McKinsey’s latest research reframes the agentic AI scaling question from a technical challenge into a cost-structure problem, and the numbers are striking. Nearly 60 percent of what enterprises spend running AI agents goes not toward generating answers but toward verifying and correcting them. Agentic tasks can consume up to 1,000 times more tokens than standard chat or code-reasoning workloads, making per-token pricing essentially meaningless as a budget tool. The full breakdown of agentic AI cost drivers<\/a> covers six distinct factors, from context length to orchestration overhead to language-based token fragmentation.<\/p>\n The organizations most exposed here are the ones that moved fast in 2024 and 2025, shipping AI agents into production before anyone built a real cost-accounting layer around them. If your team is still measuring agentic AI spend in per-token terms, you’re running a business on a meter that doesn’t reflect what you’re actually buying. The signal that you’re on the right side of this: your infrastructure team can tell you, for a given agent workflow, what percentage of compute cost is refinement versus initial generation.<\/p>\n McKinsey’s framing, coming from a firm whose consulting revenues depend on enterprises continuing to scale AI rather than pump the brakes, tilts naturally toward a “solve the economics and push forward” conclusion rather than a “question the scope” one. That’s a real tilt worth naming, but it doesn’t invalidate the underlying observation. Response refinement as the dominant cost category is structurally true of any system where outputs have to meet a quality bar before they’re usable. The 60 percent figure is the kind of specific that survives the incentive distortion.<\/p>\n The deeper issue is that four of McKinsey’s six cost drivers, context length, refinement loops, unnecessary advanced reasoning, and prompt design, are all things the buying enterprise controls, not the model vendor. That shifts the optimization locus inward. Enterprises that treat agentic AI costs as a vendor pricing problem to negotiate away will keep losing. The ones that treat it as an internal engineering and workflow design problem will actually bend the curve. That’s the budget conversation worth having with your CFO before the next renewal cycle, not whether the tokens cost less, but whether the architecture is generating them unnecessarily.<\/p>\n A token is the basic unit an AI model reads and charges for, roughly three to four characters of English text. Non-English languages often require more tokens to express the same meaning because their characters don’t map as cleanly to the model’s vocabulary, the way a single English word might split into four or five pieces in another script. For global enterprises running agents across multiple languages, this isn’t a minor rounding error; it’s a systematic cost multiplier baked into every non-English workflow.<\/p>\nWhat this means for your business<\/h2>\n
Concept deep-dive: Token fragmentation<\/h2>\n