Every CFO I talk to can defend their software stack to the dollar. Salesforce has a number. Workday has a number. The cost per signed contract in DocuSign — there's a number, and somebody owns it.
Then there's AI. The line item that doubled this year. The one nobody can defend in a board meeting.
For a while that was just a vague discomfort. This spring it turned into headlines. Microsoft pulled most of its internal Claude Code licenses after heavy usage ran $500 to $2,000 a month per engineer. Uber's CTO said the company burned its entire 2026 AI coding budget in four months — roughly three times over plan. That wasn't a spreadsheet error. The consumption curve of agentic tools simply bent far steeper than the seat-based budget assumed it would.
Here's the pattern worth sitting with: none of these companies failed at adoption. They succeeded at it. People used the tools, usage exploded, and that is exactly what blew up the number.
The trap was baked in from the start. Flat per-seat pricing taught everyone to treat AI as a fixed cost. It isn't. The real cost is metered underneath, per token, and it scales with every win — the more value a team gets, the bigger the bill, and almost nobody is watching the meter in real time.
And cheaper tokens won't save you. Gartner expects inference on a frontier model to cost about 90% less by 2030 than it did in 2025. In the same window, Goldman expects agentic AI to drive something like a 24-fold jump in token consumption. Price per unit falls; volume runs away from it. EY now pegs a single orchestrated agentic interaction at roughly 30 times the cost of a simple one. Gartner's read on the next two years is blunt: at least half of GenAI projects will overrun their budgets — not from bad luck, but from poor architecture and no operating discipline.
So what does the discipline actually look like?
Total spend is a vanity metric. The decisions live one level down — spend per team, per workflow, per outcome. That's what tells you which licenses to expand, which to cut, and which workflows are genuinely leveraged versus just expensive.
The unit of account changed. The seat was the unit of the SaaS era. The task is the unit of this one. You cannot forecast a per-seat line for something that bills per outcome, and the finance teams that keep trying will keep getting blindsided.
You don't need to spend less. You need to spend legibly — every dollar traceable to a team, a task, and a result, watched as it happens, not discovered when the invoice lands.
This is why FinOps teams now rank AI cost management as their top forward-looking priority, and why Deloitte shipped a CFO guide to token economics — a topic that didn't exist on a finance radar eighteen months ago.
The companies that win this won't be the frugal ones. They'll be the legible ones. When the invoice lands, there are no surprises, because they were watching the meter the whole way down.
You can't manage what you can't measure. The AI line is about to teach an entire market that lesson, one surprise invoice at a time.
— Tobias