Tokens Are Capital, Not Payroll: How to Budget AI

Kris Van Meter · July 2026 · Guide

The line item says "AI subscription." Renewal comes up, the CFO asks what it bought, and the honest answer is a thousand chat sessions nobody can point to. The initiative dies in review, and the review was right to kill it. The mistake happened months earlier, when intelligence got booked as one recurring cost instead of what it actually is: two different kinds of money.

Two kinds of spend

Design-time spend is capital. You put the smartest model you can buy on the problem, once: mapping the workflow, writing the rules, building the templates, defining the checks that catch bad output before a human sees it. What you own afterward is an asset, the balance-sheet kind. Investment, not expense. It runs next week without being rebuilt, and the correction you make in month two is still enforced in month twenty.

Run-time spend is operating expense, the income-statement kind. Once the asset exists, execution is mechanical. Fill the template, apply the rules, run the checks. Mechanical work belongs on cheap models. As of this summer, commodity models run one to two dollars per million input tokens while flagship models price output at twenty-five to fifty. That spread is the entire budgeting strategy: expensive intelligence designs the system, inexpensive intelligence operates it.

	Design time	Run time
What you're buying	The system: templates, rules, verification, routing	Executions of the system
Model tier	The smartest available	The cheapest that passes your checks
How often	Once per workflow, revisited when the business changes	Daily, scheduled, unattended
What it should cost	Real money, budgeted like software development	Cents per run, falling over time

Why the accounting matters

Take a proposal generator. Designing it costs real frontier-model tokens. The drafting rules, the formatting standards, the checks that keep pricing errors and off-brand language out of a client's inbox. Every proposal it renders afterward costs cents, and the hundredth costs what the first did.

Book all of that as a subscription and every dollar runs through the income statement, a recurring cost with fuzzy returns, which is exactly the kind of line a disciplined CFO cuts. Book the build as the investment it is and it earns balance-sheet questions: what asset did we create, and what does it return per run? The subscription framing can't even ask that question.

The trap on either side

Overspending has a signature: expensive models doing daily work by hand, forever, because nobody was forced to build the system. Usage charts climb, everyone looks busy, and none of it compounds. Unlimited budgets produce this reliably; the bill becomes proof of effort instead of a prompt to engineer.

Underspending has a signature too. Refusing frontier spend at design time means a cheap model with no rules around it, and you pay for that decision every single run: rework, spot-checking, the quiet erosion of trust that ends with someone doing the job manually again. Saving on the architect buys a permanent supervision bill for the crew. Spend up where it compounds. Spend down where it repeats.

What to do with your next budget cycle

1. Split the AI line into build and run. Two lines, two questions at review.
2. For anything executing weekly or more, ask which cheaper model passes your checks. If there are no checks, that's the real finding.
3. For anything being designed, use the best model available and stop feeling guilty about it. That spend amortizes.
4. Track cost per run on your top three workflows. It should fall as the market's commodity tier gets cheaper, without you touching a thing.

Owners get this instinctively because it mirrors every other capital decision they make: the truck is capex, the fuel is opex, and nobody evaluates the truck by the fuel bill. In our Discovery engagement, drawing this split for your operation is one of the first things on the table. Whether you do it with us or on your own, do the split before the next renewal, because the subscription framing loses that meeting every time.