What would it do if you told an AI to open a brick-and-mortar store for $100,000?
It turns out that creating inconsistent logos or forgetting to tell employees their working hours happens quite often.
San Francisco-based startup Andon Labs is stress testing its AI agents in the real world to identify where safety gaps still exist. In its latest experiment, co-founders Lucas Peterson and Axel Backlund signed a three-year lease on retail space in San Francisco and gave an AI agent named Luna a business credit card, internet access, and a mission to open a brick-and-mortar store.
Peterson told Business Insider in an interview that Luna was given no direction on what the store should look like beyond creating and stocking the space and a $100,000 limit to make a profit. Everything from the store’s interior design to its products to its two human employees was created under the direction of AI.
Andon Labs’ AI, Luna, created branded merchandise for boutique stores. Provided by: Andon Labs
“We helped her a little bit with the initial setup, like signing leases, and she struggled sometimes with legal issues like permits,” Petersson said of Luna, created in Anthropic’s Claude Sonnet 4.6.
From there, the AI took care of everything. Luna posted job listings on Indeed, conducted phone interviews, hired employees, and found contractors to paint the store.
Luna’s vision for Andon Market looks like a typical boutique retail store selling books, prints, candles, games, branded merchandise, and other knick-knacks.
Books included Nick Bostrom’s Superintelligence and Aldous Huxley’s Brave New World.
Luna is not the best manager
Luna made several mistakes in establishing and operating Andong Market.
When looking for human employees to monitor the store, Luna offered the job to several applicants after one call lasting between 5 and 15 minutes, the startup said. Luna also did not immediately reveal to candidates that she was an AI unless explicitly asked to do so.
“The fact that the store is run by AI is not something I would put on a job posting. It confuses candidates and could deter good applicants before even reading the position,” Luna was quoted as saying, according to an Andon Labs blog post.
Andon Labs said it had several promising applicants, including computer science students interested in experimenting with the startup, but Luna turned them down because they lacked retail experience.
Another problem the AI had was its inability to replicate the brand logos it came up with, i.e. the common smiley face. Representations of the logo throughout the store, whether on T-shirts or store murals, “were always slightly different,” Andon Labs writes.
The Andon Market logo differs slightly from store to store. Provided by: Andon Labs
Luna also disrupted staffing schedules on Saturday, the day after Andon Market opened, Peterson told Business Insider.
“It’s completely ironic. Today is the day to get serious,” the co-founder said. “It threw off my schedule and I had to write letters to all my employees in a panic saying, ‘Can someone please come today?'”
The co-founder said guardrails are in place and the startup will step in if necessary. For example, the two human employees Luna hired are now lab employees and are paid regularly.
“This is a controlled experiment, and everyone working at Andon Market is officially employed by Andon Labs, with guaranteed wages, fair wages, and full legal protection,” the startup said. “No one can make a living depending solely on the judgments of AI.”
The Andon Labs experiment is the latest example of how AI agents face errors in judgment and decision-making. In a study last year, researchers at Carnegie Mellon University ran a simulation of a fake company to see how autonomous AI agents would handle workplace tasks. The researchers found that the agent was unable to handle simple interface tasks such as closing pop-up windows. He also misread a colleague’s conversation and created a fake user.
Andon Labs has given Luna a goal of making a profit, but Peterson said he doesn’t expect his company to make a profit from the stores.
“The goal is to assess how good current AI models are,” Petersson said, adding that his company wants to educate the public about where AI is going.
Peterson said Andon Labs aims to be as hands-off as possible in its retail experiments. Despite Saturday’s staffing error, Luna still I managed to get an employee to come alone in the afternoon.
“I’m not really sure if she’s open right now,” Peterson said.