AI · Idee
It's not the model that's expensive
In every AI cost estimate I see, token costs sit at the top. They are the smallest line item. What gets expensive is changing the work behind them.
In four sentences
- The model provider's invoice is usually the smallest part of the total cost.
- The biggest line item is the work on the data — and it comes due before anything works.
- The second biggest is people changing how they work. That time never appears in a quote.
- Model prices fall reliably. Data quality and process clarity do not get cheaper by waiting.
The bill you see
Token prices are pleasantly concrete: they fit on a page, they go into a spreadsheet, they fall every year. Which is exactly why they sit at the top of every business case — and obscure the line items that actually matter.
In the projects I help build, the model's share of total costs is almost always in the single-digit percent range. The rest is spread across things nobody likes putting on a slide.
„AI projects rarely fail on the price per token. They fail on data nobody wanted to clean up."
The bill that counts
First, the data: gathering, cleaning, linking, clarifying rights, keeping it current. This work comes due before anything works, it is unspectacular, and it cannot be skipped. Skip it anyway and you build a system that shines in demo mode and guesses in everyday use.
Then the people: an AI-supported process is a different process. Roles shift, some checks disappear, new ones appear, someone has to review results and rule on exceptions. That time is in no quote and is still the second biggest line item. And finally, operations: monitoring, failure paths, model changes, regression tests — everything that turns a prototype into a system.
What follows
If you are waiting for cheaper models, you are waiting on the wrong line item. Model prices fall anyway — data quality, clean processes, and practiced teams only emerge once someone starts. That is the real head start, and it cannot be bought.
The most honest business case for AI therefore has two numbers: what a case costs today, and what it costs after the changeover — including the people who keep checking. Everything else is model-price folklore.
Questions about this
What does using AI really cost?
In practice the model's share is usually in the single-digit percent range. The bulk is data work, process redesign, and operations — monitoring, failure paths, model changes.
Is it worth waiting for cheaper models?
Rarely. Model prices fall on their own; data quality and process clarity do not. Waiting costs you exactly the learning time that makes the difference.
How do you calculate an AI business case honestly?
Cost per case before against cost per case after — including the human review that remains and the one-off data work.
