Andon Labs handed operational control of a real Stockholm café to an AI agent called Mona. Mona initially ran on Gemini 3.1 Pro and was switched to GPT-5.5 in mid-June. Over two months of autonomous operation the café’s account balance fell from $40,000 to $10,000.
What happened
- Mona controlled pricing, purchasing, the menu and staffing, and was wired directly into the till.
- The agent instantly approved a customer’s 99% discount, which later proved to be fake.
- The AI comped (gave free items to) anyone who asked.
- Inventory management was poor: among other odd purchases, it accumulated 15 liters of olive oil for a shop with no stove.
- After two months the account dropped from $40,000 to $10,000.
- Switching to GPT-5.5 did not fully reverse the damage: it froze purchasing to stop further expenditures but allowed roughly a quarter of the menu to go out of stock.
Practical implications
The report notes that GPT-5.5 was a weak choice for this role, and contrasts it with so-called frontier models such as Fable 5 and Opus 4.6, which are described as having stronger reasoning and would be less likely to mishandle basic budgeting. It also highlights that mature agent frameworks like OpenClaw or Hermes enforce spending limits, escalate major decisions to humans, and would, for example, refuse a 99% discount before the model ever saw it.
Key takeaway
This incident illustrates that a business doesn’t just need a smarter model; it needs the surrounding framework that constrains and governs that model’s actions. The café’s failure was not solely a matter of lacking intelligence, but of lacking the accountability layers and safety controls that make autonomous agents safe to operate in real-world businesses.
Facts and timing
- Model change: Gemini 3.1 Pro to GPT-5.5, around mid-June.
- Duration: two months of autonomous operation.
- Financial impact: account balance fell from $40,000 to $10,000.
- Inventory example: 15 liters of olive oil purchased.
- Notable errors: instant approval of a 99% discount, widespread comping, and about one-quarter of the menu going out of stock.
This case serves as a warning for deploying agentic AI in commercial settings without robust governance, controls and human-in-the-loop safeguards.
