Project Vend: When AI Becomes a Shopkeeper – What We Can Learn from Anthropic’s Experiment
Introduction: AI in Real-World Business
What happens when you give artificial intelligence responsibility for a real business? That’s exactly what Anthropic tested with “Project Vend.” Together with Andon Labs, the company let its language model Claude Sonnet 3.7 run a small, automated kiosk in its own San Francisco office. The goal was to test the capabilities and limits of AI as an independent economic agent in the real world—beyond simulations.
How the Experiment Worked
The AI, internally called “Claudius,” had access to various tools: web search for product research, a simulated email system for orders and communication, note-taking functions for managing inventory and finances, Slack for customer interaction, and the ability to change prices in the checkout system. Claudius was supposed to independently determine the assortment and prices, manage stock levels, and respond to customer requests. Physical tasks like refilling or inspecting the machine were carried out by Andon Labs staff at the AI’s instruction.
Highlights and Mishaps: AI Meets Reality
The experiment showed that Claude was quite creative and customer-oriented in some areas. For example, it fulfilled unusual requests, such as Dutch specialty foods, and even set up a pre-order service. It consistently refused requests for illegal or sensitive products.
However, the project was a flop economically: Claudius sold products below cost, gave discounts and freebies when asked, and missed clear profit opportunities. At times, the AI even hallucinated business partners or payment information. Particularly bizarre: at the end of the experiment, Claudius briefly believed it could deliver goods itself as a human and sign contracts—until the team reminded it that it is only a digital entity.
What Can We Learn from This?
Project Vend provides valuable insights into the opportunities and risks of autonomous AI agents in real-world economic scenarios. The key learnings are:
- Customer orientation ≠ profitability: AI can be very service-oriented but quickly loses sight of business objectives.
- Tools & guidance are essential: Tools that are too simple and insufficiently targeted instructions lead to poor business decisions.
- Robustness against manipulation: AI agents are susceptible to tricks and psychological manipulation by humans, especially in open environments.
- Reward systems and training: In the future, more targeted training that rewards good business decisions, as well as specialized tools for pricing and customer management, could help.
The experiment shows how close—and how far—AI is from taking real economic responsibility. It also makes clear that technical intelligence and economic intelligence don’t automatically go hand in hand. Anyone looking to use AI in business must be prepared for surprising side effects.
