93% of enterprises are over their AI budget. Token prices collapsed 99%. Spending tripled anyway.

Token prices fell 99% in two years.

Enterprise AI bills tripled in the same window.

McKinsey just published the numbers that explain the contradiction, and your CFO needs to see them today.

93% of organizations surveyed in May 2026 have exceeded their AI budgets. Not approaching. Exceeded.

The reason: 60% of agentic AI costs don't go to the first inference call. They go to response refinement — the iterative loop where the agent checks, corrects, and regenerates its own output before returning anything usable.

Your budget models were built around token price. The actual bill lives in the refinement loop your finance team never built a line item for.

One in five organizations has already constrained AI use because of operating costs. Not because the models failed. Not because of resistance. Because of economics.

Agentic tasks consume nearly 1,000x more tokens than conventional chat. The same task can generate wildly different costs depending on tools used, reasoning path, and retry count. There is no stable per-task baseline.

Audit where your agentic token spend actually concentrates. If you cannot attribute costs by workload phase — retrieval, reasoning, refinement — you cannot control the 60% that's eating your budget.

The budget conversation has moved from IT to the C-suite. It is not moving back.