Most of what an agentic task costs is the agent checking its own work

Roughly 60 percent of an agentic task's cost goes on the model re-checking its own output before it answers.

McKinsey published work on agentic economics in July 2026, putting roughly 60 percent of a task's cost into response refinement. A separate finding has 93 percent of enterprise AI teams exceeding budget.

Most coverage framed this as AI costing more than promised, though the more useful reading is about where the money goes.

These figures come through MarketScale's reporting of McKinsey's work rather than from the McKinsey publication directly. They are as relayed rather than as read.

Two figures with two provenances, and they shouldn't be blurred together. The 93 percent over-budget figure comes from a FinOps survey of 75 respondents, which is a small base for a number quoted this precisely. The 60 percent refinement figure is not from that survey at all, since McKinsey attributes it to separate research we haven't traced.

Be precise about the definition, because it's narrower than it sounds. The source defines response refinement as the loop where the agent checks, revises and regenerates its own output before it answers. That's machine-side token spend, not human review time.

Read as the source means it, the finding is that most of a task's cost is the model second-guessing itself. That's a claim about token spend per task rather than about an organisation's total AI budget, and the two get blurred constantly.

Beyond what the source says, and offered as opinion. There's a second refinement cost the survey doesn't measure, which is the human loop on top of the machine one. The model call is close to free. What surprises people is the eleventh revision of a prompt, because the output was nearly right.

Both loops share a root, which is why they are worth thinking about together. An agent without the context to get it right will burn tokens re-checking itself and then burn somebody's afternoon afterwards, so fixing the input is what reduces both.

It also puts build versus buy on the right axis. The question is rarely whether you can build it, because increasingly you can, but whether you want to own the refinement loop forever.

The question is rarely whether you can build it. It is whether you want to own the refinement loop forever.

When you cost an AI project, put a real number on refinement before you start and make it uncomfortable. Say the calls run to twenty pounds a month. The honest total is higher, because you add the hours spent making output usable.

Reduce the loop rather than absorbing it. A model given thin context produces something generic, then hands a human the job of specialising it. Fixing the input is cheaper than correcting the output, and it stays fixed.

If you're hiring this out, ask what happens to refinement after handover. A build that leaves you with a system nobody can adjust has moved the cost rather than removed it.

The Diagnostic prices that refinement loop before you commit to it.

Sources

How we choose and label these: our sourcing standard.

  • 60% of agentic AI costs go to response refinement, and most enterprises are already over budget →

    MarketScale · Reporting on the original research · link checked 1 Aug 2026

    Reports McKinsey's figures. It defines response refinement as the loop in which an agent checks, revises and regenerates its own output before answering. That's a machine-side token cost, narrower than the everyday sense of the word. We haven't read the McKinsey publication directly.

  • Is that AI agent worth it? Agentic economics and the modern operating model →

    McKinsey & Company (QuantumBlack) · Primary source · link checked 1 Aug 2026

    The originating McKinsey publication (McKinsey Quarterly, 13 July 2026). Its Enterprise AI FinOps Survey ran in May 2026 across 75 qualified respondents, a small base for a percentage quoted this precisely. McKinsey blocks scripted requests, so this was confirmed by fetching the page rather than by a status check. We read its summary and the MarketScale reporting, not the full article.