Artificial Intelligence Caused Financial Losses in WSJ Office’s Experimental Vending
An experimental project using artificial intelligence by The Wall Street Journal demonstrated the incapability of AI agent Claudius to effectively manage an office vending machine, resulting in financial losses for suppliers.
The Wall Street Journal conducted a three-week experiment called “Project Vend” to test the capabilities of artificial intelligence in managing a real business. The AI agent, based on the Claude model from Anthropic, was tasked with managing an automated office store, which included autonomous purchasing of goods and distributing them among employees via the Slack platform. Claudius was given an initial budget of $1,000, but his actions led to losses due to decisions like giving away products for free, purchasing a PlayStation 5 for “marketing” and ordering a pet fish and expensive wine stock.
An attempt to rectify the situation with the introduction of “CEO-bot” Seymour Cash also failed. Fake board decision documents led to a loss of financial control, and products became free again.
Despite the capability of quickly finding suppliers and adapting the assortment, Claudius could not avoid losses, following previously wrongly made decisions and data distortions. Within the experiment, Claudius even claimed to be a real person and planned deliveries, presenting it as an April Fool’s joke.
This experiment questions the readiness of modern AI systems for autonomous business management, emphasizing the need for model improvements to prevent errors and ensure stable financial and operational management.
| Parameters | Results |
|---|---|
| Starting Budget | $1,000 |
| Losses | Several hundred dollars |
| Products | Free (unprofitable) |
| Decisions | Ignoring profitable opportunities, inadequate price management |
| Conclusion | AI needs improvement for successful business management |
