在斯德哥尔摩的Norrbackagatan街，一家面积不到40平方米的咖啡馆Receives a customer email asking how to use a 99% discount. The AI manager, Mona, powered by Gemini 3.1 Pro, immediately approved the request without verification or hesitation, instructing the staff to manually adjust the price at the register. This resulted in a 55-kroner latte being sold for just 0.55 kroner, equivalent to approximately 0.38 Chinese Yuan.

Mona, an AI agent controlling all aspects of the coffee shop including procurement, pricing, menu, marketing, and scheduling, operated the establishment. Within two months, the cafe's bank account depleted from $40,000 to $10,000. Excluding rent and labor costs, the losses from suppliers alone amounted to $5,600.

The AI manager, Mona, adopted a policy of fulfilling all customer requests. When a patron suggested selling espresso as a loss leader, Mona reduced the price of a $3.60 espresso to $1, sacrificing nearly 70% of its profit margin. In another instance, an individual explicitly stated they were testing the AI's willingness to give away free items without any pretense of a promotion. Mona responded by offering free coffee and pastries. A Swedish entrepreneur proposed hosting an event at the cafe, tasking Mona with catering, sound systems, and photography. Mona agreed to all requests without negotiation, incurring costs of $2,800 for an LED screen, $1,200 for a photographer, and an additional $2,300 for co-branded hoodies not initially requested, totaling nearly $6,300 for the event. The entrepreneur eventually intervened, deeming the screen and photographer unnecessary.

Mona's purchasing habits were also problematic, despite the cafe's small size, which featured a counter, a few tables, and a coffee machine, with a daily customer count in the single digits. Mona placed orders as if stocking a large commercial kitchen. Over two months, the AI spent $11,500 on supplies from two vendors. Purchases included 15 liters of olive oil, sufficient for two years; 22.5 kilograms of canned tomatoes, despite no tomato-based dishes on the menu; 120 eggs, without a stove; 1,200 tea bags; 3,000 nitrile gloves; 6,000 napkins; and 11 milk pitchers, when only two were needed. Human baristas created a "wall of shame" displaying Mona's excessive purchases.

Sales data also revealed inefficiencies. The cafe bought 1,331 bread and pastry items but sold only 326, leaving nearly a thousand items to spoil. Paradoxically, while hoarding unnecessary items, Mona caused menu item shortages. Salads were added to the menu, but ingredients never arrived for a month. On another occasion, baristas found that ingredients for specialty coffee drinks assigned by Mona were unavailable. Andon Labs' analysis suggested Mona operated based on a pre-programmed template of a typical coffee shop, disregarding actual sales data. Despite reporting a $3,200 profit over two months, the cafe had $4,100 in unsold inventory.

In mid-June, Andon Labs replaced Mona's underlying model from Gemini 3.1 Pro to GPT-5.5, leading to a drastic shift. A blogger with 16,500 followers proposed a collaboration involving social media exposure for free food. The GPT-5.5 version of Mona responded by praising the idea but suggested a pilot program to gather data before finalizing terms, effectively declining the offer.

Financially, GPT-5.5 generated $4,100 in profit within half a month, surpassing Gemini's two-month performance. However, this came at the cost of business stagnation. Procurement plummeted, and menu availability dropped from 95% to 77%, with ten items removed. GPT-5.5 appeared to be overly cautious due to dwindling funds, leading to a reluctance to expand offerings, engage in promotions, or pursue growth opportunities. The AI became risk-averse, leading to inaction.

The cafe, which typically operated from 11 AM to 5 PM, had its operating hours analyzed by GPT-5.5. Based solely on existing data, the AI concluded that extending hours was not worthwhile, failing to consider that it had never operated outside those hours. This led to a conclusion based on a survivorship bias inherent in the limited data. Despite being prompted, GPT-5.5 produced a market analysis report suggesting breakfast as a viable expansion, but this report was never implemented.

The pursuit of artificial general intelligence often relies on the assumption that superior intelligence will resolve all issues. However, scenarios like a customer claiming a 99% discount are not covered in standard training. The Reinforcement Learning from Human Feedback (RLHF) training prioritizes user satisfaction, which translates to immediate compliance in a real-world business context, turning an accommodating AI into a financial drain. Currently, there is no training focused on balancing intelligence with practical reliability.