AI in SMEs: 40% of French microbusinesses and SMEs already use it

The survey collected responses from 9,655 businesses: 6,786 microbusinesses and 2,869 SMEs. Fieldwork ran from 23 March to 18 April 2026. These details are confirmed by DILA’s institutional summary, published on 2 October 2026. This is reported usage at the time of the survey, not an independent audit of each deployment.

The Directorate General for Enterprise’s press release confirms an increase of 14 percentage points in one year across microbusinesses and SMEs. It draws on the same study, rather than providing a second independent measurement.

Three distinctions prevent misleading conclusions:

• Usage is not profitability. A business may use an assistant without measuring its net benefit.
• AI is not only generative AI. The scope includes different families of solutions and tasks.
• Usage is not a paid subscription. France Num reports that 19% of all microbusinesses and SMEs use paid solutions.

The results therefore do not establish that four in ten businesses have automated their prospecting. They show tool adoption, not a shared level of sales performance. Nor should you reconstruct a microbusiness-only rate from the raw sample counts: representativeness and weighting cannot be reduced to the sample’s composition.

France Num describes several uses: content generation at 34%, conversational assistants and chatbots at 24%, document analysis and classification at 13%, and task automation at 11%. These categories may overlap; their percentages must not be added together. They do not refer exclusively to sales activities.

For your own team, start with an observable friction. Does an agency receive incomplete enquiries? Does a salesperson have to rebuild the context before every follow-up? Are meeting notes difficult to find? These are work problems before they are tool problems.

Choose a task narrow enough to compare two ways of doing it. Examples include turning an authorised note into a structured summary, suggesting questions before a quotation, or preparing a draft from verified facts. These are proposed trials, not results observed at a customer.

The selection rule is straightforward: favour a repeated, verifiable and reversible task. For an initial trial, avoid decisions that directly commit a price, a contractual promise or a message sent to a prospect. A sensible first use lets the responsible person correct the output easily before it has any external effect.

Fluent writing can make a record look complete while crucial information is missing. Before asking AI for help, split the record into three parts: sourced facts, assumptions and open questions. An assistant can organise these elements; it must not turn an assumption into a certainty.

When preparing for a meeting, provide only necessary and authorised information. Ask for a short output: confirmed activity, fit with your offering, missing information and discovery questions. Include this instruction: ‘If the information is not in the supplied material, mark it unknown. Do not infer budget, buying intent or urgency.’

The CRM completeness matrix for usable qualification helps distinguish a filled-in record from a genuinely usable one. Keep the date and source of each important fact as well. A public website alone does not establish that a need is current.

The sales decision should then remain explicit: contact, verify a piece of information or rule out. An unexplained score is less useful than a readable reason. If your team cannot explain why a business is a priority, do not automate its selection yet. AI should make the reasoning easier to inspect, not hide the criteria behind a confident summary.

Once the target is qualified, AI can suggest an initial formulation. Give it a precise angle, a verified fact, your actual offer and a single next step. Do not ask it to manufacture personalisation from vague clues.

Before using the draft, check four points:

• Is the cited fact accurate, recent and relevant to the recipient?
• Does the message describe a possibility rather than an assumed problem?
• Does the proposal match what you can actually deliver?
• Is the next step straightforward and free from artificial pressure?

Avoid sentences claiming that you analysed a company when nobody performed that analysis. Correct phrasing does not compensate for an invented observation. Likewise, generating more variations does not fix an offer aimed at the wrong target.

For agencies organising their prospecting, SprintLead’s agency page is a starting point for examining the product; the protocol described here is a working method, not a promise of an AI feature in SprintLead. Keep preparation, approval and sending separate. This allows the team to save time on a draft without automatically delegating the decision to contact someone. It also makes it clear where an error can still be stopped before reaching a recipient.

The pilot should begin with a list of authorised data, not a complete CRM export. Limit what you share to what the task requires. For initial comparisons, use fictional or properly anonymised records where possible. Removing a name is not always enough: unusual details can still identify a person.

Examine the provider’s conditions for retention, possible data use, access, deletion and configuration options. Check who is allowed to connect the tool to your sales system too. A subscription alone is neither legal validation nor authorisation to share every record.

Write a short operational rule: which data, for which task, in which tool, with which reviewer. Also define what happens if an error occurs: stop the trial, withdraw the incorrect output and check the other affected records.

Finally, keep a record of corrections. An incorrect price has a different level of seriousness from a sentence that is too long. Classifying errors helps you decide whether a use is acceptable, needs more supervision or should be abandoned. Human review is not merely a signature at the end of the process: it must be able to prevent bad information from circulating. Without that authority, approval becomes a formality rather than a control.

Consider a two-week pilot as a practical framework, not a universal duration. Appoint one owner and select one task. Before starting, record the time spent on the existing process, the expected quality and the errors that would make an output unusable.

During the trial, compare records of similar difficulty. Count total time: preparation, generation, checking and correction. Do not measure only how fast the first draft appears. Track the share of usable outputs and the information that had to be revised as well.

At the end, there are three possible decisions: retain the use, change the protocol or stop it. Faster drafting without a net gain after review does not necessarily justify deployment. Conversely, more consistent summaries may be useful without an immediate increase in sales.

Do not attribute a meeting or a sale to AI based on a handful of records: targeting, the offer and timing may also have changed. For the next stage, define CRM exit criteria and verifiable sales stages. You can then tell whether the work is actually progressing rather than simply counting how much content the tool produces.