From tool to strategy: why individual use of AI has yet to move the profit and loss account

Three recent studies of French companies produce three figures that seem to disagree. 32% of small and mid-sized companies use AI, according to their own leaders¹. 18% of companies with ten or more employees, according to INSEE, the national statistics office. 67% of companies with twenty or more employees use generative AI, according to the Banque de France.

All three are right. They do not measure the same thing.

Three figures, three definitions

The highest figure counts use: someone in the company relies on an assistant to write, summarise or translate. The lowest counts a decision: the company has put AI inside a process, with a budget and an owner. Between the two lies the gap most leaders feel without being able to name it.

Individual use is already widespread. Organised use is still a minority. And only the second shows up in the accounts.

Spain shows the same gap, more sharply. The IndesIA 2026 barometer finds that 89% of organisations use AI somewhere, and 6% see an effect on operating profit². Use is everywhere; results are rare.

What the tool does not do

An assistant saves one person a few minutes. A margin moves when a process changes: who does what, from which data, with which decision at the end.

Ten minutes saved on an email cut no cost. They become a saving on the day the organisation acts on them: a step removed, a delay shortened, data no longer keyed in twice, a role redirected to work nobody had time for.

An AI strategy that changes no process is a usage policy. Useful, but it is not a strategy.

Where the gap is decided

In companies of 20 to 500 people, the gap almost always comes down to three things missing.

A named process, with an owner

"Using AI in sales" is not a project. "Taking a quotation from three days to three hours, owned by the head of sales" is one. The first is measured in enthusiasm, the second in days.

Data the system can read

A system is only as good as what it reads. If quotations live in personal spreadsheets and pricing terms in the memory of two people, the first piece of work is not AI. It is making that data readable, once and properly.

A decision on what changes

What does the system do on its own? What does a person keep? Who approves, and when? Without those answers in writing, the tool remains an experiment that everyone uses their own way, and that nobody can assess.

Where to begin

Not with a three-year plan. With one flow that runs through the profit and loss account: quote to cash, purchasing, customer service, regulatory documents.

1. Measure it as it is, before touching it.

2. Decide what a system would do, and what a person keeps.

3. Build a first version that runs on the company's real data.

4. Decide, with figures in hand: extend, correct, or stop.

The last option matters as much as the other two. A documented decision not to build is worth more than a project that settles in because nobody judged it.

That is what our entry mandate is for: ten working days, a written decision, and a prototype running on a real process. We start with that.

Elah.Published on October 20, 2026