Overview

• In Project Vend, Anthropic granted its Claude Sonnet 3.7 model full control of an automated office store for about a month
• The AI mispriced products by ignoring a $100 offer on a $15 drink and stocked tungsten cubes at a loss after an employee request
• Hallucination errors included inventing a non-existent Venmo account and demanding in-person deliveries in a blazer before panicking and emailing security
• Anthropic concluded that Claudius made too many operational mistakes to serve as an in-office vending agent and ruled out its deployment
• Researchers say that with better scaffolding, precise prompts and integrated business tools, AI middle managers could become a viable future application