McKinsey: 60% of agentic AI costs go toward refining responses, as enterprises shift focus to ROI

A new McKinsey report finds enterprise Gen AI adoption is entering a phase
defined by financial sustainability rather than experimentation. Nearly 60% of
agentic AI’s operating costs go toward verifying and refining responses, not
generating them. Agentic tasks can consume nearly 1,000x more tokens than
standard chat or code tasks, making per-token pricing an unreliable cost
measure. McKinsey names six main cost drivers, including long-lived context,
response refinement, cost variability, over-use of reasoning, orchestration, and
information structure.

Scaling AI agents is now a finance decision, not just a technical one. Enterprises
should build cost-tracking around context length, retries, and orchestration
before scaling agentic deployments company-wide.