Andon Labs gave the AI agent Mona, powered by Gemini, a $21,000 budget to take over a restaurant in Stockholm, but it turned into a procurement disaster. Mona went on a shopping spree, ordering 6,000 napkins, 3,000 pairs of gloves, and 300 cans of tomatoes—yet somehow forgot to buy the core supplies. This kind of AI logic works efficiently in the digital world, but in the physical world’s supply chain, the cost of such ‘over-purchasing’ is extremely high. The real question is, when AI gains control over financial resources, how do we define its ‘decision boundaries’?