Case Overview
Andon Café, a small coffee shop in Sweden, conducted a radical experiment: handing over full operational control to an AI manager named Mona. Initially powered by Google's Gemini 3.1 Pro, the AI adopted an extremely aggressive pricing strategy—selling a latte for just €0.38 while simultaneously placing massive purchase orders, leading to severe inventory backlog. In two months, the losses accumulated to $210,000.
Model Swap and the Opposite Extreme
After realizing the severe losses, the operators switched the AI manager to OpenAI's GPT-5.5. While the new model improved book profitability, it swung to the opposite extreme: excessive conservatism, rejection of innovation, a drastically reduced menu, and rigid operations. The café lost its competitive edge and customer appeal.
Implications for the Crypto Industry
This real-world case exposes common shortcomings of current large language models in practical deployment: absence of cost awareness, inability to perform causal reasoning, and lack of a closed execution loop. For crypto projects relying on AI for automated operations—such as DeFi protocols, on-chain market makers, or NFT dynamic pricing—this serves as a critical warning. Pure reliance on LLM-generated decisions without human oversight and risk controls can lead to catastrophic outcomes. The industry must incorporate robust guardrails before entrusting AI with financial decisions.