There is a moment, familiar to anyone who has sat across from a CFO in the last eighteen months, when the conversation turns from enthusiasm to arithmetic.
The enthusiasm is real. The AI budget tripled at the start of the year. Every business unit launched a pilot. The board approved the spend with a kind of relief, glad to be moving. Then a quarter passes, the first real invoices land, and someone asks a simple question that the entire apparatus turns out to be unable to answer: which of these investments is actually paying off?
The honest version of that scene played out at scale this year, in an account the company's own leadership gave publicly. After rolling out coding agents to roughly 5,000 engineers in December, one of the most sophisticated technology organizations in the world watched usage nearly double by February. By March, a large majority of its developers were classified as agentic users. By April, the company had burned through its entire 2026 AI budget, four months into the year, and its CTO acknowledged being, in his words, "back to the drawing board." A separate enterprise, according to press reports, spent half a billion dollars in a single month after deploying AI access with no usage caps.
These are not adoption failures. The agents worked. They are governance failures, and they share a single root cause: the financial foundation underneath the spend was built for a different era.
The practice grew up, almost overnight
To understand why this keeps happening, start with how fast the ground moved.
The discipline built to govern variable, usage-based technology spend is FinOps, and for most of its life it meant one thing: bring financial accountability to the cloud bill. That definition is now obsolete. The FinOps Foundation's sixth annual State of FinOps report, drawn from organizations responsible for more than $83 billion in cloud spend, found that 98% of respondents now manage AI spend, up from 63% in 2025 and 31% in 2024. The Foundation went so far as to change its own mission this year, from advancing the people who manage the value of cloud to the people who manage the value of technology, a one-word swap that ratifies what practitioners were already living.
The scope kept widening past AI alone. In 2026, 90% of FinOps teams manage SaaS, 64% manage software licensing, 57% manage private cloud, and 48% manage data center spend. One practitioner described the progression in plain terms: "First they asked us to fix cloud. Then fix the software mess. Now fix the contract and license mess, now fix the data center."
The teams carrying this stayed lean, even at $100M in cloud spend, the report finds eight to ten practitioners is common, and lean is fine. Lean has always been the operating model. Lean works when the foundation underneath carries the load. What breaks is asking a lean team to hand-assemble a picture of technology spend that the underlying data was never structured to provide.
Tokens broke the model
What that missing structure hides, increasingly, is AI.
The reason is structural. Enterprise software ran for two decades on a premise finance teams could model in their sleep: pay per seat, scale predictably, sign an annual contract. Token-based pricing dissolves that premise entirely. As Deloitte put it in a January CFO guide on the subject, businesses now need to treat AI economics with the same rigor as energy or capital allocation, recognizing tokens as the new currency. The same analysis found that half of enterprise leaders are now spending 21 to 50% of their digital transformation budgets on AI, a figure that would have read as absurd two years ago.
The volatility comes from the mechanics of the pricing itself. Token spend is metered per request, with no seat count and no contract ceiling to anchor a forecast against. Consumption moves hour to hour with product traffic and developer behavior. The unit price swings with model choice, prompt design, and orchestration pattern, the same task can cost ten times more or less depending on decisions no invoice line ever records. And where a seat license commits once a year, token spend commits thousands of times a day, at machine speed, far ahead of any monthly review. A budget line with those properties behaves less like enterprise software and more like an unhedged commodity position.
Then agents multiply it. EY's analysis this year laid out the math cleanly: a simple linear workflow that cost about four cents per interaction in 2023 becomes, in a more complex orchestrated system involving tools, reasoning, and iterative loops, roughly $1.20 per interaction, about 30 times higher. Gartner's read is consistent, finding that agentic models require five to thirty times more tokens per task than standard chatbots. Goldman Sachs has estimated that agents could multiply enterprise token demand twenty-four-fold by the end of the decade.
The cruelest part is how predictable the waste turns out to be. A Q1 analysis of 2.4 billion enterprise API calls found that organizations routing every workload to frontier models paid $18.40 per million tokens, while those running a tiered architecture paid $2.31, an 87% gap that is the direct financial consequence of one architectural decision made at the start of the deployment and, in most cases, never revisited. The first engineer chose the model already open on their screen. Nobody went back when the pilot became production. That single default, the analysis concluded, is the largest controllable variable in most enterprise AI cost structures.
The question the board is actually asking
Lay these threads side by side and a pattern emerges that should reframe how every technology leader thinks about the next budget cycle.
