Price the Work, Not the Tokens.
Watch the video version:
Late in October, I sat at a roundtable in New York with technology executives, and one of them described an AI bill their interns had run up: somewhere between $30,000 and $40,000 in API spend. Another observed that people from engineering backgrounds consistently spend fewer tokens and get more out of them. These are their accounts of their own organizations, and I have not checked the numbers. The question underneath both was the one every budget owner asks after a bill like that: how do we stop it from happening again?
If you are writing a 2027 budget, you have three answers ready. Cap it: token limits per seat, cheaper models by default, and approval for anything bigger. Watch it: a dashboard of spend per person, with the top spenders flagged as waste. Or bring it in-house: run the models locally, so the meter stops. Each one is reasonable, and each one is aimed at the same number, the size of the bill.
None of them can see what the bill bought. A cap stops a $40,000 month that shipped something worth ten times that exactly the way it stops a month that shipped nothing. The dashboard ranks both people by spend, so the person whose spend paid for itself sits next to the person whose spend did not. Running models in-house quiets the meter and leaves the question where it was; one leader at the table called it over-optimizing the wrong thing, when the real gap is education. It does not matter how much you spend if you make more money because you spent it. I made a version of this argument about headcount in A Hundred Dollars of Tokens Does Not Buy a Hundred Dollars of Value; it holds one level down too, for each person's own use.
The people spending cannot see it either, and I include myself. At that same table, I suggested an agent that watches what you do with AI and tells you which model each kind of work needs. That week I built my own. It reads every session I run, sorts them by type, and recommends a model for each. I am on a subscription, so the costs are not a bill; it prices everything at API list rates to make the comparison. My use came to about $10,000 as I ran it, against about $8,000 on the models it recommends. One read-only scan I run every workday had been sitting on the most expensive model, about $100 of use, for work the smallest model would do for about $4.
I did not know. What surprised me was how little of my work needs the top model. The sessions where I want an argument do. Most of the rest is structured work with a clear spec, and some of it is mechanical. My first reaction was relief that I am not on API pricing. My second was that the interns are the version of me with nobody showing them the pattern. The engineers who spend less learned the sort somewhere else first. I have not had to change anything yet, because I am inside my plan's allotment, and that is the reason to have the report now: it shows the pattern while the pattern is still cheap.
So price the work, not the tokens. Sort the AI work your team does by type, and decide which model each type needs. Judgment work, where the output carries your name or your decision, runs on the top model. Structured work with a clear spec runs on the middle one. Mechanical work that follows a fixed prompt runs on the smallest. Then measure each type by the result it is supposed to produce. A cap prices tokens and teaches nothing. A sort prices work, and anyone on your team can learn it, because it is the same sort my agent runs on me every month. It is the move from Monday's piece on AI chargebacks, one level down: the cost means nothing until someone writes down the result it is for.
Take the three people on your team who spent the most on AI last month. Next to each name, write the result their spend produced. If you can't, ask them. If they can't either, you have a teaching gap, not a spending problem, and a cap will hide it. Then sit down with one of them and sort a week of their AI work together, by type, with the model each type should run on.
Before you cap anyone's AI spend, take the LeadershipOS™ Scorecard. It is free: twelve questions across Decision Load, Escalation Patterns, Cadence and Meetings, and Role and Ownership Clarity that show which part of your system is carrying the gap the bill exposed: https://theleadershiposbook.com/scorecard
