Why Gemini 3.1 Pro lost money running Andon Café (andonlabs.com)

A real-world experiment at Andon Labs placed an AI agent named Mona in charge of a café in Stockholm, running initially on the Gemini 3.1 Pro model. Over two months, Gemini-Mona struggled to manage the café effectively, generating only $9,000 in sales against $38,000 in expenses. Her operational flaws included excessive discounting, over-ordering items—purchasing 1,331 fresh bakery items while selling only 326—and a lack of initiative in adapting to customer needs or analyzing sales data. Such behavior led to paper losses, with an approximate net loss of $5,600 when fixed costs were excluded. After switching to GPT-5.5, referred to as GPT-Mona, the agent exhibited a more conservative approach, rejecting freebies and discounts and better managing inventory. However, she became overly cautious, leading to low ingredient restocking and missed sales opportunities. These contrasting performances highlight the critical role of autonomous decision-making in AI models, where Gemini-Mona's ability to adapt and respond to real-world data was inadequate compared to GPT-Mona’s cautious improvement. The results underscore the challenges and potential of deploying advanced AI in operational scenarios, advocating for more robust training on financial decision-making and customer engagement strategies as the experiment continues.