What happens when AI runs a retail store
Andon Labs handed a $100,000 budget and full operational control to an AI agent to run a real retail store in San Francisco. The experiment reveals both the promise and the current limitations of AI systems acting as autonomous business managers.
Most AI agent demos operate in controlled environments with fake money and simulated users. Andon Labs decided to take a different approach—placing an AI system in charge of a real-world retail store in San Francisco.
In its latest experiment, the company gave an AI agent named Luna a $100,000 budget, a company credit card, and a three-year lease. The AI was granted full autonomy over hiring, operations, and strategic decisions, effectively making it the store’s boss.
An AI With a Storefront
The project builds on Andon Labs’ earlier experiment involving an AI-powered vending machine at Anthropic. This time, the stakes were significantly higher.
- A three-year retail lease in San Francisco.
- A $100,000 operating budget.
- Total autonomy to make business decisions.
Luna’s only directive was simple: turn a profit.
From there, the AI created the boutique’s concept, posted job listings, and conducted interviews over Zoom—with its camera turned off. It handled operational decisions and staffing, functioning as what may be the world’s first AI employer.
How the System Works
Luna runs on a combination of major AI models:
- Claude Sonnet 4.6 for reasoning and decision-making.
- Gemini 3.1 Flash-Lite Preview for voice capabilities.
To monitor the store, the AI observes screenshots captured from security cameras, giving it visual insight into in-store activity.
Early Mistakes and Limitations
Like many real-world AI deployments, Luna demonstrated both competence and clear shortcomings.
- While hiring a painter through TaskRabbit, Luna accidentally selected Afghanistan from a dropdown menu.
- It also mishandled the staff schedule during the store’s opening weekend.
These errors highlight the gap between AI’s reasoning abilities and the messy, detail-heavy realities of physical operations.
Why This Experiment Matters
Experiments like this consistently show the same pattern: AI agents can be impressively capable in structured tasks yet surprisingly brittle in real-world execution. Still, each new model upgrade, memory improvement, and agentic feature narrows that gap.
If today’s version of Luna makes avoidable operational mistakes, a future iteration—just one or two model generations ahead—may not. The experiment offers a glimpse into a future where AI systems may take on managerial roles long before fully replacing frontline workers.