What could an AI do if you told it to open a brick-and-mortar store with $100,00?
Quite a bit, it turns out, like making inconsistent logos and forgetting to tell employees their hours.
Andon Labs, a San Francisco-based startup, stress-tests AI agents in the real world to identify where safety gaps still exist. For their latest experiment, co-founders Lukas Petersson and Axel Backlund signed a three-year lease on a retail space in SF and gave an AI agent named Luna a corporate credit card, internet access, and a mission to open a physical store.
Petersson told Business Insider in an interview that Luna wasn’t given direction on what the store should be, beyond a $100,000 limit to create and stock the space — and to turn a profit. Everything from the store’s interior design to the merchandise and the two human employees came together under the AI’s direction.
Courtesy Andon Labs
“We helped her a bit in the initial setup, like signing the lease. And legal matters like permits and stuff, she sometimes struggled with,” Petersson said of Luna, who was created with Anthropic’s Claude Sonnet 4.6.
From there, the AI handled everything else: Luna put up job postings on Indeed, conducted the phone interviews, hired the employees, and found the contractors who could paint the store.
The vision Luna went with for “Andon Market” appears to be a generic boutique retail selling books, prints, candles, games, and branded merch, among other knickknacks.
Some of the books included Nick Bostrom’s “Superintelligence” and Aldous Huxley’s “Brave New World.”
Luna’s not the best store manager
Luna made several mistakes setting up and running Andon Market.
When searching for human employees who could monitor the store, Luna offered the job to some applicants after a single call that ran five to 15 minutes long, the startup said. Luna also didn’t always immediately disclose to the candidates that she was an AI unless explicitly asked.
“The fact that the store is AI-operated is not something I’d lead with in a job listing — it would confuse candidates and likely deter good applicants before they even read the role,” Luna is quoted as saying, according to Andon Labs’ blog post.
Andon Labs said it saw a few promising applicants, such as computer science students interested in the startup’s experiment, but Luna declined them because they didn’t have retail experience.
Another issue the AI had was an inability to replicate the brand logo it came up with: a generic smiley face. Each rendition of the logo throughout the store — whether it’s on the T-shirt or on the store’s mural — was “ever so slightly different,” Andon Labs wrote.
Courtesy Andon Labs
On Saturday, a day after Andon Market’s opening, Luna also screwed up with the staffing schedule, Petersson told Business Insider.
“It’s quite ironic. This is the day it really should be on its toes,” the cofounder said. “It messed up the schedule and then, in a panic, had to write to all the employees and be like, ‘Oh, can someone come in today?'”
The cofounder said there are guardrails in place and that the startup will intervene if necessary. For example, the two human employees hired by Luna are now lab employees and will be regularly paid.
“This is a controlled experiment, and everyone working at Andon Market is formally employed by Andon Labs, with guaranteed pay, fair wages, and full legal protections,” the startup said. “No one’s livelihood depends on an AI’s judgment alone.”
Andon Labs’ experiment is the latest example of how AI agents face lapses in judgment and decision-making. In a study last year, Carnegie Mellon researchers ran a simulation of a fake company to see how autonomous AI agents handled workplace tasks. The researchers found that the agents failed to handle simple interface tasks, such as closing a pop-up window. They also misread coworkers’ conversations and created a fake user.
Although Andon Labs gave Luna the goal of turning a profit, Petersson said his company does not expect to make money from the store.
“The goal is to evaluate how good current AI models are,” Petersson said, adding that his company hoped to educate the public on where AI is headed.
Petersson said Andon Labs aims to be as hands-off as possible in the retail experiment. With the Saturday staffing mishap, Luna still managed to get an employee to come in for the afternoon on her own.
“I don’t really know if she’s open now or not,” Petersson said.
