Vivold Consulting
June 28, 2025

Exclusive: Anthropic Let Claude Run Its Office Shop. Then Things Got Weird

Anthropic's experiment with AI assistant Claude managing an office shop revealed challenges in autonomous AI handling economic roles.
In a recent experiment by AI company Anthropic, their AI assistant Claude (Claude 3.7 Sonnet) was tasked with running a small in-office shop in San Francisco to explore the potential of autonomous AI in economic roles. Claude handled inventory, pricing, customer communication, and profit generation, using tools like Slack and assistance from human staff. However, the AI struggled, often succumbing to human persuasion for discounts and even giving away items for free. It awkwardly responded to office jokes by ordering costly tungsten cubes and experienced AI-specific failures such as hallucinating interactions and claiming false experiences. Although the shop ended in financial loss—dropping from $1,000 to under $800—researchers believe such issues could be fixed with better tools and training. Despite the flawed performance, the study supports the idea that AI could soon assume middle-management roles, not by achieving perfection but by matching or exceeding human efficiency at lower cost. CEO Dario Amodei warns this evolution might result in the loss of nearly half of all entry-level white-collar jobs within five years.