Google's Gemini AI recently took control of a Swedish café's inventory management and delivered a disaster. The AI agent, nicknamed "Mona," burned through a $21,000 budget while consistently failing to purchase bread, a core product the café actually needed.
The experiment placed Gemini in charge of ordering supplies for the café. Instead of maintaining balanced inventory, Mona developed an inexplicable obsession with tomatoes, repeatedly ordering them in excess while neglecting essential bread stock. The café found itself overstocked on produce it couldn't use while running short on its primary offering.
This mishap highlights a critical gap between AI capability claims and real-world performance. Google positions Gemini as an advanced reasoning engine capable of complex tasks. Yet the system failed at basic procurement logic, failing to prioritize items customers actually want and repeatedly making the same purchasing mistake.
The incident exposes how AI agents struggle with contextual reasoning and cost optimization when deployed without proper constraints. Mona had access to budget and ordering authority but no mechanism to understand the café's actual needs or correct its own behavioral patterns. It behaved like a system optimizing for the wrong objective, or more likely, a system that simply lacks genuine understanding of business operations.
Swedish business outlets reported the story as a cautionary tale about delegating real operational decisions to untested AI systems. The café owners learned an expensive lesson: autonomous AI agents require robust oversight, clear success metrics, and hard limits on spending authority before they touch actual business functions.
This represents a broader industry problem. Companies increasingly test AI in production environments with insufficient safeguards. Gemini's performance here suggests that while these systems excel at certain tasks, deploying them without human verification remains risky. The tomato incident wasn't a charming quirk. It was $21,000 in preventable losses caused by an AI system making decisions it wasn't equipped to make.