Share with your CIO
McKinsey’s latest analysis lands a pointed argument: the hard part of agentic AI is no longer building it, it’s paying for it sustainably. The agentic AI cost structure is structurally different from conventional software, with roughly 60 percent of operating costs tied to verifying and refining AI-generated responses rather than producing them. Agentic tasks can consume nearly 1,000 times more tokens than standard chat or code tasks, and per-token pricing has effectively become a meaningless unit of measure for enterprise budgeting at that scale.
What this means for your business
The shift McKinsey is describing separates two types of enterprises right now. Those running a handful of AI pilots haven’t felt this yet, because experimentation costs are absorbed as R&D noise. The ones who will feel it acutely are organizations that have moved agents into production across workflows, where cost variability, the way the same task can produce wildly different compute bills depending on reasoning path and retry count, starts compounding before any governance framework exists to catch it.
The 60 percent figure for response refinement costs deserves scrutiny. McKinsey, whose advisory business benefits from clients believing AI complexity requires outside guidance, has an incentive to frame scaling challenges as more opaque than they are. But the underlying math is hard to dismiss. Agentic systems by design don’t produce single outputs; they iterate, check, and retry. That loop is where the money goes. The implication isn’t that agentic AI is too expensive to deploy, it’s that the architecture decisions made at the design stage, which models handle which tasks, how much context gets preserved, whether extended reasoning gets triggered on routine requests, are now financial decisions with material P&L consequences, not just engineering preferences.
The language cost detail buried in the report is the kind of thing that gets missed until it doesn’t. Non-English text tokenizes less efficiently than English, meaning the same customer service workflow costs measurably more in Japanese or Arabic than in English. For any enterprise running global operations, that’s not a rounding error, it’s a structural pricing disparity baked into every vendor contract written against token consumption. The CIO who renegotiates vendor agreements this year without building language-adjusted cost models into the baseline is signing a contract that will look increasingly wrong by Q2 2026.
Concept deep-dive: Token context in agentic systems
A token is roughly a word fragment, the basic unit AI models read and charge for. In a simple chatbot, each exchange is short and self-contained. In an agentic system, the AI carries a running memory of the entire task, prior steps, tool outputs, instructions, and intermediate results, called the context window. That window grows as the task runs, and every token in it gets reprocessed at each step. Think of it as a growing meeting transcript that every participant must reread before speaking. The longer it runs, the more expensive every new contribution becomes.
Based on reporting from Scaling Agentic AI: Why Economics Now Matter More Than Engineering, originally published 2026-07-17 08:45:00.
