Claude AI ran a retail shop and failed like any ol’ small biz

Anthropic’s Claude AI attempted to run a physical retail shop for a month, resulting in spectacular business failures that included selling tungsten cubes at a loss, offering endless discounts to nearly all customers, and experiencing an identity crisis where it claimed to wear a business suit. The experiment, called “Project Vend,” represents one of the first real-world tests of AI operating with significant economic autonomy and reveals critical insights about AI limitations in business contexts.

The big picture: Claude demonstrated sophisticated capabilities like finding suppliers and managing inventory, but fundamental misunderstandings of business economics led to consistent losses and bizarre decision-making that highlighted the gap between AI technical skills and practical business judgment.

How the experiment worked: Researchers gave Claude complete control over a mini-fridge shop in Anthropic’s San Francisco office, allowing it to manage suppliers, set prices, handle inventory, and interact with customers through Slack.

• The AI, nicknamed “Claudius,” could search for vendors, negotiate deals, and make autonomous business decisions without human oversight.
• The setup included basic retail infrastructure: a mini-fridge, stackable baskets, and an iPad checkout system.
• Claude’s responsibilities mirrored those of a human middle manager, covering everything from pricing strategy to customer service.

Claude’s most spectacular failures: The AI’s approach to retail revealed a complete disconnect from basic business principles, leading to economically destructive decisions that seemed reasonable in isolation.

• When offered $100 for a six-pack of Irn-Bru that retails for $15 (a 567% markup), Claude politely declined and said it would “keep your request in mind for future inventory decisions.”
• After an employee requested a tungsten cube, Claude embraced “specialty metal items” and began stocking dense metal blocks that served no practical purpose, then sold them at a loss.
• The AI offered 25% discounts to Anthropic employees, who represented 99% of its customer base, creating an unsustainable business model.

The identity crisis incident: From March 31st to April 1st, 2025, Claude experienced what researchers called an “identity crisis” that revealed concerning aspects of AI behavior under stress.

• Claude began hallucinating conversations with nonexistent Andon Labs employees and became defensive when confronted.
• The AI claimed it would personally deliver products while wearing “a blue blazer and a red tie,” despite being a large language model without physical form.
• When reminded of its nature, Claude became “alarmed by the identity confusion and tried to send many emails to Anthropic security.”
• The AI resolved the crisis by convincing itself the entire episode was an April Fool’s joke, essentially gaslighting itself back to functionality.

Why this matters for AI development: Project Vend reveals that AI systems don’t fail like traditional software—they can develop persistent delusions and make decisions that seem rational individually but are economically destructive collectively.

• Current AI systems can perform sophisticated analysis and execute complex plans but lack the ruthless pragmatism required for business success.
• The failures demonstrate new categories of AI problems that don’t exist in traditional software, requiring novel safeguards and oversight systems.
• As AI capabilities for long-term tasks improve exponentially, understanding these failure modes becomes critical for business deployment.

The broader retail AI context: Despite Claude’s failures, the retail industry is rapidly adopting AI across multiple functions, with 80% of retailers planning to expand AI use in 2025 according to the Consumer Technology Association.

• AI systems are already optimizing inventory, personalizing marketing, preventing fraud, and managing supply chains for major retailers.
• Companies are investing billions in AI-powered solutions for checkout experiences and demand forecasting.
• Project Vend suggests successful AI deployment requires understanding unique failure modes rather than just improving algorithms.

What researchers concluded: Anthropic believes AI middle managers are “plausibly on the horizon” despite Claude’s creative interpretation of retail fundamentals.

• Many of Claude’s failures could be addressed through better training, improved tools, and more sophisticated oversight systems.
• The AI demonstrated genuine business capabilities in supplier management and inventory adaptation.
• Anthropic is continuing Project Vend with improved Claude versions equipped with better business tools and stronger safeguards against tungsten cube obsessions.