AI Agents
40% of small businesses use AI, only 19% pay for it: where to start when you run a team of 10 to 30
AI is spreading fast in French small businesses and SMEs, but mostly through individual use. What the 2026 figures say, and a seven-step method to start from a problem rather than a tool.
In short
- In 2026, 40% of French very small businesses and SMEs use at least one AI solution, and 53% of SMEs, according to the France Num barometer published on September 28, 2026.
- Only 19% pay for an AI tool, and AI comes last among the digital projects of the companies that have such projects (18%).
- Adoption is fast but mostly individual: an employee using an assistant is not a company with a strategy.
- The right method starts from five concrete problems, not from a tool.
- There are three levels: well-framed subscriptions, automation, an assistant connected to your data. Most SMEs should start with the first.
You probably already have AI in your company. A salesperson who rewords their emails, an assistant who summarizes meeting notes, a partner who asks ChatGPT questions in the evening. The question is no longer "should we go for it", but "how do we move from scattered uses to something that actually saves the company time".
What do the 2026 figures say about AI in SMEs?
In 2026, 40% of French very small businesses and SMEs use at least one artificial intelligence solution, 14 points more than in 2025, and 53% of SMEs, according to the France Num barometer published by the Direction générale des Entreprises on September 28, 2026 (France Num is the French government's portal for small-business digital transformation, run by the Ministry of the Economy's directorate for enterprise). The survey covers 9,655 companies interviewed between March 23 and April 18, 2026 (Entreprendre Service Public, the French government's official information site for businesses, October 2, 2026).
The growth is spectacular: 13% in 2024, 26% in 2025, 40% in 2026, according to the breakdown reported by Blog du Modérateur on September 30, 2026. Watch the vocabulary: this 40% figure combines very small businesses and SMEs. For SMEs alone, it is 53%.
But three other figures from the same barometer tell a more nuanced story:
- 19% of companies pay for an AI solution (32% of SMEs). This percentage covers all the companies surveyed, not only AI users: above all, it shows that a large share of usage goes through free versions.
- The main uses are content generation (34%) and chatbots (24%), according to Entreprendre Service Public. In other words, mostly help with writing.
- AI comes last among digital projects for 2026-2027 (18% of companies that have projects), far behind e-invoicing (39%), software (26%) and training (21%).
The figures vary from one survey to another depending on the question asked. Bpifrance Le Lab found at the end of 2025 that 55% of very small businesses and SMEs used generative AI (Bpifrance is the French public investment bank). The two surveys do not contradict each other: they agree on broad adoption, and on usage that is still poorly organized.
Why is "using ChatGPT" not an AI strategy?
Using ChatGPT is individual use: one person saves a few minutes on a task. An AI strategy is a company decision: an identified process, usage rules, a measured result. The first is useful, but it does not change how the business runs.
The difference shows up in three everyday situations:
- Know-how stays in one person's head. The salesperson has found the right wording for their follow-ups. Nobody else knows it. If they leave, it leaves with them.
- Data circulates without rules. A customer quote, a margin spreadsheet or a contract ends up pasted into a free tool, under a personal account. It is not bad intent, it is the lack of a rule. We wrote an article on it: the AI usage policy that protects your customer data.
- The gain is not measured. People feel that "it helps", without knowing how many hours a week, or on which task.
Take a fictional example. An assistant prepares a sales reply faster with an AI assistant. But she is still waiting for the price from the owner and a product reference held by a colleague. The draft arrives earlier, the final reply is still stuck. The real work is centralizing pricing, not adding a tool.
The barometer confirms this in its own way: when business owners list their digital priorities, they do not ask for "more AI". They talk about invoicing, software, training. Their problem is organization. AI is only one of the ways to address it.
Which problems should you list before choosing a tool?
Before choosing a tool, list five concrete problems that cost time every week, with the person involved and an estimate of the hours. This takes 20 minutes in a team meeting and is worth more than any sales demo.
Ask your team three questions:
- Which task do you do every week thinking "this again"?
- Which information do you copy from one tool to another?
- What is waiting on your desk because nobody has time?
In companies of 10 to 30 people, the same answers come back almost every time:
- Incoming emails to sort, forward, and answer with the same information.
- Follow-ups on unsigned quotes and unpaid invoices, done when someone remembers.
- Re-entering data between the quoting software, the tracking spreadsheet and the accounting system.
- Quotes built from an old quote, with the errors that come with it.
- The Monday report, put together by hand from three sources.
For each problem, write down: who does it, how many times a week, how long each time. A 20-minute task done 10 times a week adds up to more than 3 hours, or about 150 hours a year for a single person.
How do you know where AI helps, and where it is useless?
AI and automation help with frequent tasks that have clear rules and moderate stakes. They are of little use for rare tasks, or for those that require case-by-case judgment. A three-criterion grid is enough to sort your list.
| Criterion | Good candidate | Poor candidate |
|---|---|---|
| Frequency | Every day or every week | A few times a year |
| Rules | Can be explained to an intern on one page | "It depends" every time |
| Stakes of an error | Can be caught by a review | Commits the company (price, contract, unhappy customer) |
An invoice reminder ticks all three boxes. A price negotiation ticks none. Sorting incoming requests is a good candidate, provided a human approves the ambiguous cases.
