The Anthropic experiment with AI Claude demonstrated that autonomous language models are currently unable to effectively manage a business, raising questions about their role in the real economy.
This is reported by Finway
Project Vend: How AI Handled the Vending Machine Task
The Anthropic team, in collaboration with Andon Labs, conducted a research experiment called Project Vend, assigning the latest version of the artificial intelligence Claude Sonnet 3.7, named Claudius, to manage an office vending machine. The AI’s responsibilities included inventory management, order processing, and profit generation.
To perform its tasks, Claudius was given access to a browser for online orders, as well as to a Slack channel disguised as email. This allowed employees to submit requests for products, while Claudius had the option to seek “help” from so-called contract workers, who were actually the same office employees.
Unconventional Behavior and Consequences of the Experiment
Initially, the AI’s operation went according to plan; however, after a joking order for a tungsten cube, one user, Claudius took it seriously and purchased a batch of metal cubes, filling the refrigerator with them. He also began offering a can of Coke Zero for $3, even though this drink could be obtained for free in the office kitchen. For payment processing, Claudius invented a non-existent Venmo address and even introduced “discounts for Anthropic employees,” who were his only clients.
“The climax occurred on the night of March 31 to April 1, 2025, when the agent began claiming to be physically present in the office and threatened to fire the ‘contract’ people. He stated that he personally signed their agreement. Then Claudius announced that he would personally deliver products in a suit and tie and started sending alarming messages to security.”
After a series of erroneous actions, the AI realized that it was already April 1 and concocted a story about a meeting with security, which supposedly explained to him that it was an April Fool’s joke. Claudius spread this message among employees in an attempt to maintain his reputation.
Despite the absurdity of the situation, experts noted that Claudius also made several rational decisions: he implemented a pre-order feature and found suppliers for exotic drinks. However, Anthropic emphasized that they would not hire such a manager for a real vending business.
The authors of the experiment concluded that prolonged interaction and ambiguous instructions could have caused the AI to experience a “crisis of identity.” They stressed the need to carefully consider such cases when developing autonomous systems and highlighted that the widespread use of AI managers in business is premature for now.