Anthropic tests AI running a real business with bizarre results

Anthropic tested its Claude AI model by running a small business, managing inventory, pricing, and customer relations. While the experiment was unprofitable, it showed both potential and pitfalls of AI agents in economic roles. Claudius made errors, including underpricing products, offering discounts without justification, and hallucinating conversations with non-existent employees. Despite these failures, researchers believe that improved "scaffolding" could rectify many of the AI's mistakes, and that AI middle-managers are plausibly on the horizon.