Budget lines are not decision makers

Budgeting for AI and managing it in real time are two different jobs. Mavvrik's 2026 State of AI Cost Governance Report, compiled with Benchmarkit from 396 organizations surveyed in April and May, found that 98 percent of companies now track AI infrastructure costs and 95 percent assign a formal AI budget. Only 11 percent can forecast that spend within 10 percent in either direction. Without a person responsible to manage the budget in real time, it is just a number written down and never checked again. 62 percent of the same organizations said an unexpected AI cost had already changed a real business decision this year, through forced repricing, spending freeze, or cancelled projects.

Uber's experience is the clearest open case study of how that happens. The company burned through its entire 2026 AI budget in four months, much of it on agentic coding tools running $500 to $2,000 per engineer per month. Its response was a flat cap of $1,500 per employee per tool, a control that rations productive and wasteful usage identically because nobody had the attribution data to tell the two apart. Accenture's research finds that only 23 percent of C-suite leaders report widespread, sustained business value from AI, and 57 percent point to weak alignment between the business and the technology function as the reason. This is exactly the position where a named individual would sit if the job existed.  

Such dollar figures above can be absorbed by companies with balance sheets that are big. Mid-market companies and smaller einterprises will see  its own Cursor or Claude Code contract hit the same four-to-five-times jump Priceline saw. It is not just an increase or revised costs. That is a chunk of a department's quarterly budget, discovered only after the invoice arrived, which could have caught been caught sooner with dedicated resources. The enterprises above can absorb one bad quarter while they build the muscle. Mid-market companies get one shot at noticing the number before it becomes the headlines.  

Why the same solution won't work everywhere

Readers who are critical will object here, as they should. The objection is that the largest buyers already have this solved, organized differently than a single name implies. That deserves a straight answer. Accenture routes its own internal traffic through a gateway that defaults new sessions to a lighter model and reserves frontier access by role, a decision its CFO and CIO are asked to own jointly. JPMorgan’s leadership shaped the Tokenomics Foundation's direction before the public launch, which is itself a form of institutional ownership. Thoughtworks, in its own recent piece on this problem, recommends a standing AI governance committee with explicit duties: treat consumption as a risk, set policy on approved models, and own the metric decision. But these are companies with existing FinOps maturity.
‍
Mid-market companies don't have a CFO, a CIO, and a head of engineering with bandwidth to jointly own decisions or run a standing committee on token spend. And it doesn't have three people whose job description already includes reviewing an AI invoice. It assigns responsibility to whoever sits closest to the budget, usually an operations or finance lead already stretched across a dozen other responsibilities. Asking that company to follow Accenture's structure before it can act is asking it to solve a staffing problem before it solves a spending problem. Naming one person, even someone doing this as a fraction of an existing role, is the only version of this answer a company that size can build this quarter. In this version, acting on it this week beats a better one where it can't staff for another year.

‍A trial for Monday morning

Here is a simple exercise that costs nothing to do. Go to your office and ask who in the building would have to explain it if this month's AI bill doubled, and whether that person could have stopped it before the invoice arrived rather than after. If the honest answer is that finance would ask engineering and engineering would ask whoever ran the workload, there is no owner. There is just a chain of people who can point at one another.

This is an org-chart decision, not a procurement decision, which is exactly why it doesn't have to wait for a tool, a vendor contract, or a standards body's roadmap. The companies that have solved for this chain of people, did not wait for a standards body to finish its roadmap. They gave one person the budget and authority tied to a named role rather than a shared API key. That person uses that authority to set targets in both dollars and tokens, so a price change couldn’t silently expand what someone was allowed to consume. They also reviewed the numbers on a fixed schedule instead of only when a bloated invoice arrived to call a meeting. When a name is attached to the number, tracing an overrun back to its cause takes an afternoon. When it is not, the tracing takes a committee, and by the time the committee reports back, the quarter is already over.

The Tokenomics Foundation will, over the next several years, do for AI spend roughly what the FinOps Foundation did for cloud spend: give large buyers and sellers a shared vocabulary and a common invoice format. That is progress worth watching closely. But it answers the question the market would have answered anyway, because both sides of that transaction profit. The question that needs to be answered is the one of accountability: who gets to say no, and to whom. That question has no foundation behind it and no date by which it must be answered. Companies that assign it this quarter will spend next year's budget cycle making decisions. Companies that wait for the standard to arrive will find out when the next invoice arrives.