The model is the cheap part.

“Understanding AI” series, note 7 of 12. The previous one gave the order of the steps. This one puts a price on each of them, and on what is not in the proposal.

First day of the month, a fiduciary firm in St. Gallen, twenty employees. The controller opens the invoice from the provider of the AI model they have been using for three months to read and classify client documents: CHF 61. She stares at the figure, because the project cost CHF 24,000 in consulting and development, and because the partner who championed it told the board that “AI is expensive”. Sixty-one francs. So where is the money?

The scene is imaginary, with imaginary figures. The proportion is not. In almost every project I have seen, the model is the smallest item, and the controller’s surprise has an explanation in seven parts. I give them to you in order, with the public prices you can check, and then we do the return calculation with an example you can copy with your own numbers.

You pay for tokens, for hours and for attention. Tokens are cheap. Hours appear in the proposal. Attention is put on no invoice by anyone, and it is what decides the return.

Item 1. The model, by the token.

You will remember from note 2 that models read and write in tokens, and that the provider charges per input token and per output token. The prices are public. Anthropic’s, in October 2026 and per million tokens: its mid-range model, Sonnet 5.5, costs $2 for input and $10 for output; the small one, Haiku 5.5, $0.10 and $0.50; the large one, Opus 5.5, $4 and $20; and the most capable, Fable 5.1, $10 and $50. In addition, whatever repeats in every call, such as the instructions and the fixed documents, can be cached and is charged at a fraction, between 5% and 10% of the normal price, when it is reused. Other providers have similar tables.

Translate that to your company. A two-page invoice is in the order of a thousand tokens. Reading it and extracting its data with the small model costs a fraction of a cent; with the mid-range one, a few tenths of a cent. Ten thousand documents a month come to a few tens of francs with the mid-range model, and under ten with the small one. The controller was right to be surprised: at that price, the model is never what makes a project expensive. What makes it expensive is everything else.

Two nuances worth noting. First, the price per token has been falling for years and versions change; what does not fall is the number of tokens your process consumes, so the useful figure is “tokens per case”, not “dollars per million”. Second, language: as I saw in note 2, the same text in German, French or Italian turns into more tokens than in English, and therefore costs more. It is a percentage, not a multiple, but in a system that serves four languages it is worth measuring.

Item 2. Licences, per person.

The other way of paying for the model is the subscription: so many francs per person per month for a general-purpose assistant. It is the simplest form and the most wasted, because you pay for the person, not for the use, and in the clinic from note 6 sixty licences were being paid for four users. My rule: subscriptions for the general use of those who really use them, and the model by the token for automated processes, where you pay exactly for what is processed.

Item 3. The build.

Here is most of the fiduciary firm’s CHF 24,000. Connecting the model to your systems, building the workflow from note 5, setting up retrieval over your documents, programming the checks between steps, running the fifty cases, doing shadow mode. It is people’s work, it is charged by the hour or at a fixed price, and it usually represents between half and three quarters of the first year’s cost. It is also the item that varies most between providers, and the one in which a vague proposal hides the most surprises. Ask for it broken down by step.

Item 4. The data.

Step 2 from the previous note, cleaning and ordering the documents, is done by your people, and nobody puts it in a proposal because nobody invoices it. It costs afternoons. In companies with years of duplicated files it costs weeks. It is the most underestimated item and, when it is skipped, the one that turns a fast system into a fast and wrong one.

Item 5. Attention, by the hour.

This is the item that decides the return and the one that appears on no invoice. A system with risk-proportional review needs people to review. If the system processes a thousand invoices a month and 15% go to a person, that is one hundred and fifty reviews; at four minutes each, ten hours. Those hours are subtracted from the saving. A provider who presents the gross saving, “it eliminates 80 hours a month”, without subtracting the review is not lying to you; they are telling you half.

There is a second, subtler form of attention: the one the system demands of its users so that they do not accept wrong proposals. You will remember the jagged frontier from note 4. A team that accepts everything the model proposes is very productive until the day the error reaches a customer. The training from step 8 and the minutes of checking every day are a cost. A small, permanent cost that buys most of the value.

Item 6. Maintenance.

Models change version and are retired, documents expire, providers’ screens move. Someone has to run the fifty cases again with every change, update the archive and fix what breaks. In my experience it is between 15% and 25% of the build cost per year, and it is the item that explains the most abandoned projects: the system nobody maintains degrades in silence, one day someone stops trusting it and goes back to doing it by hand. When that happens, the money from item 3 has been lost in full.

