AI in small businesses: five uses that save time in the first month

Why French small businesses are taking it up now

Because the cost of entry has collapsed, and because use came before organisation. According to the business survey run by Bpifrance Le Lab, 55% of micro, small and mid-sized businesses said they were using generative AI at the end of 2025, against 31% a year earlier. The Baromètre France Num 2025, published by the Directorate-General for Enterprise, notes for its part that the share of small firms drawing on AI solutions has doubled in a year, reaching 26%. The gap between the two figures says everything that matters: a great many employees are using AI on their own initiative, in a browser, without the company having decided or framed anything. That is both an opportunity — adoption is already there, so training will be quicker — and a risk, since nobody is controlling what leaves the company in those exchanges. The question is therefore no longer whether to get started, but how to frame a use that already exists inside the business.

Sources: Bpifrance Le Lab and Baromètre France Num 2025, Directorate-General for Enterprise (both in French).

The five uses that live up to the promise

One — document processing: supplier invoices, delivery notes and contracts read automatically, with amounts, dates and references extracted and pushed into the accounts or the ERP. Two — customer replies: draft answers to recurring requests, prepared from your own templates and approved by a person before they go out. Three — data entry: meeting notes dictated then structured, customer records completed, data copied from one tool to another without anyone touching it. Four — reporting: a sales or finance dashboard regenerated every Monday morning from your own files, with a written summary of the variances. Five — internal search: asking a question in plain language rather than digging through shared folders. Not one of the five requires you to change software. They graft onto the mailbox, the sales system and the shared folders your teams already open every morning.

What these five uses have in common: the task is frequent, repetitive, text-based and low-risk. When one of those four criteria is missing, the gain melts away.

What does not work — and gets asked for anyway

Three requests come back regularly, and we turn them down. “An assistant that answers customers on its own”: without human approval, a mistake goes out to a customer under your company letterhead, and you carry the responsibility for it. “An AI that decides” — granting payment terms, shortlisting a candidate, adjusting a price: these are decisions that commit the business, sometimes in law, and they must stay with people. “An AI across all our data, straight away”: with no sorting beforehand, the system serves up the obsolete internal memo just as readily as the contract in force, and the teams lose confidence within a week. What these failures have in common is always the same thing: the project started from a tool instead of starting from a measurable task. We would rather lose a sale than deliver a system we know will be abandoned within three months.

How long before a visible result

Two to four weeks for a first use case, provided you tackle only one. Our sequence: a half-day scoping session with you and two or three people from the shop floor, to list the time-consuming tasks and put hours per week against them; a proof of concept of two to four weeks, built on real data and used by real users, with a measurement before and after; a rollout integrated into the tools already in place, to avoid any pointless change of habit; short training, one session per department and a one-page document. If the gain is not there at the end of the proof of concept, the project stops there — and you will have committed a few weeks, not an annual budget. We measure the gain a second time a month after go-live, because a gain measured in the launch week means nothing.

Key points

• One use case at a time, quantified in hours per week before anything starts.
• A person always approves before a document goes out to a customer.
• If nobody can put a figure on the time spent today, the project is not ready.

And where does the GDPR fit in?

It is compatible, on three conditions dealt with at the scoping stage. Know where the data goes: models hosted in the European Union, or run on your own infrastructure when the data is sensitive, with a contract that explicitly rules out any reuse of your data for training. Limit what is sent: personal data pseudonymised before processing, the document scope restricted to what is strictly necessary, retention periods defined. Document it: the processing recorded in the register, the people concerned informed, and an impact assessment where the processing warrants one. The real risk today is not the framed project: it is the employee who pastes a customer file into a consumer service, for want of a clear instruction and an approved tool, without realising that this innocuous gesture sends data out of the company. A one-page written instruction and an approved tool settle that point better than any blanket ban.