A finance director at a 400-person software company pulled up the OpenAI and Copilot line items on last month's technology spend review and found they had grown 40% in ten weeks. Nobody on the team could say why. Headcount hadn't changed. Usage, as far as anyone could tell, hadn't obviously changed either. And the per-token price for the models in use had actually gone down twice in that same window.
That contradiction, prices falling while the bill climbs, is the shape of AI spend right now. It's also why a growing number of finance and technology leaders have started treating AI cost governance as its own discipline, separate from the general SaaS and cloud spend problem it grew out of.
What AI cost governance actually means
AI cost governance is the practice of tracking, forecasting, and controlling what an organization spends on AI models and the tools built on top of them. The unit that matters is the token, not the seat. That's the whole shift.
A SaaS contract is priced by seat count, decided once at signing and reviewed once a year at renewal. An AI bill is priced by consumption instead: how many tokens a model processes, how many API calls an application makes, how many times an agent loops before it returns an answer. Nobody signs a contract for 40 million tokens a month. The number just happens, based on how employees, applications, and increasingly autonomous agents use whatever models are available to them.
CostRoom has written before about how AI tools enter company budgets the way SaaS did five years ago, skipping procurement thresholds because a single developer with a credit card can turn on an API key. AI cost governance picks up after that point, once the tool is already in use. It comes down to one question: whether anyone can explain, after the fact, why the bill looks the way it does. Most companies can't.
This sits next to, not inside, CostRoom's broader guide to AI, SaaS, and cloud spend management. That guide covers AI as one more spend category among several a finance team has to watch. This piece is about the part that behaves differently: token pricing, model choice, and the specific ways an AI bill breaks the forecasting habits most finance teams already learned from managing SaaS.
Why falling token prices haven't lowered your AI bill
Deloitte's technology practice put this plainly in a January 2026 piece: AI is now the fastest-growing expense in corporate technology budgets. Some firms report it now consumes up to half of their IT spend. The same piece cited Deloitte's own 2025 US Tech Value survey. Only 28% of finance leaders reported clear, measurable value from their AI investments, even as nearly half expected it would take up to three years to see ROI from basic AI automation. The gap between spend and proof is wide.
McKinsey's research into enterprise AI spend, published in July 2026, found the same pattern from a different angle. As organizations move from isolated AI use cases to spend across the whole enterprise, AI spend increases nearly fourfold. Ninety-three percent of the companies McKinsey surveyed report exceeding their AI budgets in the process. Almost none saw it coming.
Put the two findings together and the paradox stops looking like a paradox. Per-token prices have fallen consistently as model providers compete on price. But consumption, more users, more workflows, more autonomous agents making their own model calls, is growing faster than the price is falling. A cheaper unit multiplied by a much larger number of units still produces a bigger bill. McKinsey's own August 2026 survey found that one in five companies now say AI operating costs, including token costs, are already constraining how much they can use it. Most still plan to spend more next year regardless.
What your AI spend actually looks like across every provider
CostRoom maps AI vendor spend, token consumption, and where AI costs are hiding inside other renewals, before any optimization work begins.
Three things breaking your AI cost forecast
Most finance teams built their forecasting habits on SaaS: a seat count, a renewal date, a predictable line item. AI breaks all three assumptions, for three distinct reasons.
Usage itself is wildly unpredictable. The same task, run twice, can produce dramatically different token counts, depending on which model answered it, how many tool calls an agent made along the way, and how long the resulting chain of reasoning ran. Stanford's Digital Economy Lab, in research McKinsey cites directly, found token usage can vary by up to 30 times when a model executes what looks like the same task. Agentic workflows make this worse. One user request can now trigger dozens of model calls instead of one.
AI costs are also increasingly embedded inside SaaS subscriptions a finance team already reviewed and approved months ago. Microsoft 365 Copilot licensing. AI features quietly added into a CRM or a support tool at renewal. A coding assistant bundled into a developer platform contract. None of these show up as a separate AI line item. They show up as a renewal that looks routine, because on paper it is a renewal, right up until the usage-based portion of it starts moving on its own.
And then there's the AI spend nobody approved in the first place. Employees can now build AI-powered workflows and small autonomous agents with almost no engineering effort, which is exactly what makes this category hard to see coming (the same person who once asked IT for a new tool can now just build one over a weekend). A marketing analyst wiring together a weekly report generator. A support team running a chatbot nobody in finance signed off on. This is uncontracted AI spend, and it behaves nothing like the uncontracted SaaS spend most finance teams already know how to find. None of this is anyone's fault, exactly. It's just how consumption-based pricing behaves once dozens of people, and a growing number of agents, are making the calls instead of one procurement form.
What AI cost governance requires that most companies don't have
Four things, in practice, separate a company that has a handle on AI spend from one still finding out about it at the end of the month. None of the four is complicated.
