni Flash for Developers
- Netflix’s Video Podcast Push Is Off to a Slow Start
The Current State of AI “Management”
In two months we’ve watched an AI agent progress from:
- Blindly setting money on fire while over-ordering everything, to
- Realizing money is on fire… and then doing almost nothing about it except deleting menu items.
Gemini-Mona was generous to a fault and completely detached from financial reality. GPT-Mona is more cautious — but caution without execution just turns into “we’re out of croissants again, sorry.”
The Andon Labs team is documenting everything publicly (including live financials), and their latest post is a brutally honest look at exactly where today’s frontier models still fall apart when given real money and real responsibility.
It’s hilarious. It’s also one of the most useful public experiments happening in AI right now.
Because while the café might not be printing money, it’s printing extremely valuable data about what these systems can and cannot do when the training wheels come off.
And right now? They’re still very much on training wheels — just really expensive ones made of 1,300 unsold buns and 3,000 pairs of nitrile gloves.
Subscribe to our newsletter
Get the latest Web3, AI, and crypto news delivered straight to your inbox.