AI Agents Could 30x Your Bill, McKinsey Warns

Last updated: September 29, 2026

When companies first wired up large language models, the IT invoice line item barely registered. A few million tokens, a few dollars, on to the next category. That mental model is breaking. According to businessinsider.com, McKinsey & Company is telling corporate buyers that the cost of running AI agents can swing by a factor of thirty depending on how a task is carried out, and that this spending is no longer rounding error inside a technology budget.

The warning lands just as adoption has shifted from pilot to default. McKinsey's 2026 State of AI survey, cited in the same reporting, puts roughly one in three organizations above the ten-percent mark of their technology spend, and most respondents plan to spend even more in the year ahead.

At a glance

• McKinsey says completion costs can differ by up to thirty times between AI agents doing the same job.
• About a third of organizations now devote more than a tenth of technology outlay to AI.
• Sixty percent of surveyed companies plan to lift AI spending next year.
• One in five respondents say AI costs are already squeezing operating budgets.
• Coding-focused agents are singled out as the most token-hungry workloads.

What changed: from chat windows to autonomous workflows

For most of the past two years, corporate AI use meant a chat box. An employee typed, got an answer, moved on. Tokens were cheap, usage was bounded by typing speed, and finance could ignore it.

Agents behave differently. They read files, call APIs, write code, run commands, and review their own output, often in loops. Every loop consumes tokens, and loops multiply consumption. The same business goal, such as triaging a ticket or refactoring a module, has many possible paths, and the paths do not cost the same. That is where the thirty-times spread comes from.

The spending shape changes too. A chatbot is a recurring subscription you can forecast. An agent fleet behaves like a metered utility, where one inefficient workflow can drain a quarter's budget in a week.

Why this matters for project and engineering leaders

McKinsey's framing, as reported, is blunt: roughly a third of organizations have crossed the ten-percent threshold, and one in five already feel AI costs as a real constraint. For anyone running agile delivery or a product engineering group, this is the moment when AI moves from productivity footnote to steering-committee slide.

The second-order signal is harder to dismiss. Coinbase, Salesforce, and Amazon have each taken visible steps to throttle or monitor AI use as invoices climbed. Amazon reportedly retired an internal leaderboard that some staff were gaming purely to push their token counts higher. When employees start optimizing for usage metrics instead of outcomes, you have both a cost problem and a behavioral problem in the same room.

For leaders used to agile ways of working, the implication is uncomfortable but familiar: you cannot iteratively improve what you cannot measure. Per-task cost, not per-seat subscription, is the unit that needs to land on the team dashboard.

Our analysis

The most useful way to read McKinsey's warning is not "AI is getting expensive" but "AI is getting unevenly expensive." The median spend can look reasonable while a handful of long-running, multi-step agents quietly dominate the invoice. That asymmetry is what makes the thirty-times figure credible rather than alarmist.

Who benefits? Vendors who bundle predictable, capped pricing around agent runs, FinOps-style tooling vendors, and consultancies that can install cost-discipline programs the way they once installed cloud-cost programs. Who loses? Teams that built impressive demos on uncapped usage and now have to defend renewal numbers, and any leader who treated AI as a flat line item in last year's plan.

It also reframes the productivity story. McKinsey's earlier work, cited in the same coverage, claims agentic systems can cut human time on transformation-office tasks by a wide margin. The new puzzle piece is that "time saved" and "tokens burned" are both first-class metrics, and they are not always correlated.

A practical scenario

Say you are a scrum master with four engineers told to use AI aggressively. Without guardrails, an agent given a vague ticket might loop for twenty minutes before producing output, while a well-specified ticket returns a usable patch in two. Across a sprint, that gap is the difference between a tool your team likes and a budget conversation with your director.

A concrete first move: for every story that touches an agent, capture three numbers at close: tokens consumed, wall-clock time, and a one-line human verdict on usability. After two sprints you have a real cost-per-story distribution, not a vendor benchmark.

What to watch next

Watch finance teams start asking the same questions of AI budgets that they learned to ask of cloud a decade ago. Watch vendors move toward per-task pricing instead of per-token. And watch your own dashboards: if you cannot yet explain, on one page, what each agent run cost and what it produced, you are not ready to scale usage, whatever the survey headlines say. The bottom line from McKinsey is not that AI is too expensive; it is that without measurement, you cannot tell.

Key takeaways

• Treat AI spend as a per-task FinOps problem, not a flat subscription, because the same goal can cost thirty times more depending on the agent's path.
• Put tokens consumed, wall-clock time, and a human quality verdict on the same dashboard for every AI-assisted story.
• Expect budget conversations to focus on the ratio of agent output to agent cost, not on seat counts.
• Be cautious scaling agent use in software development, where workflows are flagged as especially token-hungry.

Source: businessinsider.com