Two years ago an assistant answered a product question for a few cents. Now that same question passes through a planning step, retrieves documents, runs intermediate steps and hands the work to a second agent. Run that a few thousand times a day and the bill at the end of the month is not the number anyone approved.
EY recently put it on paper in a whitepaper on the total cost of agents: a customer interaction that cost around four cents in 2023 now costs about $1.20, roughly thirty times more. And the token bill is only the visible part.
Most of it sits out of view
The bulk of what an agent truly costs never appears on the model vendor's invoice: the infrastructure keeping it running, the oversight that makes its output safe enough to trust, the redesign of the work around it, and the rare error you have to be able to absorb. Costs that only surface once an agent is already scaling, spread across budgets that do not talk to each other.
The model vendor's invoice is the smallest part of what an agent costs.
That explains a prediction quoted widely this summer. Gartner expects more than forty percent of agentic AI projects to be cancelled before the end of 2027. The causes: rising costs, value nobody can demonstrate, and controls that are missing.
Governance is not the cost item
The advice that keeps coming back is sound. Give every cost line an owner before the money is spent, build hard limits before you scale, and put the full cost into the investment case instead of only the model price. But there is a deeper layer underneath. In the usual cost model, governance appears as a line item that grows with every agent you add. Read from the ground up it is exactly the other way around. Discipline in the foundation is precisely what stops the other costs from growing.
An agent does not only cost more because it burns more tokens. It costs more because on a foundation nobody entirely trusts, every task has to go to the heaviest and most expensive model just to be safe enough to ship.
You pay frontier-model prices to buy back trust the foundation itself should be providing.
And the market makes that deliberately expensive: the price of the heaviest models rose through 2026, while lighter models keep getting cheaper.
Pick the model that fits the task
Most work inside an operational organisation is routine and high in volume: matching, pulling data out of documents, classifying, spotting the exception. That work does not need a frontier model, it needs a foundation to stand on. A small or private model on uncontrolled data invents answers; the same model on a grounded foundation, checked against meaning the organisation owns itself, delivers frontier-level results at a fraction of the price, without the data ever leaving the organisation. Pick the model that fits the task, and keep the heavy model for the reasoning that genuinely earns it.
High-volume routine work needs a foundation, not the most expensive model.
First the ground, then the steering
You cannot steer on an outcome you cannot see. The cost of an agent is invisible until you make it visible, and that is a statement about the foundation, not about the model. Lean, Six Sigma and Agile each began by putting the ground in order before steering on what came off it. Agentic AI is no different.
The foundation comes first, or the costs decide themselves.
The advice to give every cost line an owner and build the brake before you scale is right. On a grounded foundation those are simply not procedures you staple to the invoice afterwards. Ownership becomes a property of the ground itself: every number carries who owns it and what it is allowed to do. The brake is the governance that was already there before the first agent ran.
Meaning, ownership and limits are in place before the first agent runs.
The organisations that survive the next cycle are not the ones with the most agents or the largest model budget. What an agent costs is decided by the ground it runs on. You lay that ground once, and before the first agent runs.
Sources
EY, "Unlocking agentic value: a new investment discipline for the agentic era", EY Total Cost of Agents series, 1 June 2026. https://www.ey.com/en_us/insights/ai/agentic-ai-token-costs
Gartner, "Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027", 25 June 2025. https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027
