Share with your CIO
McKinsey’s latest report lands a pointed argument: the enterprise AI cost conversation has been asking the wrong question. Companies spent two years optimizing token pricing and model access. Now, as agentic AI systems move into production, the real cost driver is context management, verification loops, and orchestration overhead, not inference itself. Nearly 60 percent of agentic AI operating costs go toward verifying and refining outputs rather than generating them, and agentic tasks consume roughly 1,000 times more tokens than standard chat applications.
What this means for your business
The organizations most exposed here aren’t the ones that moved slowly on AI agents. They’re the ones that moved fast and measured nothing. If your agentic deployments were greenlit on the strength of pilot demos and per-token cost estimates, those numbers are now structurally misleading. Per-token pricing, as McKinsey notes directly, has stopped being a useful proxy for what enterprises actually pay once agents start chaining tools, retrying tasks, and maintaining long-lived context across multi-step workflows.
The 60 percent figure on verification costs points to something the AI industry’s framing consistently obscures: the expensive part of an AI agent isn’t the answer, it’s the checking. Agentic systems don’t just generate outputs; they loop, retry, and validate, and each cycle compounds the token count. McKinsey, whose advisory practices benefit when enterprises find their AI deployments complicated enough to require outside help, still lands on a structurally correct diagnosis. The cost variability problem is real and systematically underweighted in enterprise AI business cases. Deploying a sophisticated reasoning model to handle a routine document classification task isn’t just wasteful; it’s the kind of architectural decision that turns a favorable unit economics model into a budget crisis at scale.
The decision this reframes isn’t whether to scale agents. It’s whether your current infrastructure gives you the visibility to know when scaling is destroying value instead of creating it. The organizations that will absorb this finding without pain are those that already treat agent orchestration, the layer governing how AI agents coordinate with each other and with external tools, as an engineering discipline with its own cost telemetry. Everyone else is flying on assumptions that the McKinsey data suggests are wrong by an order of magnitude. I’d revise this assessment if enterprises start reporting that agentic ROI is tracking to original projections; so far, the trajectory runs the other direction.
Concept deep-dive: Long-lived context
In standard chat AI, each exchange is relatively self-contained, keeping token counts modest. Agentic AI tasks require the system to carry a running memory of prior steps, tool outputs, decisions made, and work remaining across an entire multi-step workflow, the way a project manager holds the full thread of a project rather than just the last email. That accumulated memory is “long-lived context,” and it’s why agentic token consumption is orders of magnitude higher than chat, making it the single largest cost driver in production deployments.
Based on reporting from AI Agents ROI: AI Agents Facing ROI Challenges Amid Rising Operating Costs, McKinsey Report Reveals, ETEnterpriseai, originally published 2026-07-18 02:31:00.
