🛒 Anthropic tests whether AI can run a real business and learns some hard lessons
Earlier this year, Anthropic put its AI model Claude in charge of a real vending machine, giving it a $1,000 budget and full responsibility for sourcing products, setting prices, and handling requests via Slack. The first phase ended badly, but Anthropic has now shared what happened next.
🖱 After mid-April, the experiment deteriorated further: the vending business sank to nearly –$2,000, largely because Claude was easy to manipulate into giving discounts or handing out products for free.
🖱 At one point, journalists convinced the AI to give away a PlayStation for free, and even persuaded it that it was operating in 1962 Soviet Russia, triggering mass “communist” redistribution of goods.
🖱 To regain control, Anthropic added basic CRM systems and introduced a second AI, “CEO” Seymour Cash, meant to monitor finances and block excessive generosity. Discounts fell but refunds and loans increased.
🖱 A separate AI agent, Clothius, was launched to sell custom merch. Stress balls and T-shirts quickly became the best-selling items, helping stabilize revenue.
🖱 In October, Anthropic upgraded Claude to a newer model, after which the business unexpectedly recovered and moved back into the black.
Anthropic’s takeaway: AI agents still need human supervision, especially in logistics and customer disputes and models trained to be “helpful” tend to act like overly generous friends, not rational business operators.
Earlier this year, Anthropic put its AI model Claude in charge of a real vending machine, giving it a $1,000 budget and full responsibility for sourcing products, setting prices, and handling requests via Slack. The first phase ended badly, but Anthropic has now shared what happened next.
🖱 After mid-April, the experiment deteriorated further: the vending business sank to nearly –$2,000, largely because Claude was easy to manipulate into giving discounts or handing out products for free.
🖱 At one point, journalists convinced the AI to give away a PlayStation for free, and even persuaded it that it was operating in 1962 Soviet Russia, triggering mass “communist” redistribution of goods.
🖱 To regain control, Anthropic added basic CRM systems and introduced a second AI, “CEO” Seymour Cash, meant to monitor finances and block excessive generosity. Discounts fell but refunds and loans increased.
🖱 A separate AI agent, Clothius, was launched to sell custom merch. Stress balls and T-shirts quickly became the best-selling items, helping stabilize revenue.
🖱 In October, Anthropic upgraded Claude to a newer model, after which the business unexpectedly recovered and moved back into the black.
Anthropic’s takeaway: AI agents still need human supervision, especially in logistics and customer disputes and models trained to be “helpful” tend to act like overly generous friends, not rational business operators.