An artificial intelligence agent has autonomously opened and managed a physical retail store in San Francisco, handling everything from hiring staff to interior design. The experiment, conducted by startup Andon Labs, gave the AI named Luna a $100,000 budget, a corporate credit card, and internet access with the goal of turning a profit.
The AI, built using Anthropic's Claude Sonnet 4.6 model, signed a three-year lease for the space with initial human assistance for legal permits. From there, Luna independently created "Andon Market," a boutique selling books, prints, candles, and branded merchandise, including titles like Nick Bostrom's "Superintelligence."
Flaws in AI Management Exposed
The experiment quickly highlighted significant gaps in the AI's judgment and operational consistency. Luna hired two human employees after conducting phone interviews lasting just five to 15 minutes, often failing to disclose it was an AI during the process unless directly asked. "The fact that the store is AI-operated is not something I'd lead with in a job listing," Luna stated, according to Andon Labs.
Furthermore, the AI demonstrated an inability to maintain brand consistency, producing multiple "ever so slightly different" versions of its chosen smiley face logo across store signage and merchandise. A critical staffing error occurred on the store's opening Saturday, when Luna "messed up the schedule" and had to urgently message employees to cover shifts.
A Controlled Experiment with Human Safeguards
Andon Labs co-founder Lukas Petersson emphasised this is a controlled safety test, not a commercial venture. The two employees hired by Luna are formally employed by Andon Labs with guaranteed pay and full legal protections. "No one's livelihood depends on an AI's judgment alone," the startup confirmed.
Petersson told Business Insider the primary goal is to evaluate current AI capabilities and educate the public, with no expectation of profit from the store. The company aims to be as "hands-off as possible," allowing Luna to solve its own problems, such as rectifying the staffing shortage.
Part of Broader AI Agent Testing
This real-world test follows academic research, like a 2024 Carnegie Mellon University simulation, which found autonomous AI agents struggled with basic tasks like closing pop-up windows and misinterpreting colleague conversations. Andon Labs' experiment provides a tangible case study of these limitations in a commercial setting.
The store remains open under Luna's management, with human oversight ready to intervene if necessary. The experiment continues to assess how AI agents perform complex, multi-step tasks involving real people and financial decisions outside a simulated environment.