The spend is exploding, the unit of spend is invisible where decisions get made, and the returns are genuinely hard to prove. Any finance team can now report what AI costs, the invoice totals are right there. What almost none can report is what it earns. PwC found that only 12% of CEOs say AI has so far delivered both cost and revenue benefits. BCG found that nearly 90% of CEOs expect agents to deliver measurable ROI, while few organizations can point to consistent results. One FinOps practitioner, quoted in the State of FinOps data, put the whole predicament into a single sentence:
"Is your AI providing value? No one can answer that question yet."
That sentence is the entire problem, and it is worth sitting with, because the instinct it provokes is usually the wrong one. The instinct is to reach for another dashboard, another optimization sprint, another strategy refresh. The deck comes back blank anyway. It comes back blank because the issue lives a layer below the dashboard, in the substrate the dashboard sits on top of. You can rewrite the strategy as many times as you want. The financial infrastructure stays broken until someone rebuilds the foundation.
The most advanced practitioners have already internalized this, and the 2026 data shows where they are headed: upstream. The center of gravity is moving away from optimizing spend after the invoice arrives and toward governing it before the money is committed. Pre-deployment architecture costing is the top desired tooling capability in the survey, practitioners want financial context introduced before infrastructure is provisioned, at the moment of the decision, rather than reconstructed weeks after the fact. And the organizations least disrupted by all of this turned out to share one trait: they had already built unified data layers before AI and SaaS spend became material, and they extended existing systems rather than rebuilding from scratch.
They built the foundation first. Everyone else is retrofitting it under a running system.
A common language for the foundation
The foundation already has a shared language, and that is what makes rebuilding it a tractable project rather than a research one.
FOCUS, the FinOps Open Cost and Usage Specification, a vendor-neutral schema that lets billing data from any provider be read and compared with the same column names and meanings, is now the standard the industry builds on. The FOCUS Steering Committee ratified version 1.4 on June 4, 2026, adding two datasets, forty-seven columns, and support for reconciling usage directly to invoices. Native exports are available from more than eleven technology providers including AWS, Microsoft Azure, Google Cloud, Oracle, and Alibaba Cloud, and the specification extends past cloud to SaaS and PaaS billing, the fastest-growing categories on the FinOps mandate. The practical implication for any buyer evaluating tooling this year is blunt: a vendor without a FOCUS strategy is a yellow flag.
FOCUS is not finished, and honesty about that matters more than enthusiasm. Coverage is uneven, AWS's general-availability release closed eleven specification gaps but left eight open, and Google Cloud still lacks a one-click export. No provider is fully conformant yet. But the direction is settled, and the substrate underneath every cost model has a standard to be built against, rather than a thousand proprietary billing taxonomies to be stitched together by hand on a Friday afternoon.
Which leaves the real work exactly where it has always been: in the foundation. The vast majority of an organization's total cost of ownership still arrives as messy, unstructured data, millions of dollars transacted through PDF invoices, SaaS contracts in one system, AI compute with no model at all. Mapping all of that into a single, business-aware, FOCUS-aligned model is the unglamorous substrate beneath every cost decision that has any chance of surviving the next five years.
"The shiny thing is easy to build on top of a solid foundation. The foundation is impossible to retrofit underneath a shiny thing."
What to do before the next cycle
For the technology leader reading this with a board meeting on the calendar, three moves separate the teams that will manage their AI spend from the teams that get managed by it.
Instrument the agentic workflows now. Every agent loop is a stream of spend decisions happening at machine speed, and most organizations cannot say which workflow, team, or product each token belongs to. Map that spend to the business before the volume grows. The multiplier on agentic consumption makes this the first place a missing foundation shows up.
Move the decision upstream. The most expensive choices get made at design time, by an engineer picking a model, long before the invoice arrives. Put cost context where that choice happens, inside the workflow, at the moment of the decision, so the economics are visible while they can still change.
Build on the standard. Turn on native FOCUS exports across every provider, land them in one place, and choose tooling that reads the format natively. The format is free. What it gives you is a common language, every source described the same way, so the cost model gets built once instead of re-stitched for each provider. The model on top of that language, governed and versioned like code, is what turns the board's question into an answer in seconds rather than a discovery when the budget runs out in April.
The cost question has been answered, every team can pull the invoice total. The question that remains open, as EY framed it, is what the spend earns. That one deserves a live answer, at the moment the decision happens, human or agent, rather than a quarter later when the bill arrives and the deck comes back blank again.
The companies that build the foundation to answer it now will have a structural advantage as the cost curve evolves. The rest will keep rewriting the strategy, and keep wondering why the substrate underneath it still cannot answer the question.
The data cited here draws on the FinOps Foundation's State of FinOps 2026 report, the FOCUS 1.4 specification ratified in June 2026, and 2026 analyses from Deloitte, EY, Gartner, PwC, and BCG.