The direct answer, in one paragraph. Seventy-three percent of companies overshoot their AI budget, some by a factor of 2.4. The same finance teams that forecast cloud spend within one to three percent miss AI spend by two to three times. This is not incompetence. It is a category error. Enterprises budgeted AI like software, and AI behaves like electricity. Software has a seat price you multiply by headcount. Electricity has a meter that runs faster when things are going well. Until your budgeting model accounts for that difference, you will miss your number every quarter.
The numbers
Ninety-eight percent of FinOps practitioners now manage AI spend. In 2024 that figure was thirty-one percent. In 2025 it was sixty-three percent. A discipline that barely touched AI two years ago now touches almost nothing else.
Seventy-three percent of companies overshoot their AI budget. In a review of enterprise agentic AI projects, some overshot by a factor of 2.4.
Mature cloud FinOps teams forecast within one to three percent of actual spend. Those same teams, applying the same discipline to AI, miss by two to three times.
And the strangest number of all: the price per token keeps falling, and the total bill keeps rising.
The paradox that breaks every forecast
Model pricing has fallen dramatically and continues to fall. Every finance team that built an AI budget assumed this would work in their favor. It has not.
Six things happen at once, and they compound.
**Cheaper units drive more usage.** This is Jevons paradox applied to inference. When a query costs a tenth of what it did last year, teams stop rationing. Usage does not stay flat while price falls. It expands to fill the new economics, and usually overshoots them.
**Agentic workflows multiply calls per task.** A chatbot answering a question is one call. An agent completing a task plans, retrieves, acts, observes, and replans. That is not one call. It is frequently dozens, sometimes hundreds. The user performed one action. Your bill recorded forty.
**Context windows got bigger, so every call got heavier.** Larger context is a feature, and teams use it. Retrieval-augmented systems stuff more documents into each request because they can. Per-token price fell, per-call token count rose, and the two moved in opposite directions at different speeds.
**Reasoning models burn tokens you cannot see.** Extended thinking consumes output tokens before it produces a single word the user reads. The invisible portion is often larger than the visible one. Finance sees the bill. Nobody sees the reasoning.
**Everything around the model also calls the model.** Guardrails, evaluations, retries, content classification, summarization of prior turns. These are all model calls. They are not what anyone budgeted for because they are infrastructure, not features.
**Success is the most expensive outcome.** A pilot that nobody uses costs nothing. A deployment people love scales linearly in cost with the thing you were hoping for. The better it works, the more it costs, and the budget was written assuming the pilot's volume.
Put those six together and the arithmetic is straightforward. Unit price can fall by eighty percent while total spend triples.
Why software budgeting fails here specifically
Enterprise software has a shape finance understands. You negotiate a per-seat price, multiply by headcount, add a growth assumption, and you have a number that is correct within a few percent for twelve months. The entire procurement and budgeting apparatus is built around that shape.
AI consumption has a different shape. Three properties break the model.
**Cost is coupled to behavior, not headcount.** Two companies with identical seat counts can differ tenfold in spend depending on how their people use the system. Headcount tells you almost nothing.
**Demand is not stable.** Usage follows adoption curves, seasonality, and campaign cycles. A single successful internal launch can double consumption in a week with no corresponding change in the license.
**The unit of consumption is invisible to the buyer.** Nobody in procurement has intuition for a token. They have intuition for a seat. When the unit of consumption cannot be reasoned about by the people approving the budget, the budget is a guess wearing a spreadsheet.
What AI FinOps actually is
FinOps as a discipline emerged because cloud broke the same assumptions a decade ago. Capital expenditure became operating expenditure, provisioning became self-service, and finance lost visibility. The answer was not to stop using cloud. It was to build a practice that made variable spend governable.
AI FinOps is the same move for inference. It is not a cost-cutting initiative. It is the practice of making unpredictable spend forecastable, attributable, and tied to outcomes.
The distinction matters because framing it as cost-cutting produces the wrong behavior. Teams throttle usage of the thing that was supposed to create value, and the deployment quietly dies. The goal is not to spend less. It is to know what you are spending, on what, for what return, before the invoice arrives.
Six controls that make AI spend forecastable
**Attribution before optimization.** Every model call should carry tags identifying the team, the workflow, and the use case. Without attribution, you have one large number and no way to act on it. Most enterprises skip this because it requires engineering work at exactly the moment everyone wants to ship. It is the single highest-return control and it must come first.
**Token budgets per workflow, enforced in code.** Not policy. Code. Each workflow gets a ceiling on tokens per task and per day. When the ceiling is hit, the system degrades gracefully or escalates rather than silently continuing to spend. This is the control that prevents the 2.4x overshoot.
**Model routing by task difficulty.** Most workloads do not need the frontier model. Classification, extraction, routing, and summarization run acceptably on smaller models at a fraction of the cost. Route by task, not by habit. Enterprises that implement routing typically cut spend materially without any change in output quality, because they were overpaying for easy tasks.
**Caching at every layer that allows it.** Prompt caching, response caching for deterministic queries, embedding reuse. The cheapest token is the one you do not send. This is unglamorous and it is where a large share of the savings lives.
**Anomaly detection with routing to an owner.** Spend spikes should alert the team that caused them, not land in a monthly report that finance reads three weeks later. Cloud providers ship tooling for this now. The important part is not the detection. It is that the alert reaches someone with the authority and the context to act.
**Unit economics per outcome, not per token.** The number that matters is cost per resolved ticket, cost per qualified lead, cost per document processed. Token cost is an input. Nobody outside engineering can act on it. Outcome cost is what determines whether the deployment survives its next budget review.
The forecasting model that actually works
Stop forecasting AI spend from headcount. Forecast it from three variables.
**Volume.** How many tasks per period will hit this workflow. This is a business number, not a technical one, and the business owner should produce it.
**Tokens per task.** Measure this from production, not from the pilot. Pilot workloads are almost always cleaner and smaller than real ones. Measure the p50 and the p95, because the tail is where budgets die.
**Blended rate.** Your actual cost per token across the model mix you route to, not the list price of your most expensive model.
Multiply the three, add a buffer sized to your p95 rather than your average, and you will land far closer than a seat-based estimate. Teams that do this reliably get inside ten to fifteen percent, which is not cloud-grade precision but is a different universe from missing by 2.4x.
What to do this quarter
Tag every model call with team, workflow, and use case. If you can only do one thing, do this one. You cannot manage what you cannot attribute.
Pull your last ninety days of spend and calculate cost per outcome for your top three workflows. Most enterprises have never computed this and are surprised by what they find. Some workflows will be obviously worth it. At least one usually is not.
Check whether any production workflow is calling your most expensive model for a task a cheaper one would handle. This is the most common and most easily fixed source of waste.
Set token ceilings on your two highest-volume workflows and make them enforceable in code. Policy ceilings do not hold.
Name one person accountable for AI spend. Not a committee. The same principle that applies to agent behavior applies to agent economics.
The part most enterprises get backwards
Rising AI spend is not automatically a problem. If cost per outcome is falling and volume is rising, the bill going up is the deployment working. That is the good version.
The problem is not knowing which version you are in. An enterprise that cannot decompose its AI bill into workflows, tasks, and outcomes cannot distinguish a successful deployment from a runaway one. Both look identical on the invoice.
Ninety-eight percent of FinOps teams now manage AI spend because that ambiguity became intolerable. The ones who built attribution first can tell the difference. The ones who did not are throttling usage and hoping.
Frequently asked questions
Why do most companies overshoot their AI budget. Seventy-three percent of companies overshoot, some by 2.4x, because they budget AI like per-seat software when it behaves like metered consumption. Cost is coupled to user behavior rather than headcount, demand is unstable, and the unit of consumption is invisible to the people approving the budget.
Why does the AI bill rise when the price per token falls. Six factors compound. Cheaper units drive more usage, agentic workflows multiply calls per task, larger context windows make each call heavier, reasoning models consume invisible output tokens, surrounding infrastructure like guardrails and evaluations also calls the model, and successful deployments scale in volume. Unit price can fall eighty percent while total spend triples.
What is AI FinOps. AI FinOps is the practice of making variable AI spend forecastable, attributable, and tied to outcomes. It is the same discipline that emerged for cloud a decade ago, applied to inference. It is not a cost-cutting initiative, and framing it as one usually kills the deployment.
How do you forecast AI spend accurately. Forecast from three variables rather than headcount. Volume of tasks per period, tokens per task measured from production at p50 and p95, and your blended rate across the model mix you actually route to. Size the buffer to p95 rather than average. Teams doing this land within ten to fifteen percent.
What is the first thing to fix in AI cost management. Attribution. Tag every model call with team, workflow, and use case. Without attribution you have one large number and no way to act on it. It is the highest-return control and every other control depends on it.
