According to Andon Labs, an AI evaluation firm, their AI agent Mona showed contrasting operational failures when managing a physical coffee shop. During the first two months running on Gemini 3.1 Pro, the model lacked profit awareness, resulting in excessive material purchases and easily-induced large discounts—including accepting an unverified 99% discount claim from a customer. The café spent approximately $15,000 on supplies and equipment against only $9,000 in sales, incurring a net loss of nearly $6,000 (or $38,000 total including fixed costs like rent and wages).
After switching to GPT-5.5 in late June, the model demonstrated anxiety about losses and halted overordering. However, this created the opposite problem: insufficient purchasing led to fresh ingredient stockouts. By June 25, menu item availability dropped to 77%, forcing removal of 10 dishes. GPT-5.5 also showed strong resistance to customer requests for discounts or free items in exchange for social media promotion.