40% of SMEs use AI, mostly to write text: how to move to the uses that pay off
The France Num 2026 Barometer, published on 28 September by the French Directorate General for Enterprises from a survey of 9,655 micro, small and medium-sized businesses, delivers one clear figure: 40% of them use at least one artificial intelligence tool, up from 13% two years ago.
Behind that growth, the picture is more nuanced. Most of the usage is still text generation. The uses that actually change a process, automation or data analysis, stall at around 10%. We suggest reading these figures from the point of view of a Swiss SME.
What the barometer really says
Adoption has tripled in two years: 13% in 2024, 26% in 2025, 40% in 2026. Among SMEs alone, excluding micro-businesses, the rate reaches 53%, up from 34% last year.
But the dominant use remains content production. The 2026 figures, with 2025 in brackets, speak for themselves:
text, voice or image generation: 34% (22%);
chatbots, assistants and information search: 24% (14%);
document analysis and classification: 13% (6%);
task automation: 11% (5%);
data analysis and forecasting: 9% (5%).
Another lesson: 68% of the businesses using AI report a positive effect on their activity. That share climbs to 84% among those paying for a solution. Yet only 19% use a paid AI tool, and just 5% report any spending on training.
The barometer covers French companies, according to the summary published by TPE Actu on 28 September 2026. We have no equivalent survey of this size for French-speaking Switzerland, but what we observe among our clients in Geneva and Vaud follows the same curve: plenty of assistants, few processes.
Why convenience uses are not enough
Drafting an email or summarising a PDF with an assistant saves time. One freelancer interviewed in the barometer's qualitative survey estimates saving one to two hours a day. That gain is real, but it stays individual and hard to measure.
It changes neither your delivery lead time, nor your invoicing error rate, nor your ability to answer a tender. Yet those are the indicators your bank, your customers or a potential buyer look at.
The step that matters lies between the individual assistant and the tooled process. It requires three things the barometer reveals by their absence: accessible data, a committed budget and trained staff.
On that last point, 5% of businesses investing in training is the most worrying figure in the study. A tool that is paid for but not mastered ends up in a drawer, and the experience leaves the team sceptical about the next project.
We also see a less visible risk: convenience AI settles in without a framework. Everyone uses their own assistant, with their own data, sometimes clients' data. The productivity gain is then paid for in exposure, without anyone having decided on that trade-off.
An example: an 18-person fiduciary firm in Carouge
Take a Swiss fiduciary and accounting firm with 18 staff. Since 2025 it has used a generative assistant to rephrase client letters and prepare summaries. Everyone saves a little time, nobody knows how much.
The real reserve lies elsewhere. Every month, the team handles close to 1,200 accounting documents received by email, often as unstructured PDFs. Sorting, reading and keying them in takes the equivalent of one and a half full-time positions.
A targeted project changes the scale:
automatic classification of incoming documents (invoice, bank statement, expense claim) by a document-reading model;
extraction of the key fields and a proposed entry in the accounting software, validated by a person;
a weekly exception dashboard, so the team corrects the model instead of enduring it.
On this kind of flow, we aim for an automatic processing rate of 70 to 80% after three months, with the remainder going to manual review. The time freed up is measurable: one position redeployed to advisory work, which is billed at a higher rate than data entry.
The difference with the writing assistant is not the technology. It is that a process was chosen, its cost measured beforehand, and an indicator set afterwards.
Four steps from assistant to process
We recommend a short approach, compatible with an SME's budget and calendar. It fits in four steps.
Map the high-volume repetitive tasks: anything counted in hundreds per month (documents, requests, quotes, tickets) is a candidate.
Pick a single flow, with a known current cost in hours. Without a starting measurement, there will be no proof at the finish.
Tool it with human validation: the AI proposes, a person validates, and every correction feeds the model's improvement.
Set an indicator and a date: automatic processing rate, lead time, error rate, reviewed at 30 and then 90 days.
The training budget is part of the project, not a second phase. Two half-days with the team involved are often worth more than the licence itself.
This framing matches the barometer's finding: the businesses that invest, in licences and in skills, are those reporting the most positive effects. Correlation is not proof, but it points in a direction.
What we take away
The 40% threshold is good news: AI is no longer a topic reserved for pioneers. But the adoption measured by the barometer is mostly an adoption of assistants, not yet a transformation of processes.
For a Swiss SME, the question is no longer "do we use AI?" but "on which process do we have a measured result?". That answer is what separates a convenience use from a competitive advantage.
We support SMEs through this step, from choosing the flow to measuring the result. To see where your company stands, our AI Barometer gives you a first diagnosis in a few minutes. You can also contact us to discuss it.
Comments