McKinsey has released a report on the economics of agentic AI, and its central number should unsettle anyone budgeting an AI rollout: about 60 percent of what an agentic task costs is spent after the first answer is generated, on verifying and refining it. As reported by ANI via The Economic Times on July 17, the firm argues that the next phase of enterprise AI adoption will be decided by unit economics, not by model capability.
The timing is not accidental. For two years most companies focused on getting access to models, running pilots and shipping first deployments. Now agents are moving into production, and the people asking the questions have changed. According to the report, CFOs and CIOs increasingly demand evidence that AI investments deliver measurable returns, while many companies still lack systems to track the business impact of AI-driven decisions.
Scaling an agent is now a financial decision
"The decision to scale an agent is increasingly becoming a complex and fast-changing economics decision, not a technical one," the report states. That is a quiet but real shift in who owns AI inside a company. When the blocker was accuracy or latency, engineering owned the roadmap. When the blocker is cost per completed task, finance sits at the table, and the projects that survive are the ones that can show their math.
Why per-token pricing stopped working
The report notes that agentic tasks can consume nearly 1,000 times more tokens than conventional chat or code-reasoning tasks. An agent does not answer once: it plans, calls tools, reads results, retries and carries long-lived context across every step. "Per-token pricing has stopped being a useful measure for what enterprises actually pay for gen AI," the report says. In other words, the meter most buyers still look at measures the wrong thing.
The six cost drivers McKinsey identifies
The report breaks agentic operating expenditure into six factors:
Interested in implementing similar AI solutions? Discover how PATech Labs can help your business leverage cutting-edge artificial intelligence.
Learn About Our Services- Long-lived context. Carrying history and working memory across steps is one of the biggest single contributors to cost.
- Refinement over generation. Modifying an AI-generated answer costs more than producing the first draft: about 60 percent of a task's cost is tied to refining answers.
- Cost variability. The same task can cost different amounts depending on which tools the agent picks, which reasoning path it takes and how many retries it needs.
- Advanced reasoning on simple tasks. Extended reasoning pays off on hard problems and burns money on routine ones.
- Orchestration. How agents coordinate with tools, models and one another materially moves the bill. Efficient task allocation can cut costs without hurting outcomes.
- Information structure. Prompt design, context length, formatting and even language change token consumption. "Non-English text, for example, gets fragmented into more tokens per meaning, so the same conversation costs more in some languages than others," the report notes.
The same task never costs the same twice
The variability point deserves emphasis, because it breaks conventional IT budgeting. Classic software has a known cost per transaction. An autonomous agent does not: a lucky run resolves in three steps, an unlucky one loops through retries and tool calls. Coverage of the report by EdexLive frames managing token consumption and reasoning paths as a new core skill for AI engineers, and that matches what the numbers imply: cost control is becoming an engineering discipline of its own.
What this means for businesses deploying AI agents
At PATech we run agents in production every day: voice agents that answer business calls and a content pipeline that researches, writes and renders on its own. The report matches what we see in our own bills, so here is our practical read.
First, measure cost per completed task, not per token or per request. A cheap model that needs five retries is more expensive than a capable one that finishes in one pass. Second, route by difficulty: reserve deep reasoning for steps that genuinely need it and push routine steps to smaller, faster models. Third, put a hard budget on the refinement loop. Since roughly 60 percent of cost lives in verify-and-refine, capping retries and defining what counts as good enough is a direct financial control, not a quality compromise. Fourth, treat context as inventory: trim what the agent carries between steps, because every stored paragraph is billed again on every subsequent step.
None of this argues against agents. It argues against deploying them without instrumentation. The companies that win this phase will know their cost per outcome to the cent, the way logistics companies know cost per delivered package. We covered the demand side of this market in our piece on enterprise AI agent spending reaching 206 billion dollars in 2026; this report is the supply-side discipline that spending will require.
The bottom line
Agentic AI is leaving the demo stage, and the bill is arriving. McKinsey's message is not that agents are too expensive: it is that their costs behave differently and must be engineered, tracked and owned. Teams that build cost telemetry, routing and refinement budgets now will scale with predictable economics. Teams that skip it will find that their CFO, not their model, becomes the bottleneck.