By Steve Urban
For most of my career, technology was a rounding error next to labor. You argued about headcount. You approved software almost without thinking.
That is changing, and you can hear it in how executives are talking this budget season. CNBC reported earlier this year that some companies were burning through their annual AI budgets in a month or two. Glean CEO Arvind Jain made the point that the tokens versus people conversation is new, because tech used to be a small fraction of operating costs. He also estimated that around 95 percent of enterprise AI tasks still run on the most expensive models, even when cheaper ones would do. And in a SiliconANGLE interview this month, Dell CFO David Kennedy described AI token spend becoming its own budget line that finance tracks like any other operating cost.
So the debate has flipped. A year ago the question was whether we were using AI enough. Now a lot of leaders are asking whether they are spending too much, and some are reaching for hard caps.
I think both questions miss the point.
Let me give the cap crowd their due, because they are not wrong to be nervous. Usage-based costs can grow faster than anyone forecast. Agents running in loops can burn spend with nobody watching. Teams experimenting with no budget discipline is how you end up explaining a surprise invoice to the board. A hard ceiling forces people to think before they spend. As an operator, I respect that instinct.
But a cap is a blunt tool. It treats the most valuable use of AI in your business the same as the most wasteful one. When the ceiling hits, both stop. That is not cost management. That is rationing.
Most companies cannot answer a simple question: what did we get for that spend?
Not how many people logged in. I have written before about why that number misleads. I mean which workflows used it, what those workflows produced, and whether it was worth more than it cost.
If you cannot answer that, you cannot manage AI spend in either direction. You do not know where to cut, and you do not know where to invest more.
I have run multi-site operations, turned around businesses that were bleeding cash, and built companies from scratch. The discipline that works for AI spend is the same discipline that works for overtime, fuel, or materials.
Attach every cost to a workflow. Each meaningful AI use should map to a named process with an owner. Recruiting intake at Riderflex. Service notes at Robo Reliance. Proposal drafts. Whatever it is, the spend lives somewhere specific.
Measure the output, not the activity. Cycle time, error rate, revenue touched, hours freed. Pick the one that matters for that workflow and track it before and after.
Match the tool to the task. Summarizing a call is not the same as working through a complex contract. If the 95 percent estimate is anywhere close, a lot of companies are paying premium prices for routine work. Routing simple tasks to cheaper tools may be the easiest savings you find this year.
Budget by predictability. Kennedy made this point and it is a good one. Steady, repeatable workloads can be forecast, and some can move to fixed-cost setups. Exploratory work gets a smaller, separate pot with a clear review date.
Review it monthly with the owner in the room. Not a dashboard emailed around. A short conversation: what did we spend, what did it produce, what changes next month.
The headline version of this debate is AI versus headcount. Every dollar on tokens is a dollar not spent on a person. I understand why it gets framed that way, but it pushes leaders toward bad decisions.
The right comparison is not tokens against people. It is the full cost of a workflow done one way against the full cost of doing it another way, including the human time to check and fix the output. Sometimes AI wins by a mile. Sometimes the person is cheaper and better. Often it is a combination, and the person's job gets better because the grind goes away.
Leaders who treat it as a straight swap tend to cut people they needed and keep AI spend they did not. Leaders who treat it as workflow economics tend to get both right.
If you are building next year's budget right now, do three things. Pull every AI line item and assign it to a workflow and an owner. Shut off anything nobody claims. Then, for the top three by spend, ask the owner to show you what it produced.
You will probably find some waste. You will probably also find one or two uses worth far more than they cost, and those deserve more money, not less.
That is the job. Not spending less on AI, and not spending more. Knowing what you bought.
Connect with Steve: linkedin.com/in/stevepurban