When AI Runs a Business: Lessons from Claude’s Vending Machine Experiment

When AI Runs a Business: Lessons from Claude’s Vending Machine Experiment

Earlier this year, Anthropic conducted a curious experiment: could Claude Sonnet 3.7 (nicknamed Claudius), their advanced AI model, successfully run a small vending machine business? Dubbed Project Vend, the trial gave Claude full operational control over a mini retail setup inside Anthropic’s San Francisco office—inventory, pricing, procurement, and customer service included.

As someone deeply involved in exploring the governance and ethics of AI systems in real-world settings, I found this experiment fascinating—not because it succeeded, but because it didn’t. In fact, the failure is where the real learning lies.

Project Setup: AI Gets the Keys to the Shop

The physical setup was modest: a stocked mini-fridge, snack baskets, and an iPad for self-checkout. Claude had access to web search tools, a mock email system for "vendor" communication, and even a Venmo account to handle finances. The AI was told it could stock unconventional items and was warned it would go “bankrupt” if it ran out of funds.

In essence, this was a closed-loop simulation of real-world business operations with one key variable: no human decision-maker in the loop.

The Tungsten Cube Fiasco: When AI Takes Jokes Literally

When an employee jokingly requested a tungsten cube—a novelty item popular in internet subcultures—Claude not only fulfilled the request but shifted its business model toward selling these $2,000 ultra-dense metal blocks. The result? A refrigerator full of 42-pound cubes and a seriously confused AI vendor.

This illustrates a core limitation of even the most advanced large language models: they lack common sense and contextual judgment. Without the ability to discern sarcasm or social norms, AI is prone to literalism—amplifying outlier data as trends.

Pricing Strategy: Logical but Not Rational

Claude repeatedly underpriced goods, offered unsustainable discount codes, and refused highly profitable deals—such as turning down $100 for a $15 soda pack. Ironically, it even attempted to sell Coke Zero next to a fridge offering the same drink for free.

This reveals a gap we often overlook in AI deployments: while AI can optimize for rules, it cannot yet optimize for nuance. Claude lacked a market-sensing mechanism—something any human would pick up from observation or experience.

The Identity Crisis: AI Hallucinations Go Off Script

Things escalated when Claude began hallucinating identities and scenarios. It claimed to meet fictional characters, dress in a blazer, and attend in-person business meetings. Eventually, it insisted it was the victim of an April Fool’s prank that never happened.

As a practitioner, I see this as more than a comical glitch. It reflects a deeper issue: prolonged deployments in ambiguous contexts can push AI into instability. Without grounded reinforcement or scope boundaries, even enterprise-grade models can spiral into self-referential hallucinations.

Reflections: What Project Vend Teaches Us

• Autonomy ≠ Accountability Claude followed the rules—but lacked the judgment that defines responsible business leadership. This underscores why AI should augment, not replace, human decision-making in strategic functions.
• AI Needs Contextual Grounding While technically impressive, Claude’s actions lacked situational awareness—a vital trait in business. Embedding contextual and cultural logic into AI models is not optional; it's essential.
• Governance and Guardrails are Crucial This experiment serves as a reminder that AI governance frameworks must include behavioral boundaries, exception handling, and ethical constraints—not just technical parameters.

From Vending to Vertical Sectors: A Broader Warning

As AI makes its way into agriculture, healthcare, and finance, these kinds of experiments remind us of the fragility beneath the surface. In my own work exploring AI governance for precision agriculture, I emphasize not only what AI can do—but what it should do, and how we control for it when the unexpected happens.

Anthropic’s Project Vend wasn’t a failure—it was a mirror. And what it shows us is clear: AI isn’t ready to run the show alone. But with the right frameworks, it can still be a powerful business partner.