McKinsey: the same agent task can cost thirty times more one run to the next

Fortune put a McKinsey argument in front of CFOs on Wednesday, and the underlying numbers are worth separating from the headline. In “Is that AI agent worth it?”, published 13 July, McKinsey states that in agentic programming work “the same task can have a factor-of-30 variation between completions.” Not between models, and not between vendors — between two runs of the same task, because the agent takes a different path each time.

The second number reframes the first. In an August companion piece McKinsey puts token costs at “just 20 to 25 percent of the variable run costs of an AI agent”, with human oversight accounting for 70 to 75 percent. If four-fifths of the bill is people checking the work, a 90% cut in token prices moves the total by single digits.

Put those together with Epoch's finding and the apparent paradox dissolves. Intelligence at a fixed capability really is collapsing in price — McKinsey senior partner Lari Hämäläinen says so himself: “Intelligence at a certain capability level is getting a lot more affordable.” But the unit being bought changed underneath the price cut. You are no longer buying one answer; you are buying an agent that will attempt a task an unpredictable number of times and then need a person to verify it.

McKinsey's prescription is to stop governing the token. Its own phrasing: “the unit of governance is the completed business outcome, not the token, model call, or technology line item”, and more bluntly, “Tokens are not value; tokens are the bill.” For a small business signing a flat-fee or per-seat agent contract priced off token math, a 30x swing on a single task is the number that ends the argument.

• Confirmed McKinsey states “the same task can have a factor-of-30 variation between completions” in agentic programming work — a run-to-run variation, not a price difference. McKinsey QuantumBlack, 13 Jul 2026
• Confirmed Token costs are “just 20 to 25 percent of the variable run costs of an AI agent”, while human oversight accounts for 70 to 75 percent. McKinsey, 24 Aug 2026
• Reported Fortune's 23 September write-up frames the operational metric as task execution cost, agent success rate and human verification time rather than cost per token. Fortune