Item 7. What is not seen.

Three costs remain that are rarely named. The first is the cost of changing provider: if the whole system depends on the particularities of one model, the day that model is retired or goes up in price, you pay. That is why I insisted in note 5 on open connection standards. The second is the energy and environmental cost. Google published in August 2025 its measurement of a median text query to Gemini: 0.24 watt-hours, 0.03 grams of CO2 equivalent and 0.26 millilitres of water, with a 33-fold reduction in energy in one year. These are the provider’s own figures, with its own methodology, but they support one conclusion: for a service company, the environmental cost of AI is today far below that of its travel, and it is worth knowing so as to fall neither into alarm nor into indifference. The third is the cost of not measuring: a system nobody measures has no return, it has faith.

Where the return comes from.

The return arrives by four routes, and it is worth knowing which one yours comes by, because they are measured differently. Hours: the work that stops being done by hand, measured as in the study of the 5,179 agents from note 4, 14% on average and 34% for new staff. Errors: what a badly served order costs, an invoice booked twice, a missed deadline. Speed: collecting fifteen days earlier because the reminder goes out on its own, answering a request for a quote in an hour instead of three days. Capacity: serving more without hiring, which is the hardest return to measure and the largest when it exists.

And it is worth knowing what those who have already paid say. In May 2025 IBM published a survey of 2,000 chief executives: only 25% of their AI initiatives had delivered the expected return, and only 16% had been extended across the whole company. 64% admitted having invested in some technologies before properly understanding their value, for fear of falling behind. And half said the accumulation of disconnected technologies was a brake. Each of those figures has its partner among the seven items above.

The sum, with an example.

I invent a company to do the sum, and I ask you to replace each number with your own. A distributor receives 400 supplier invoices a month. Each one costs twelve minutes today between reading it, finding the order, checking and booking: 80 hours a month, which at a cost of CHF 45 an hour is CHF 3,600. The proposed system costs CHF 18,000 to build, CHF 40 a month for the model and CHF 300 a month for maintenance.

• Gross saving: CHF 3,600 a month.
• Review: 15% of invoices go to a person, 60 a month, at four minutes: 4 hours, CHF 180.
• Net saving: 3,600 minus 180 minus 340 for model and maintenance: CHF 3,080 a month.
• Payback: 18,000 divided by 3,080: a little under six months, if the system really processes 100% of the invoices from the first month.
• With 50% adoption, because half the team keeps doing it by hand “just in case”: the net saving falls to about CHF 1,400, and the investment takes thirteen months to pay back.
• Without maintenance, the system degrades, the team stops trusting it after a few months and the sum goes back to zero with CHF 18,000 less.

Notice what decides the result: it is not the price of the model, which is CHF 40, nor even the build. It is adoption and maintenance. The two things that are not in the provider’s proposal and are in the hands of whoever is in charge. The sum above is the one I ask to be done before signing, with three adoption scenarios, and the one I do again with the real numbers in the third month.

1. Before asking for a price

   Measure today’s process

   Cases a month, minutes per case, cost per hour. Without this there is no return to calculate, only promises to compare.

2. With the proposal in front of you

   Ask for the seven items separately

   Model, licences, build, data, review, maintenance and exit. Whatever is missing, ask about it. Whatever they do not know, write it down.

3. With the proposal in front of you

   Do the sum with three adoption rates

   100%, 70% and 40%. If it only pays off at 100%, it does not pay off.

4. In the third month

   Repeat the sum with the real numbers

   Cases processed, cases reviewed, real minutes. Compare with the proposal. Decide the second process on that basis.

How I approach it.

My proposals carry the seven items separately, with the review cost estimated from the fifty cases and maintenance as a monthly line, not as a surprise. The price of the model is calculated in tokens per case with your real documents, and shown with two providers. And the return calculation is done with three adoption rates, because adoption is not controlled by the provider: it is controlled by you, and it is the item worth the most.

Take the last AI proposal you received and count how many of the seven items appear separately. Then look for the one that almost never does: the hours your team spends checking what the machine does.

If it is missing, do the sum yourself, even on a napkin. And ask yourself what would happen to the promised return if only half your team used the tool. It is the item nobody can sell you, and the only one you decide.

Next note: why pilots fail. The five causes researchers find again and again, the symptoms visible from the top in time to correct them and the case of a company that was everyone’s example and then reversed course.