The moment of truth: year two
The first year of a generative AI project is often easy to fund: a pilot, a few dozen users, a modest bill. The question arises when usage takes off. And it does: according to the France Num 2026 barometer, AI use among French SMEs grew by 14 points in a year. The choice of cost model then becomes decisive.
Three models, three logics
1. The per-user licence
A monthly subscription per seat. Simple to understand, but the cost tracks headcount: rolling the tool out across the whole company multiplies the bill, which often pushes companies to reserve AI for a few people.
2. Per-token billing
Each request is billed according to the volume of text processed. This is the model of cloud APIs. It is ideal for getting started, but the bill tracks adoption: the more your teams use AI, especially agents that chain several steps, the more it climbs, and the harder it becomes to forecast.
3. The flat rate per server
A server, installed on your premises, runs the models. The cost depends on installed capacity, not on the number of requests or users. Usage becomes unlimited and the budget predictable.
What price lists do not show
Comparing listed prices is not enough. Several costs remain invisible when choosing:
- Agents consume far more than a simple question and answer: an agent that searches, reads several documents, computes and then writes can process the equivalent of dozens of requests in one task.
- Document context costs money: with per-token billing, every document excerpt sent to the model to answer is counted, for every question.
- Self-censorship has a cost: when every request is billed, teams hesitate to use the tool, and the expected gains do not materialise.
- Changing provider is often a project in its own right, rarely budgeted at the outset.
Think in cost per use
The right unit of measure is neither the user nor the token, but the business use: what does summarising minutes, answering an HR question or analysing a spreadsheet cost? And how much time does it save? This calculation, made on your real volumes, almost always points clearly to the right model.
In organisations where AI becomes a daily tool for several dozen people, the flat rate per server generally wins: the marginal cost of an extra request is zero, and each new use improves profitability instead of eroding it.
How to choose
- For a test: tokens or a few licences are enough.
- For broad, daily use, especially with agents or automation: the flat rate quickly becomes the most advantageous.
- For sensitive data: an on-premises server settles both the cost question and the sovereignty question.
Our choice
We chose a flat rate per server, never per user: MIA and our local API are used with no limit on requests. Successful adoption should never become bad budget news.
To estimate what AI would bring you before talking about cost, the AI audit quantifies the gains, use by use.
