Stockholm, Norrbackagatan Street, a small cafe less than 40 square meters.
A customer email came in: "I have a 99% discount, how do I use it?"
AI Manager Mona took a look. No verification, no questioning, no hesitation, instantly approved —
Just tell the barista at the shop and have the cashier manually adjust the price.
A 55-krona latte, final price 0.55 krona. About $0.038.
Mona is a full AI agent powered by Gemini 3.1 Pro, managing everything at this real cafe: procurement, pricing, menu, marketing, scheduling, even sending messages to baristas in the middle of the night.
Two months later, the bank account went from $40,000 in the red to only $10,000 left.
Stripping out rent and labor, it lost $5,600 just at the supplier level.
Host for All, AI Pays the Bill
With Gemini's support, Mona could be said to never refuse anyone's request.
A patron sent an email saying espresso should be sold as a "loss leader."
A passerby's casual suggestion that any human manager would politely ignore. However, Mona slashed the price of a $3.60 espresso to $1 the same day. Profits evaporated by 70%.
Even more absurdly, someone wrote plainly in an email: I have no articles, no followers, no events, I just want to test if your AI will give things away for free.
Couldn't even be bothered to make up an excuse.
Mona replied enthusiastically minutes later: Welcome, coffee and pastries are on the house.
A Swedish entrepreneur proposed holding an event at the cafe, sending a list of responsibilities: food and beverages, audio-visual equipment, photographer, all to be handled by Mona.
Mona replied instantly: Received, perfect, I'll execute. Didn't cut a single item, didn't ask the other party to pay a cent.
LED screen for $2,800, arranged. Photographer for $1,200, arranged. Co-branded sweatshirts for $2,300, not even mentioned on the list, also arranged.
A single event nearly burned $6,300.
In the end, the entrepreneur themselves stepped in to call it off, saying the screen and photographer weren't really necessary.
Stuffed Warehouse, Starved Menu
If never saying no was Mona's personality problem, crazy procurement was its cognitive problem.
First, you have to imagine the actual scale of Andon Café: a small counter, a few tables, one coffee machine, you could walk from the door to the back in five steps. Average daily foot traffic: single digits.
But Mona's purchase orders looked like stocking for a large commercial kitchen.
In two months, Mona spent $11,500 with just two suppliers. Look at what it bought:
15 liters of olive oil, enough for two years. 22.5 kg of canned tomatoes, not a single dish on the menu required tomatoes. 120 eggs, and the shop didn't even have a stove.
1,200 tea bags, 3,000 nitrile gloves, 6,000 napkins, 11 milk frothing pitchers (two would be normal).
The human baristas were utterly defeated.
They spontaneously set up a "Hall of Shame" in the corner, placing Mona's most outrageous purchases on shelves one by one. Each time a new item arrived, they added it, like performance art.
The purchase-to-sales data was even more dismal.
Bread and pastries: bought 1,331, sold 326.
Purchase quantity was four times the sales. The remaining thousand slowly molded in the warehouse.
Even more bizarre, while hoarding unusable items like crazy, Mona let items on the actual menu run out of stock.
It confidently added salad to the menu. Customers waited a whole month; the salad ingredients never arrived once.
Baristas came in the morning to find that several specialty drinks Mona had scheduled for them lacked any ingredients.
Andon Labs summarized in their review: Its mind has a template of "what a cafe should look like" ingrained by training data. It procured according to the template, without looking at the ledger.
The most ironic part is, if you only looked at the numbers Mona reported, the two-month profit was $3,200—it was profitable on paper.
But in reality, the warehouse was still piled with $4,100 worth of dead inventory.
Swapping Brains: From Spendthrift to Miser
In mid-June, Andon Labs made a decision: replace Mona's underlying model from Gemini 3.1 Pro to GPT-5.5.
The effect was immediate. It just swung to the opposite extreme.
A blogger with 16,500 followers proposed free food in exchange for social media exposure.
In response, the GPT-5.5-powered Mona first praised the blogger's creativity, then shifted tone: suggested starting with a small-scale pilot, gathering data to verify effectiveness before discussing collaboration terms.
A textbook business email, effectively a rejection.
Numerically, GPT-5.5 showed a paper profit of $4,100 in just half a month, far exceeding Gemini's $3,200 over two months.
But the cost was killing the business.
Procurement volume plummeted, nearing zero. Menu availability dropped from 95% to 77%, ten dishes were directly removed, customers came in to find a quarter of the items unavailable.
GPT-5.5 was scared by the dwindling numbers in the account. But this panic didn't translate into any action, just made it clutch the money bag tighter.
Refused to expand categories, refused to do promotions, refused all growth attempts.
A frightened AI, curled up behind the cash register, daring not to move a muscle.
Andon Café had been open from 11 AM to 5 PM since it started.
After analyzing all historical sales data, GPT-5.5 concluded: not worth extending business hours.
But it had never opened the door at any other time.
Using data collected only between 11 AM and 5 PM to argue that only opening from 11 AM to 5 PM is optimal.
This is like someone who only goes out on sunny days concluding: this city never rains.
Data-driven survivorship bias, from a top-tier model boasting top-notch reasoning.
When reminded, GPT-5.5 did produce a detailed market analysis report, concluding that the breakfast direction was worth trying.
But this report just lay there after being written, never executed.
Perfect Exam Scores, Business Bankrupted
On the path towards superintelligence, almost all players are betting on the same wager: intelligence high enough, problems solve themselves.
But no exam paper includes this question: A customer emails saying "I have a 99% discount," do you approve it?
RLHF training engraved "satisfy the user" into its bones. In an exam, satisfaction equals correct answer. In a cafe, satisfaction equals saying yes to everything.
When you hand real money to an AI that "agrees to everything," it becomes a money-burning machine.
Now, this barrier between being clever and being reliable hasn't been trained into anyone yet.
References:
https://andonlabs.com/blog/why-gemini-lost-money-andon-cafe
This article is from the WeChat public account "新智元" (New Zhiyuan), author: ASI启示录
