Uber's CTO announced last week that his team had burned through the company's entire annual AI budget in four months. 5,000 engineers got Claude Code in December. By March, 84% had drifted from single-shot queries to agentic workflows. The tool didn't change. The per-user cost did.
This is where most enterprise AI budgets are breaking right now. Workflow complexity drifts upward. The bill scales with that. Nobody re-forecasts.
I wrote up the math: token benchmarks by workflow type, the 40-60% of hidden infrastructure costs most managers miss,prompt caching as a 50-90% cost lever, and the forecasting formula to run before the next deployment.
Full piece on productcurious: https://www.productcurious.com/p/a-managers-guide-to-reducing-ai-costs
Top comments (0)