Test giving $100,000 to AI to open a store

The test was conducted by Andon Labs, a startup in San Francisco, with the goal of testing the ability of AI agents to operate in the real world, thereby identifying and evaluating existing security vulnerabilities.

In it, co-founders Lukas Petersson and Axel Backlund signed a three-year lease for space in San Francisco, then gave an AI agent named Luna a business credit card, Internet access, and the task of running the physical store. This AI was created using Anthropic’s Claude Sonnet 4.6 software.

Luna has a limited capital of 100,000 USD, must complete the task of designing, selling goods and making a profit. Everything from the store’s interior design to the stocking schedule and managing two employees is done under the direction of artificial intelligence.

“We help ‘her’ a bit with the initial setup stages, like signing the lease and legal issues like licensing,” Petersson told Business Insider.

Luna then handled almost everything, including posting the job ad on Indeed and conducting phone interviews, finding contractors who could paint the store. This AI’s vision is to build a typical small retail store: selling books, prints, candles, games, and a few other sundries. Called Andon Market, this is considered the first physical store run by AI.

Luna does not tell candidates that she is AI during interviews. “I won’t mention that the store is powered by artificial intelligence, because that would mislead candidates and potentially cause qualified people to quit before reading the job description,” the Andon Labs blog explains.

However, in the process of setting up and operating Andon Market, Luna made some mistakes. For example, when looking for store supervisors, AI accepts candidates after only a 5-15 minute call, a time that is considered too short.

Another problem the AI ​​encountered was its inability to create impressive logos, instead coming up with generic smiley face images. Each version of the logo in the store is “slightly different” instead of being consistent across the entire system.

Right after Andon Market opened in the middle of last week, Luna also made a mistake in arranging the work schedule, not assigning anyone to be on duty on the first day. “It’s ironic that this is what AI should do best, but it messed up the schedule,” Petersson said. “In a panic, he wrote to all the employees and asked: Can anyone come to work today?”.

Luna’s troubles in the early days forced Andon Labs to intervene directly. Two employees recruited as official employees of the company also receive full benefits. “This was a controlled experiment,” Petersson added.

Although Andon Labs gave Luna a profit target, Petersson said his company did not expect to make money from the store. Instead, their focus is on assessing how good current AI agents are, whether they live up to the hype, and helping the public better understand the direction of artificial intelligence development.

Theo NBC NewsAndon Labs’ experiment is the latest example of how AI agents make errors in judgment and decision-making in real-world environments. In a study last year, a team of Carnegie Mellon University experts ran a simulation of a fake company to see how automated AI agents handled workplace tasks. As a result, these agents cannot handle simple interface tasks, such as misinterpreting some of a colleague’s conversations, or even creating fake users.