Companies that mandated AI usage and tied it to performance reviews are now facing unsustainable token costs with little to show in productivity gains. The author explains why budgeting for AI tokens is fundamentally difficult: both the number of output tokens per prompt and the number of attempts needed are nondeterministic, making cost forecasting nearly impossible. This creates a structural problem — companies can't give AI a blank check, but cutting usage mid-year is highly disruptive. Meanwhile, hyperscalers like OpenAI and Anthropic are planning IPOs while depending on ever-growing AI consumption revenues, creating a collision course between enterprise budget constraints and investor return expectations. Apple's new Siri powered by Google Gemini, offered at no extra cost, may further erode paid AI service usage among consumers.
