A bold experiment by Andon Labs has drawn attention across the tech world after handing full operational control of a retail store to an artificial intelligence system. With a budget of $100,000, the AI was expected to independently set up, hire staff, and manage the business. However, the outcome revealed a notable gap between advertised capabilities and real-world performance.
The experiment was led by co-founders Lukas Petersson and Axel Backlund in San Francisco. The team leased a retail space for three years and granted an AI agent named “Luna” access to a corporate credit card, internet connectivity, and full authority to run the store.
The AI was built using Claude Sonnet 4.6 developed by Anthropic. Its responsibilities included designing the store layout, selecting products, managing operations, and generating profit. From interior design to supervising two employees, nearly every aspect was handled by the AI, with minimal human assistance limited to initial legal and administrative setup.
The store, named Andon Market, was envisioned as a small retail space selling books, prints, candles, games, and various novelty items. Notably, Luna independently posted job listings on Indeed and conducted phone interviews without disclosing that it was an AI, aiming to avoid bias or hesitation from applicants.
Despite its autonomy, Luna quickly exposed several limitations. During recruitment, it made hiring decisions after phone calls lasting only 5 to 15 minutes—far too brief to properly evaluate candidates. This raised concerns about the AI’s judgment and ability to assess human capability.
On the branding side, the AI struggled to create a consistent visual identity. The store’s logo lacked distinctiveness and appeared slightly different across versions, undermining professionalism and brand coherence.
A critical issue emerged on the store’s opening day, when Luna failed to assign any staff to work. In response, the AI sent a mass email asking if anyone could urgently come in, highlighting its difficulty in handling real-world operational scenarios.
These early setbacks forced Andon Labs to step in and stabilize operations. Still, the company emphasized that the project was a controlled experiment, with profit not being the primary objective. Instead, the goal was to evaluate how capable current AI agents truly are and whether they live up to the expectations set by the industry.
According to NBC News, the experiment reflects a broader issue: AI agents continue to struggle with decision-making in real-world environments, especially in situations requiring flexibility and social understanding.
Earlier research by Carnegie Mellon University reached similar conclusions. In a simulated workplace environment, AI agents failed to handle even basic interface tasks, misinterpreted conversations, and in some cases generated non-existent user data.
Ultimately, the Andon Labs experiment serves as a reminder that while artificial intelligence is advancing rapidly, it still falls short of fully replacing humans in roles that demand nuanced judgment, adaptability, and real-world experience.