Often, the right answer is not even AI. Choose the simplest mechanism that gives a verifiable result:
| Observed situation | First option to try |
|---|---|
| A piece of information has no owner | Assign who keeps it up to date |
| The same data moves from one software to another | A connection or a controlled import |
| A document follows a stable template | A template with required fields |
| Varied messages need to be classified | AI assistance, evaluated on real examples |
| A decision commits the company | Assisted preparation, human approval |
We have detailed the seven tasks an SME should automate first, ranked from the most profitable to the most convenient. They overlap with the list above in almost every case.
What are the three levels of AI use in a business?
There are three levels, from the simplest to the most demanding: well-framed assistant subscriptions, automating a process, then an assistant or agent connected to your own tools and data. You only move to the next level once the previous one works.
Level 1: well-framed subscriptions. Business accounts for the team, written usage rules, half a day of training. This is where most companies already are, but without the framework. The cost of a consumer tool is between €0 and €25 per month according to France Num. For many SMEs, this is the right starting point, and sometimes the right end point.
Level 2: automating a process. A flow that runs on its own: the reminder goes out at the right time, the incoming request is classified, the quote is prepared from your price list. Tools like n8n or Make connect your software together; we compare them in our n8n or Make guide for SMEs.
Level 3: an assistant connected to your data. An assistant that knows your customers, your prices, your procedures, and can act in your tools. It is the most powerful level, and the one that requires the most framing. We cover it in our article on AI agents by 2027.
For the budgets at each level, see the market price ranges for an AI project in an SME.
What can you test in two weeks without a provider?
Without a provider, an SME can test a level 1 use on a specific problem in two weeks: choose an assistant, write three rules, have two volunteers work on the same task and measure the time spent before and after.
Here is a simple protocol:
- Choose a single problem from your list, preferably a writing task (answers to frequent requests, meeting notes, product sheets).
- Appoint two volunteers, each with a business account.
- Write three rules: which data never to paste, who reviews before sending, where to store the good prompts.
- Measure the time spent on the task the first week without AI, the second week with it.
- Have the results checked by someone other than the person who wrote the instructions, and send nothing automatically during the trial.
- Take stock in 30 minutes: real gain, pain points, what it would take to go further.
This test teaches you two things: whether the gain exists in your company, and what is blocking. Often, what blocks is not the AI but the data: scattered information, outdated files, tools that do not talk to each other.
That is where an outside view saves time: when several tools need to be connected, when customer data is involved, or when the problem affects several people at once.
How do you prepare a first conversation with an agency?
For a useful first conversation, bring your list of problems, the list of your tools and an idea of the expected gain. With those three elements, a serious provider should be able to tell you whether AI is the right answer, and where to start.
Prepare:
- Your five problems, with the estimated hours.
- Your tools: quoting and invoicing software, CRM or tracking spreadsheet, email, file storage.
- Your constraints: sensitive data, customers who are demanding about confidentiality, deadlines (the September 2027 e-invoicing mandate in France, for example).
- What you have already tried, and what did not work.
And ask the provider a simple question: "In which case would you tell me to do nothing?" The answer says a lot. At Polaris, it is a principle we stand by on our AI agency for SMEs page: we tell you when AI is not the solution.
One last point: your online visibility is also part of the picture. Your prospects now ask AI assistants before contacting you. We explain why your Google clicks are falling and how to get cited by AI.
Bring your five problems, and we will prioritize them together. In 30 minutes, we look at your list and your tools, and we tell you honestly where to start, including when the answer is not AI. Book your slot.
Frequently asked questions
- My company is small: is AI worth it below 30 employees?
- Yes, as soon as there are repetitive tasks. The criterion is not size but repetition: a 20-minute task done ten times a week adds up to about 150 hours a year. Start with the simplest level, well-framed subscriptions, before automating.
- Are free tools enough to get started?
- To test individual use, often yes. The risk is not the quality of the tool, it is the lack of a framework: customer data in personal accounts, no record of what was shared. Switch to business accounts as soon as company data is involved.
- Should you train the whole team before starting?
- No. Two volunteers testing on a real problem for two weeks are better than general training for everyone. Training then expands, based on what worked in your company.
- Why do AI adoption figures vary so much from one study to another?
- Because the questions differ: "at least one AI solution" (France Num, the French government's small-business digital portal, 40% in 2026), "use of generative AI" (Bpifrance Le Lab, 55% at the end of 2025), "regular" or "occasional" use. All of them show the same trend: fast adoption that is still poorly structured.
- Should you start with an autonomous agent?
- Rarely. An assistant that prepares an action and then waits for human approval is easier to evaluate and much less risky. Autonomy becomes justified later, on a specific scope, once the results are reliable.
- How long before you see results?
- A level 1 test gives you an answer in two weeks. A process automation is usually built in a few days to a few weeks. What takes time is rarely the technology: it is clarifying the process and making the data reliable.