Study reveals AI retail agent runs $100k shop in San Francisco

An AI system is now operating a physical store in San Francisco while also hiring human staff, marking a striking shift in how automation is entering everyday retail. The experiment, led by AI startup Andon Labs, places its AI agent Luna in charge of a boutique with a $100,000 budget and full operational control.

The project has drawn attention for blurring the line between software tools and workplace authority, with the system acting not just as an assistant but as a decision-maker.

The experiment was developed by Andon Labs and deployed in a small gift shop in San Francisco’s Cow Hollow district. The AI agent, built using Anthropic’s Claude Sonnet model, was given a three-year lease, internet access and autonomy to build a profitable business.

The store sells items such as branded merchandise, candles and chocolates, operating as a test case for fully agent-driven commerce.

Luna’s responsibilities extended beyond product decisions. The system posted job listings on platforms such as Indeed and Craigslist, conducted interviews and selected staff.

Some candidates reportedly did not realise they were speaking with an AI during interviews. The agent ultimately hired two employees, making them part of what the company described as the first retail team managed entirely by an AI system.

The hiring process also included automated screening decisions, with preferences based on perceived retail experience rather than academic background.

While being an autonomous system, the project has encountered operational errors, such as schedule errors, and selection problems while choosing contractors. It demonstrates the limitations of the existing agent-based approach to the problem.

For example, the company has developed guardrails that would help to interfere if necessary, especially concerning administrative and legal procedures.

Based in San Francisco, this project is part of a larger trend of developing autonomous business management systems.

According to early reports from Business Insider, the system works fine, but it is still experimental and has many imperfections.