FOUNDER · EVERY WEDNESDAY · SEPTEMBER 9, 2026
Your first AI project is smaller than you think.
A few weeks ago the chief executive of an industrial mid-market company told me he was waiting. Not for the technology, he said. For the money. In his head, a serious AI project meant several hundred thousand euros and a program running two or three years. So it sat in next year’s plan, exactly where it had sat the year before.
He is right to be careful. He is wrong about the order of magnitude, and that mistake has cost him a year.
This spring, Siparex and Bpifrance documented twenty completed AI projects in French SMEs and mid-market companies, in a white paper called Les audacieux de l’IA. The number a chief executive should take away is this one: 70 percent of those projects required an investment between 15,000 and 100,000 euros, and 80 percent were deployed in under six months. This is not a software vendor’s brochure. It is a private equity firm and the public investment bank examining companies they finance, after three years of work alongside more than 1,500 companies.
One caveat I would rather state than hide: these twenty are cases they chose to publish, drawn from companies they support. They are successes, not a representative sample. Read them as what is achievable, not as an average.
The rest of the picture explains the hesitation. In two years, the share of French SMEs and mid-market companies that have launched an AI project went from 15 to 55 percent. Among the 110 executives Siparex surveyed at the end of 2025, 86 percent call it a strategic priority, against 48 percent twelve months earlier. The talk moved faster than the budgets.
This is not a technology problem. It is an execution problem.
Aldes makes indoor air quality and ventilation systems: 380 million euros in revenue, 1,900 people. Its problem was entirely mundane. A thousand emails a day asking for a quote, fifty technical sales people qualifying them by hand, and a promise to answer within forty-eight hours that was getting harder to keep.
What they built is just as mundane. A system that reads each incoming email, classifies it, drafts the clarifying questions when the request is incomplete, and prepares the quote when it is not. Build cost, around 40,000 euros. Build time, five months. Running cost, 60,000 euros a year, which I quote because it is the line most write-ups leave out. Estimated gain in year one, 800,000 euros, between the productivity of the sales team and the requests that stopped going cold.
Two details matter more than those numbers. They started with a proof of concept on a deliberately narrow perimeter, quotes under 3,000 euros. And they put seven key users inside the test, who became the people who sold it internally. That is the method, and it costs nothing. Figures reported by the company through its investor, and the 800,000 is an estimate, so read it as an order of magnitude.
The study’s own conclusion says the same thing in one line: the key success factor is not the technology, it is the method.
The 3 essentials for you
- The order of magnitude is 15,000 to 100,000 euros and under six months, not a multi-year program. If your first project is quoted at ten times that, the scope is wrong before the technology is.
- The projects that pay are rarely the spectacular ones. Half the panel sits in the intermediate range and a third is openly simple. Simple is where most companies start, and it is what funds the rest.
- What separates these twenty from the others is method, not money: two or three prioritized cases, a named owner, teams trained. That is a decision you can take this month, at your own scale.
What I would do in your shoes…
Stop sizing the AI budget. Size the first case instead. Take one task your teams do every week, where you already know what it costs you in hours. Ask for a quote on that single case, with data preparation and team support visible on page one, and hold it against the 15,000 to 100,000 range. If it fits, you have your answer and your scope. If it does not fit, you have learned something more useful: the scope was too wide, and nobody had said so.
Then give it a date. Not a start date, a decision date, with the starting figure you will compare the result against. The projects that never end are the ones nobody ever agreed to judge.
What we are building at YPrtnrs points the same way. An execution engine, Y, your AI chief of staff, that prepares, drafts and organizes, then leaves the decision to you. I propose, you decide.
The case for waiting is not absurd. A company whose sales history is a mess, or whose year is already committed, can reasonably decide that its first case is a January decision rather than a September one. That is defensible, and I have seen serious people make it.
What has expired is the third option: leaving the subject in next year’s plan because the budget was imagined too large. That one has already been decided twice, by default, and it will be decided a third time the same way unless somebody puts a figure on a single case.
My inbox is open for the figures that contradict this range, or for a first project that came in far above it: arnaud@yprtnrs.com. I read everything, and I answer personally.
