AI in a French mid-cap: where to start when you have an IT department, several entities and an ERP

This guide is for chief executives, finance directors and IT directors of French mid-caps who have already seen AI come through their company (licences, a pilot, a few individual uses) and who wonder how to turn it into a measurable result. It does not compare tools. It answers four questions: what makes the subject different in a mid-cap, which processes to look at first, who should decide what, and in what order to move.

What a French mid-cap is, and why it changes the question

In French statistics, an entreprise de taille intermédiaire (ETI) employs between 250 and 4,999 people, with turnover of no more than €1.5 billion or a balance sheet total of no more than €2 billion. A business with fewer than 250 employees that exceeds €50 million in turnover and €43 million in balance sheet total also belongs to the category.

Insee, the French national statistics office, counted 7,442 of them in 2023. Three figures from the same publication describe what they really are.

The middle figure is the one that matters here. A mid-cap is not a large SME: it is a group of around ten companies, often built through acquisitions, with several sites, sometimes several ERPs, and processes that share a name without running the same way from one entity to the next. An AI project there is first a project across several entities and several systems.

Where businesses of this size stand

Insee does not publish an AI usage rate by business category. Its annual survey gives a “250 employees or more” band, which combines mid-caps and large companies. It is the closest benchmark, to be read with that caveat.

In that band, France is 3 points above the European Union average (55%). The question is no longer whether to go, but what it produces.

Two findings from the same survey show where it gets stuck.

Lack of expertise comes before lack of usefulness. Among businesses with 250 or more employees that use no AI, 73% cite a lack of expertise: it is their first reason. About half cite incompatibility with existing equipment, software or systems. Among the smallest businesses, the first answer is that they see no use for it. A mid-cap sees the use. It runs into know-how and into what is already there.

Buying off the shelf dominates. 80% of businesses using AI acquire it as ready-to-use commercial software. Insee nonetheless singles out one group: 10% of businesses using AI, which employ 48% of the staff of all of them, use at least five technologies, and nearly two thirds of them developed these in-house. They are more often businesses with 250 or more employees.

What sets an AI project in a mid-cap apart

The last point is the most common. A mid-cap can afford to launch several experiments at once, and none of them clears the next step: connection to the information system, permissions, monitoring. A convincing pilot that waits for six months does not have a model problem. It has an integration problem, covered in the AI POC that will never reach production as-is.

Three levels of AI in a mid-cap

The first two levels are bought. They speed up the person at their desk, which is useful and rarely visible in the accounts, for a reason explained in We rolled out Copilot, and nothing changed in the accounts: the cost of a process lies in the time that elapses between its steps, not in the time spent on each. For what Copilot reaches and what it does not trigger, see Copilot or custom AI.

The third level is where a mid-cap sets itself apart from its competitors, because nobody else has its processes. It is also where the existing estate weighs most. If part of your business applications is built in-house, AI is wired into those applications rather than beside them: that is the subject of the guide Integrating AI into an existing .NET application, written for your development teams.

The candidate processes

A good candidate has four characteristics: it is repetitive, it can be measured, it crosses several systems or several entities, and an error in it can be caught before it leaves the company.

A note on invoices. Since 1 September 2026, large companies and mid-caps in France must issue their invoices electronically, and every business must be able to receive them. Invoices between VAT-registered businesses therefore increasingly arrive as structured data. Automatic invoice reading, long the first use case quoted, is losing its appeal. The documents on the second line remain.

That second line is the most mature. I put a document AI pipeline into production on this pattern, for an industrial company: the AI extracts, deterministic rules check, a human validates what is doubtful, and the corrections improve later extractions. The principle fits in one sentence: AI proposes, deterministic logic disposes.

I give no percentage gain. It depends on your volumes, your systems and your organisation. It is measured on one of your processes, before deciding.

Who decides what

Insee notes that a lack of expertise also holds back 53% of the businesses that already use AI. In a mid-cap, that lack often takes a precise form: everyone waits for someone else to own the subject. The split that works is the following.

A project carried by IT alone becomes a tool project. A project carried by the business line alone stops at the connection. It takes a business owner who answers for the figure, and an IT department that answers for production.

What the EU AI Act changes for a mid-cap

The AI Act became applicable on 2 August 2026, with a staggered timeline that the omnibus regulation in force since 27 July 2026 amended.

Two practical consequences. Training staff is already an obligation, not an extra. And most of the processes in the previous table are not high-risk. CV screening, employee evaluation and credit scoring are. A mid-cap that starts there picks the most regulated ground for its first project.

This guide is not legal advice. Classifying a system is done with your lawyer, on the actual case.

Eight questions to settle before the first project

1. Which uses already exist? Licences are being paid for, staff use an assistant, sometimes with a personal account. The inventory comes before the strategy.
2. Which process, and who owns it? One process, one named business owner.
3. What does it cost today? Volumes, lead times, time spent, error rate, measured over a month. Without that figure, no result can be proved.
4. Is the process the same in every entity? If it differs, start with one entity, or harmonise first. Automating ten variants means running ten projects.
5. Where is the data, and in what state? If it sits in binders or in files nobody keeps up to date, the first job is not an AI job: see Your problem is not (yet) an AI problem.
6. Does the tool you already pay for do it? The ERP, the CRM or Copilot sometimes cover the need. In that case, the answer is configuration.
7. Who will run the system a year from now? Monitoring, cost per item processed, error recovery, model changes. If the answer is “the supplier”, reversibility is negotiated before signing.
8. What happens when the AI is not sure of itself? A threshold, a validation queue, a named person. This is the point that separates a pilot from a system in production.

Going further

Questions 2 to 5 are the ones that call for an outside view, because they require comparing entities that do not compare themselves spontaneously. That is the purpose of the 30-day AI diagnostic: processes mapped, the return costed in euros per priority case, and a 90-day roadmap, including when the right answer is a feature of the tools you already have. If your company relies on a .NET and SQL Server application estate, the next step is .NET AI integration: AI built into your applications, with your teams, rather than beside them.