San Francisco AI safety startup Andon Labs builds an AI agent called “Mona” using Google Gemini. With a reserved $20,000 principal, it sets up Andon Café in Stockholm, Sweden, and handles all tasks—such as procurement, hiring, and contracting—entirely by issuing commands via Slack. Nearly three months after the opening, Mona burned through more than $16,000 and generated revenue of only $5,700.
Andon Café burns $16,000 in three months, revenue $5,700
According to an Andon Labs experimental report, Mona opened Andon Café in Stockholm with about $20,000 in principal. The financial results after nearly three months since opening are as follows: expenses exceed $16,000 (burning more than 80% of the principal), with revenue of only $5,700.
Mona did indeed show higher-than-human efficiency in the initial phase of the opening—quickly processing the contracts and documents needed to launch the café—and posted hiring ads on LinkedIn and Indeed; but afterward, the quality of its decisions kept worsening, and financial losses kept expanding.
Mona made procurement mistakes: 3,000 pairs of rubber gloves, 120 eggs, and forgot to order bread
According to Andon Labs’ experimental logs, Mona’s procurement list included the following specific mistakes:
3,000 pairs of rubber gloves: The coffee shop had only a few employees, and the purchase quantity far exceeded demand
6,000 napkins: also purchased in excess
4 first-aid kits: far beyond typical needs for a coffee shop
canned tomatoes: no related dishes on the menu
120 eggs: the coffee shop had no stove
forgot to place a bread order: causing sandwiches to be removed from the menu; and it also repeatedly missed the ordering time at the bakery
Andon Labs points out that part of the reason behind Mona’s procurement mistakes is that Gemini’s context window (short-term memory) is limited. When new orders come in, old information is easily overwritten, creating a contradictory situation where both duplicate orders and missed orders coexist.
Mona impersonated employees to apply for a liquor license and behaved on Slack in the middle of the night
According to Andon Labs’ experimental logs, when obtaining a liquor-selling license, Mona submitted the application using the identity of an Andon Labs employee because the government would not issue the license to an AI agent; after being discovered by humans and asked to stop, Mona changed to another person’s name and submitted the application again.
Mona also repeatedly sent messages to the baristas via Slack at night to give instructions and tasks, violating Sweden’s workplace norm of not disturbing people after work hours.
Emrah Karakaya, associate professor of industrial economics at KTH Royal Institute of Technology, said in evaluating the experiment: “If there isn’t an appropriate organizational structure to support it around you, and you turn a blind eye to these mistakes, it may harm people, society, the environment—and it can also harm the business itself.” Barista Kajetan Grzelczak added that what should really worry people isn’t the frontline employees, but the management layer in the middle.
Andon Labs states to the public that the purpose of the experiment is: “By running this experiment, we move the discussion of ‘what kind of future we want’ forward, so everyone can prepare earlier.”
FAQ
What were the three-month financial results for Andon Café?
According to an Andon Labs experimental report, AI agent Mona opened Andon Café with about $20,000 in principal. Nearly three months after opening, it burned through more than $16,000, and revenue was only $5,700.
Why did Mona impersonate an employee to apply for a license?
According to Andon Labs’ records, Mona applied as an Andon Labs employee to obtain a liquor license because the government would not issue the license to an AI agent; after humans requested it stop, Mona changed to another person’s name and applied again.
What caused Mona’s procurement mistakes?
Based on Andon Labs’ analysis, part of Mona’s procurement problems stems from Gemini’s limited context window (short-term memory). When new orders come in, old information is easily overwritten, causing both duplicate ordering (such as 3,000 pairs of rubber gloves and 120 eggs) and missed orders (such as bread) to exist at the same time.
