Token economics for managers: the bill explodes in month three

There is a pattern I see in almost every organisation.

Month one: a small pilot, a bill of tens of dollars, everyone pleased. Month two: limited expansion, hundreds, no problem. Month three: the project worked, everyone started using it, and the bill is ten thousand.

And the painful part: the bill exploded because the project succeeded. That is the pattern nobody explains before you sign.

Why this differs from every piece of software you have bought#

Everything you are used to buying is priced per seat or per licence: a fixed number, paid whether it is used or not. Your marginal cost for additional use is zero.

AI is priced by consumption. Every use costs, and every success increases use.

The line items that never appear in the vendor's deck#

The fourth row is what flips the business case. Any system with real consequences needs someone reviewing its output, those are human hours with a cost, and their absence is the most common error I see in feasibility studies.

The simple model that is enough#

You do not need a complicated spreadsheet. You need four numbers.

Operations per month. How often will this be used? Be optimistic, because success is the expensive scenario.

Cost per operation. Divide the pilot's bill by the number of operations it ran. That is more accurate than any theoretical estimate.

Growth factor. If it works, how much will usage multiply? Three times is conservative.

Human review cost. How many minutes does a person need per output, and what is their hour worth?

Three levers that cut cost immediately#

Route by task. Do not run everything on the strongest model. Simple work on the economy tier usually saves more than half the bill.

Caching. If part of your prompt is fixed and repeated, many providers price repeated content cheaper. This matters for anyone sending the same instructions thousands of times a day.

Control output length. Output costs more than input. Specifying "answer in three sentences" is not a style preference. It is a cost decision.

The honest part#

This article focuses on cost, which is only half the picture.

A system costing a thousand dollars a month and saving eighty hours of work is cheap, and anyone rejecting it on the number is as wrong as anyone approving it with no arithmetic. The question is not what it costs but what it costs relative to what.

And harder still, some benefits do not measure easily: a faster decision, an avoided error, an employee who did not resign out of boredom with a repetitive task.

In closing#

The bill explodes on success rather than failure, which is the strangest fact about this technology and the most neglected in planning.

Start today with one question to your team: what does one operation cost in our system? If that has no numerical answer, you do not have a cost model. You have a hope.

And if you want to build that model with your team on a real project, it is one of the first sessions in the leadership programme.

Common questions#

Why does an AI bill explode on success?

Because pricing is by consumption rather than licence. Traditional software has zero marginal cost after purchase; here every additional use costs money, and success means more use.

Which line item do managers usually forget?

Human review of output, which can be 20 to 40 percent of true cost. Any system with consequences needs reviewing, and those are human hours with a price.

How do I estimate cost reliably?

Divide your pilot's bill by its number of operations to get cost per operation, multiply by expected volume, multiply by three as a success scenario, then add human review.

What is the fastest way to reduce the bill?

Routing by task: the economy model for simple repetitive work and the strong one for anything needing reasoning. That alone usually saves more than half.

Is cost the only criterion?

No. A system costing a thousand dollars and saving eighty hours is cheap. The right question is not what it costs but what it costs relative to what, and some benefits do not measure easily.