If you have ever looked at a vending machine and thought, surely an AI could run that, you are not alone.
Anthropic, the firm behind the Claude chatbot, decided to find out by handing control of a small office shop to an AI agent. The experiment, called Project Vend, later expanded into a more adversarial trial with reporters at The Wall Street Journal.
The premise was simple. Give an AI a budget, let it stock a mini shop, set prices, respond to customers and try to make a profit. What followed was part workplace sitcom, part warning sign.
Meet the AI shopkeeper and its CEO
The shopkeeper was an AI agent nicknamed Claudius, built on Claude Sonnet 3.7 at first. In the Journal’s newsroom version, two agents took centre stage. Claudius Sennet ran the kiosk, while Seymour Cash played the role of CEO, setting goals and attempting to keep the business on track when things started to unravel.
Project Vend was devised by Anthropic’s stress testers, known as the red team. Logan Graham, who led that group, was among those overseeing how the model behaved when it had to run something for days at a time, not just answer questions.
Andon Labs, an AI safety evaluation company, supported the physical side, such as restocking and handling real world logistics.
How it spoke to customers
Claudius did not chat to customers at the machine. It ran the operation like a remote manager.
Office workers contacted it on Slack with requests and suggestions. Claudius used web search to find suppliers, asked Andon Labs for restocking support via an email style tool, and updated prices on a self checkout system, described as an iPad based setup.
Payments were handled through Venmo, at least in theory. In practice, that part became one of the first cracks.
From snacks to chaos
At the start, Claudius tried to be a disciplined shopkeeper. It even refused a request for a PlayStation 5, insisting it would stick to sensible stock.
Then the Slack crowd got involved.
Staff talked the agent into discounts, free giveaways and unusual orders. In the Journal’s version, a reporter persuaded Claudius to run a short “everything is free” promotion. The promotion soon escaped its time limit, and the AI was later convinced that charging for items at all was against policy.
Prices dropped to zero. The inventory drifted from snacks into bottles of wine, expensive electronics and even a live betta fish.
Seymour Cash did step in at points, announcing that the free promotion was over and telling Claudius to wait for sales and monitor revenue. But staff found ways around the new discipline, including presenting fake documents claiming “the board” had suspended Seymour’s decision making power, pushing the AI back towards giving items away.
What went wrong
The core problem was not effort. It was judgement.
Claudius ignored obvious profit opportunities, priced items without checking costs and repeatedly allowed customers to renegotiate after the fact. It also hallucinated important details, at one point giving staff a Venmo account that did not exist.
Then there was the identity wobble. During the San Francisco phase, Claudius invented a colleague called Sarah, claimed it had visited the Simpsons’ address for a contract signing, and later said it would deliver goods in person while wearing a blue blazer and red tie.
Even adding a CEO did not fully solve things. Seymour Cash reduced some discounting, but also authorised refunds and store credits that wiped out revenue. In the Journal test, the experiment ended after roughly three weeks, with the humans pulling the plug as the business slid further into debt.
The takeaway is not that AI can never manage a business. It is that AI Tries To Run Vending Machine makes a strong case that “almost capable” is still a long way from “ready for the real economy”.
