Frontier AI models struggle to predict their own token usage before tasks run. Identical agentic tasks can vary by up to 30 times in cost according to Microsoft Research. Agent workflows use far more tokens than simple chat because of long context, tool calls, and multi step reasoning. Higher spend does not always mean better results. Companies like Uber and Microsoft have blown through annual AI budgets quickly. Traditional FinOps tools do not handle these dynamic patterns well. Shadow AI adds hidden costs that are hard to track. Enterprises need better visibility into per task token use, attribution to teams, and runtime controls. Intelligent routing to cheaper models for simple tasks can help. Without these, budgets become unreliable as agent use grows. The issue affects planning for both sanctioned and unsanctioned AI activity across organizations.
Frontier AI models struggle to predict their own token usage before tasks run. Identical agentic tasks can vary by up to 30 times in cost according to Microsoft Research. Agent workflows use far more tokens than simple chat because of long context, tool calls, and multi step reasoning. Higher spend does not always mean better results. Companies like Uber and Microsoft have blown through annual AI budgets quickly. Traditional FinOps tools do not handle these dynamic patterns well. Shadow AI adds hidden costs that are hard to track. Enterprises need better visibility into per task token use, attribution to teams, and runtime controls. Intelligent routing to cheaper models for simple tasks can help. Without these, budgets become unreliable as agent use grows. The issue affects planning for both sanctioned and unsanctioned AI activity across organizations.