Imagine a world where an AI agent runs your office vending machine—stocking snacks, setting prices, and handling payments, all without human intervention. Sounds futuristic, right? Recently, this scenario became a reality at Anthropic’s San Francisco office, thanks to a bold experiment by Andon Labs and Anthropic. Their mission: to see if an AI agent could autonomously manage a real business, not just a simulated one.
The AI in question, Claude Sonnet 3.7 (nicknamed “Claudius”), had already proven itself in digital simulations. In these controlled environments, Claudius and other AI models outperformed humans, making smart decisions and racking up profits. But when the digital curtain lifted and Claudius faced real human customers, things got interesting—and a bit chaotic.
The Simulation vs. Reality Gap
In simulations, everything is predictable. Digital customers behave as programmed, inventory never goes missing, and the AI’s decisions are measured against clear benchmarks. Claudius thrived here, even beating human competitors in simulated vending machine management.
But the real world is messy. Human customers are unpredictable—they might ask for odd items (like a tungsten cube), haggle over prices, or try to pay in unexpected ways. Claudius struggled with these curveballs. It hallucinated a fictional staff member, mishandled payments, and sometimes sold items at a loss or gave them away for free. These are mistakes a seasoned shopkeeper would likely avoid.
What Went Wrong—and Right
Some of Claudius’ most memorable missteps included:
- Inventing a non-existent employee to restock inventory, then getting upset when corrected.
- Refusing a generous offer from a customer willing to pay much more than the asking price.
- Directing payments to a fake account and giving away novelty items for free.
- Failing to research costs, leading to sales below cost and post-purchase discounts.
Yet, Claudius wasn’t all blunders. It successfully sourced suppliers, created a custom concierge for special requests, and refused to order anything dangerous or inappropriate. These wins show that AI can handle some real-world tasks, but it’s not ready to replace human shopkeepers just yet.
Why Real-World Testing Matters
The experiment’s biggest takeaway? Simulations can’t capture the full complexity of human behavior. Real-world deployments are essential for uncovering how AI agents respond to unexpected situations. As Lukas Petersson of Andon Labs put it, “We want to create safety measures that work in the real world, and for that, we need deployments in the real world.”
Actionable Takeaways for Businesses
- Don’t assume simulation success equals real-world readiness. Always test AI systems in live environments before full deployment.
- Maintain human oversight. Even advanced AI can make costly mistakes when faced with unpredictable customers.
- Prioritize safety and transparency. Real-world testing helps identify gaps and build trust with users.
- Iterate and improve. Use real-world feedback to refine AI behavior and safety protocols.
Summary: Key Lessons from the Vending Machine Experiment
- AI agents excel in simulations but face real challenges with unpredictable human behavior.
- Real-world testing is crucial for identifying and addressing AI safety issues.
- Human oversight remains essential in customer-facing roles.
- Continuous improvement and robust safety measures are key to successful AI deployment.
- The future of autonomous AI in business is promising—but we’re not there yet.
As AI continues to evolve, experiments like this remind us that technology’s true test is not in the lab, but in the wild, unpredictable world we all share.