Last week, a director asked me how much his organisation spends on AI each month. He thought it was one number. It was 14, spread across invoices for ChatGPT, Claude and twelve other tools. I understood his surprise.
That is not an accounting detail. It is where every AI cost problem begins.
You cannot control what you cannot see. And with AI, almost nobody sees the total. Usage is hidden in separate subscriptions, SaaS tools with an AI button and API keys a team once requested. Added together, it is often a substantial amount. But nobody adds it up, so nobody manages it.
What FinOps is, and why AI needs it again
FinOps is a discipline that emerged around cloud costs in recent years. The idea is that cloud usage is unpredictable because you pay per use, and that usage changes continuously. The solution is not to set a budget once, but to have technical teams, finance and management jointly control what is being consumed. Visibility, allocation, control. In that order.
AI has exactly the same problem, only more acutely. You pay per token, consumption scales with every use and prices move. The FinOps Foundation (opens in new window) has therefore added an explicit branch for AI. Not because AI costs are fundamentally different from cloud costs, but because they are even less predictable and even more deeply hidden.
The heart of the matter is that AI costs only become manageable when three things are in order. I will take you through them as I approach them in practice.
Step one: visibility, or that one number
Before you can manage anything, you need to be able to see it. That sounds obvious, yet it is almost always skipped.
Visibility means knowing how much your organisation spent on AI this month as a single number. Not per tool, not per team, just the total. In practice, that number does not exist because spending is spread across different invoices, departments and contracts. The AI in your customer service software, marketing's ChatGPT subscription, the API key for that one project: nowhere is it all brought together.
The first assignment is therefore tedious and crucial. Collect all AI spending from the past three months in one place. Include SaaS tools with an AI component, because they often hide more cost than your direct API usage. The first time you see that number, it will usually shock you. That is good. Shock is the beginning of control.
Step two: allocation, or where it comes from
One total figure is a starting point, not the finish. As soon as you see the total, you want to know where it came from.
Allocation means being able to connect costs to a team, an application or a customer. Which project consumes the most? Which department? Which application keeps running in the background even though nobody looks at it anymore? This is the point where AI costs change from a mystery into something manageable.
This is where an uncomfortable finding appears. Research by the MIT NANDA initiative (opens in new window) shows that more than half of generative AI budgets went to sales and marketing tools, while the greatest ROI was found in the back office. The money went to visible, attractive applications; the returns were in the tedious processes. Without allocation, you do not see that. With allocation, you see exactly where you spend a lot and get little in return.
Step three: control, or limits you are willing to set
Visibility and allocation are diagnosis. Control is treatment.
Control means setting limits based on what you see. Quotas per user. Limits per application. A deliberate choice about which model runs where, because you do not need the most expensive model everywhere. And perhaps most importantly, one person who is ultimately accountable for the AI budget. One person, not a committee.
That choice of model deserves a separate note because it often offers quick gains. Not every task needs the most powerful and expensive model. A simple summary can run on a cheaper model, while you reserve the top model for work where it really matters. Anyone who uses the heaviest model by default is paying for capacity most tasks do not use. It is the same logic as not driving a lorry to the supermarket. A large part of control is putting the right model in the right place.
Let me immediately explain what control is not. It is not banning AI. Blocking without visibility pushes usage into the shadows, leaving you with neither control over costs nor sound governance. Control is the opposite: you make usage visible and set boundaries around it, so people can continue working while you know what it costs.
The trap I see most often
One mistake occurs so often that I want to address it separately. Organisations respond to invisible AI costs by restricting access. No more ChatGPT, no separate tools, everything through one central channel. Cost problem solved, they think.
The opposite happens. If you block AI without providing a good alternative, usage does not disappear; it disappears from view. People use a private account, their phone or a tool nobody has approved. Your bill falls, but only because you can no longer see it. At the same time, you create a governance problem because data is now going to places you do not know.
Control and blocking are not the same. Control makes usage visible and sets boundaries around it. Blocking makes usage invisible and builds a wall that people walk around. The first gives you control. The second gives you a number that looks better than reality. Anyone who wants to manage costs should make sure people have good, paid tools, because that is what keeps usage visible.
Follow me on LinkedIn
The order is not optional
These three steps are a sequence, not a menu. I see organisations start with step three: they set limits before they know what is being consumed. That is steering with your eyes closed. You randomly restrict something and hope it was the right place.
The correct order is to measure, allocate and only then control. First see what is there, then understand where it comes from, and only then intervene. After that, repeat the process, because AI usage is not a photograph but a film. What was correct last quarter is no longer correct this quarter because new tools appear, usage grows and prices move.
Why this is urgent now
You might think that your AI bill is still small and this can wait. That is exactly the mistake.
Adoption is growing rapidly. According to the Stanford AI Index 2025 (opens in new window), 78% of organisations used AI in 2024, up from 55% a year earlier, while the use of generative AI more than doubled. Your bill may be small now, but it grows with usage, and usage grows faster than you think. Introducing FinOps after the bill has already spiralled out of control is mopping the floor while the tap is running. Introducing FinOps now, while things are still manageable, is the difference between staying in control and being shocked later.
I saw that difference up close when AI budgets at large technology companies spun out of control. Not because AI failed, but because programmers used far more than budgeted and nobody saw it in time. The lesson was not "use less AI". The lesson was "make it visible before it becomes large".
FinOps is not a tool, it is a habit
Tools exist that measure AI usage, along with dashboards that break down your costs, and they help. But I want to clear up one misconception: FinOps is not software you buy and switch on. It is a habit you introduce.
The habit is to make costs a standard part of every conversation about AI. Someone asks how much each new application will consume. The AI budget becomes as natural a part of the planning and control cycle as every other budget. Someone checks every quarter whether usage still matches the value. A dashboard without that habit is a beautiful screen nobody looks at. The habit without a dashboard still works, only more slowly.
So do not start by choosing a tool. Start with that one number and the agreement that someone is responsible for it. The tool comes later, once you know what you want to measure. The reverse rarely works.
The self-check you can do today
You do not need to build a FinOps department to know where you stand. Go through these questions.
- Do you know what your organisation spent on AI this month as one number?
- Can you break that number down by team or application?
- Is one person ultimately accountable for the AI budget?
- Are there quotas or limits per user or application?
- Do you know which application consumes the most and whether it also delivers the most?
- Do you have a scenario for a supplier doubling its price?
If you cannot answer more than two of these questions, you are steering by instinct. That is not a disaster, and there is no shame in it. But it is the first thing you should repair before rolling out more AI.
The director from my opening anecdote had fourteen invoices and no total. A month later, he had one number, one person responsible and a list of three applications that cost a lot and did very little. He did not spend less on AI. He spent it more consciously. That is the whole point of FinOps for AI, and it is far less complicated than the word suggests.
The broader framework, from calculating the full cost to measuring returns, is in my guide to AI costs and ROI.
Sources
* FinOps for AI as an official branch of the FinOps framework: finops.org (opens in new window) * Budgets went to sales and marketing while ROI was found in the back office (MIT NANDA, via Fortune): fortune.com (opens in new window) * Growth of AI adoption among organisations (Stanford AI Index 2025): hai.stanford.edu (opens in new window)