Full visibility into where the spend is actually going: every model provider, every AI feature bundled into another SaaS contract, and every team running its own usage, in one place rather than scattered across a dozen invoices. Unit economics that tie a dollar figure back to something a CFO recognizes, cost per task, cost per interaction, rather than a single vendor total with no way to see what drove it. Clear ownership: someone who reviews AI spend on a fixed cadence, the same way a renewal calendar gets reviewed, instead of everyone assuming finance is watching it because the invoice technically goes through finance. And a standing habit of matching model choice to the task at hand, because a lot of AI spend is simply an expensive frontier model doing work a cheaper one could handle just as well.
None of this requires the company to become an AI company first. It requires treating AI spend with the same discipline SaaS and cloud spend already get, once someone decides to apply it.
Why you don't need to hire a dedicated AI FinOps team
McKinsey's own numbers make the honest case here. Only 20 to 25% of companies it surveyed have mature AI FinOps practices in place. The ones that do report real results: cutting AI costs by 20 to 30% once they actively optimize spend rather than just watching it grow. Building that capability from scratch usually means new hires, in a discipline barely two years old, learning on a live budget while the bill keeps moving. Hiring rarely fixes that fast enough. Few companies can afford to learn that way.
The function matters more than the team that runs it. That's the same point CostRoom has made about cloud FinOps for companies with 200-500+ employees, and it holds here too. Visibility, unit economics, and ownership don't require a headcount line. They require someone who already knows what a normal token bill looks like across enough companies to spot an abnormal one fast.
This is a different problem from the one covered in CostRoom's guide to AI tool governance. That guide, and its own first piece on AI tool risk, are about data processing agreements, the EU AI Act, HIPAA and PCI exposure, and vendor risk. Cost governance and compliance governance often come up in the same meeting. The two rarely overlap. They answer different questions, and a plan built for one will not cover the other.
CostRoom's platform already tracks spend across nine AI tools and counting, including OpenAI, GitHub Copilot, and Anthropic, alongside the SaaS and cloud accounts most finance teams already have visibility into. The AI spend analysis below runs on the same category-manager model CostRoom already uses for SaaS and cloud. It's expert-led, vendor-agnostic, and built to run as a service, not another dashboard someone in-house has to operate. Nothing about that requires more headcount.
Getting started with an AI spend analysis
An AI spend analysis is the starting point CostRoom uses instead of a sales pitch. It maps current AI vendor contracts and usage data across every model and tool in use. It maps token and API consumption patterns over the past several months, where AI costs are hiding inside other SaaS renewals, and how current usage compares to what a right-sized setup would cost against the same tasks. The spend analysis and optimization service covers how this works in full.
None of this requires switching AI providers. It doesn't require pausing anything already in production, either. It requires access to billing data and usage logs, and a short conversation with whoever currently owns the AI tools in question, often two or three people at a company this size rather than one central team.
The output is a specific list. It shows which model choices are costing more than the task requires, which AI features are quietly riding along inside another SaaS renewal, and which usage patterns are worth a policy before the next budget cycle rather than after it. CostRoom's expert-led spend analysis and optimization service runs this the same way for SaaS, cloud, and AI spend, under one category manager instead of three separate conversations.
AI spend is not going to slow down on its own, and per-token prices falling further will not change that. What changes it is someone deciding, on purpose, which model runs which task and why, and reviewing that decision on a cadence instead of at the end of the fiscal year.
The bill will keep moving either way.
The only real choice is whether anyone explains it before finance has to ask.
Start with an AI spend analysis
CostRoom maps your full AI vendor footprint before any optimization work begins.
Frequently asked questions
What is AI cost governance?
AI cost governance is the practice of tracking, forecasting, and controlling what an organization spends on AI models and AI-enabled tools, tied to actual token and API usage rather than a fixed subscription.
Why do AI bills keep growing if token prices are falling?
Per-token prices have fallen as model providers compete on price, but consumption is growing faster than the price is dropping. More users, more workflows, and more autonomous agents making their own model calls add up to a larger bill even at a lower unit price.
Is AI cost governance the same as AI tool governance?
No. AI tool governance covers compliance and risk: data agreements, regulatory exposure, and vendor risk. AI cost governance covers spend: visibility, unit economics, and which model is running which task at what cost.
Do we need to hire a dedicated AI FinOps specialist?
Not necessarily. The functions a dedicated AI FinOps hire would cover, visibility, unit economics, and ownership, can be delivered as an expert-led service instead of an internal role, the same way CostRoom already runs this for SaaS and cloud spend.
What does an AI spend analysis cover?
An AI spend analysis maps current AI vendor contracts and usage, token and API consumption patterns, AI costs hidden inside other SaaS renewals, and how current usage compares to a right-sized setup, before any optimization work starts.
Related reading:
