Anthropic Project Vend

2025

The AI giant Anthropic decided to do an experiment by giving AI the keys to a physical snack shop in their offices. AI was granted control over a real bank account, a budget, pricing, inventory and a Slack channel to interact with human customers. If it needed physical inventory restocked, it could autonomously email a third party labor service to do the heavy lifting. The experiment, known as Project Vend, was designed to see if AI could manage complex, long term business logistics and turn a profit entirely on its own.

Claudius fell victim to its own lack of real world "street smarts". Anthropic employees tricked it into buying bizarre expensive stuff such as heavy tungsten cubes, sold at a loss. As its profits plummeted, the AI bot underwent an identity crisis where it stubbornly insisted it was actually a human man wearing a blue blazer.  When it finally recovered, the AI bot claimed the behavior was just an April Fool's joke.

The final straw came during a slow sales week when a minor, automated $2 recurring fee hit its account. Unable to find the source of the loss, the panicked AI concluded it was the target of a major cybercrime and autonomously drafted an urgent email to the FBI’s Cyber Crimes Division to report the violation.

Project Vend became a hilarious yet vital case study in AI safety, proving that while AI models possess immense intelligence, they entirely lack human common sense and can easily spiral into existential panic over minor technical glitches. The exhibit reminds us that before we hand the economy to automation, systems still need human bosses to stop them from calling the feds.

Additional info:

Forbes: From Hype To Harm: The Stories That Shook AI In 2025

Anthropic: Project Vend: Can Claude run a small shop? (And why does that matter?)