AI FinOps: The Financial Discipline For The Intelligence Economy

For most companies, the AI bill keeps climbing every month and never shows what the business gets for the money. Sandeep Shilawat, a Partner and Hybrid Cloud CTO at IBM, thinks they are tracking that spending without any way to control it. His answer is FinOps, the cost practice cloud teams built during the last budget crisis, applied to AI and eventually handed to AI to run.

That view comes from years of guiding organizations through each technology shift, first to cloud and now to AI. At IBM, he leads the AI and Automation Practice for the federal market. His recent book, Trustworthy AI: Red Teaming, Risk and Architecture of Secure Intelligence, makes the case that trust in a system has to be proven and enforced. He wants the same evidence behind what companies spend on it.

"Intelligent efficiency should be measured in terms of outcomes. Right now, it is strictly viewed as an operational cost," Shilawat said. He sets the problem out as an equation.

AI ROI = Business Outcomes ÷ (Infrastructure + Models + Tokens + Operations + Governance + AI Security).

Every item on the bottom half shows up on an invoice. The top half is harder, because the industry has no benchmark for what a given piece of AI work is worth. Generating code eats through tokens faster than drafting plain English, and neither carries an agreed price for the value it returns.

• Token math: The rates behind AI spending are not settled. Providers bill by the token, the small chunks of text a model reads and writes, and those prices come out of a young, capital-intensive market where a few well-funded companies compete hard for share. Enterprises that build multiyear plans around today's rates are working from a number that has not held still. Model choice changes the total as well. The same task can cost several times more on a large model than on a small one. "While these models are useful, using them for basic tasks is like using a Rolls-Royce to get groceries," Shilawat said.

The cost is already changing how companies behave. Some cancel their AI licenses after developers burn through a year's budget in a couple of months. Others cap usage or send requests through an approval queue, and the delay lands on the people the tools were meant to help.

• Cloud hangover: The habit of overspending comes from the cloud, where paying only for what you use makes computing power feel endless. That arrangement teaches teams to add capacity freely and settle up later. Applied to AI, the same reflex lets spending run without check. 37signals, the maker of Basecamp, left its cloud provider after a yearly bill above $3 million and moved the work onto hardware it owns. AI played no part in the decision. The spending was cloud economics, growing a little at a time in a system where each increase was easy to approve, and the total only registered once someone added it up for the year. "The utility model is an OpEx model that emerged during the cloud era, and it psyched us," Shilawat said.

• Fighting fire with FinOps: Cloud teams have long used a practice called FinOps to track and control what they spend on computing. The same discipline applied to AI runs into a problem of pace, because AI costs move by the hour. Monthly reviews and spreadsheets cannot follow that. Software that watches costs as they land can show what a periodic review misses, giving finance a running account of where the money goes. "While organizations are applying FinOps discipline to artificial intelligence, there is another part of it, using AI to do FinOps," Shilawat said.

An agent keeps working after the person who launched it has moved on, calling models, pulling data, hitting APIs, and holding compute until its task finishes. Every agent in production draws on the budget with nobody at the keyboard. Deciding what an agent may do is also deciding what it will cost.

• A thousand cuts: Companies are adding AI agents faster than they can govern them. Nobody reviews most of what those agents do. An agent that misfires bills the company for work it got wrong. Across thousands of them, those unchecked actions become a serious risk. "Every bad or dysfunctional agent is a cut, and you do not want death by a thousand cuts. You need agent governance, agent management, and a centralized control plane," Shilawat said.

• A garden of models: Standardizing on a single AI model is the next mistake in the making. Tasks differ in complexity, latency, security, and the quality the output has to reach, and those differences decide what each one costs to run. Sending a workload to a model that meets its requirements and nothing more holds the bill down, and that only works when the cost shows up while the work runs. "This is what I call a model monoculture issue, relying on a single model for everything. I do not think that will survive," Shilawat said. Models running side by side each build their own understanding of the business, and none of them share it. A company operates on one connected picture of itself, and no model holds more than a piece of it. How to give them a shared view remains unsolved.

The cost of a single unit of intelligence keeps falling. Spending rises anyway, because each drop in price makes work worth attempting that nobody would have funded a quarter earlier, and consumption grows faster than the price comes down. That is the condition finance has to manage as the intelligence economy takes shape, and no team can track it by hand. "The science of FinOps itself needs to be handled by AI, which means companies must adopt AI technology to generate real-time data for financial analysts," Shilawat said.