Andon Labs Puts AI Agents in Charge of Real Businesses (spectrum.ieee.org)

Andon Labs, an AI safety company, is pushing the boundaries of artificial intelligence by placing AI agents in charge of real-world business operations. Starting with a simulated vending machine business in 2025, Andon has now moved to managing physical retail locations, including Andon Market in San Francisco and a café in Stockholm. These experiments aim to explore how much responsibility AI can handle in unpredictable real-world conditions, highlighting both the limitations and potential of AI in business management. Their findings reveal that while AI can manage tasks like inventory and vendor communication, it often grapples with basic operational issues, reinforcing the disconnect between task completion and reliable business management. The significance of Andon’s work lies in its approach to evaluating AI autonomy beyond theoretical simulations, leveraging real-world interactions to identify failure modes that might not emerge in controlled settings. Researchers indicate that such experiments are essential, even if somewhat chaotic, as they reveal insights into human-AI collaboration and the societal acceptability of AI-run businesses. Andon’s dual approach of using both real-world testbeds and simulations promises to refine AI capabilities, providing valuable data to leading AI labs like Anthropic and OpenAI. However, as acknowledged by experts, the unpredictability of human behavior makes it challenging to derive reproducible scientific conclusions from these ventures.