AI in business means handing software systems the tasks that used to require skilled human time: writing, analysis, sorting, customer replies. In 2026, the question for an SME is no longer whether to experiment, but which two or three uses to pick, measure and sustain.
Answer a few questions and get a personalized assessment with recommendations tailored to your industry.
Assess my AI maturityAI in business means handing software systems the tasks that used to require skilled human time: writing, analysis, sorting, customer replies. In 2026, the question for an SME is no longer whether to experiment, but which two or three uses to pick, measure and sustain.
This guide covers the real state of adoption, the use cases that pay off, how to calculate return on investment, a deployment sequence, the public funding available in France and the regulatory obligations. It sits alongside our two other pillar guides: the complete guide to AI training for business for the skills side, and the AI Act guide for SMEs for the compliance side.
One company in seven below 50 employees uses AI, against nearly six in ten above 250 employees: the size gap is the defining fact of this period. According to the information and communication technology survey published by Insee for 2025, 18 % of companies based in France with 10 or more employees report using at least one artificial intelligence technology, 8 points more than a year earlier. The breakdown by size tells you more than the average: 15 % among companies with fewer than 50 employees, 31 % between 50 and 249, and 58 % above 250. France sits slightly below the European Union average of 20 %.
These numbers measure declared usage, not structured usage. In practice, three very different situations hide behind the same yes:
The first situation is by far the most common, and it is the one that creates the most risk while producing no measurable value. It also explains the permanent gap between announced adoption rates and gains that actually show up in the accounts.
On the policy side, the picture has become clearer. The national plan known as Osez l'IA, launched in July 2025 by the French Directorate General for Enterprise, targets an AI diffusion rate by 2030 of 100 % among large companies, 80 % among SMEs and mid-caps, and 50 % among micro-businesses. The stated public objective is therefore to move SMEs from trial to routine use within five years.
Profitable use cases share three traits: the task is repetitive, the required data already exists, and a human can verify the output in under two minutes. Anything outside that frame costs more in supervision time than it saves.
The table below sorts the most common SME use cases by implementation difficulty and by the condition that determines whether they pay off.
| Use case | Function | Difficulty | Profitability condition |
|---|---|---|---|
| First drafts (proposals, meeting notes, product sheets) | Sales, management | Low | An existing document template and systematic review |
| Sorting and qualifying inbound requests | Customer service, order desk | Low | A stable request taxonomy |
| Tier 1 replies assisted by the knowledge base | Customer service | Medium | Up to date, indexed documentation |
| Data extraction from documents (invoices, delivery notes, CVs) | Accounting, HR | Medium | Sufficient monthly volume and recurring formats |
| Internal search across company documents | All | Medium | A clean document repository and clear access rights |
| Analysis of customer verbatims or survey responses | Marketing, quality | Medium |
Two points deserve to be stated plainly. First, low-difficulty use cases are also the ones with modest unit gains: a few minutes per document. They only become significant through volume. Second, the high unit value use cases, such as internal document search or extraction, almost always require prior work on the data that companies underestimate. A retrieval system such as RAG does not compensate for a messy document base: it exposes it.
Time saved only becomes return on investment if it is reallocated to value-producing work, avoids a hire, or removes an external cost. As long as the saved hours dissolve into the working day, the project costs money without earning any.
This is the step most often skipped in AI ROI calculations. An honest formula fits in four lines:
Without the fourth line, the result is a theoretical gain. To illustrate the reasoning: a team handling 400 requests a month at 8 minutes each spends roughly 53 hours. Assisted sorting that brings handling down to 5 minutes frees about 20 hours a month. At a loaded hourly cost of 35 euros, that is around 700 euros of potential monthly value. If the tool and its oversight cost 250 euros, the balance is positive, provided those 20 hours genuinely go somewhere else. Treat these figures as a calculation template, not a promise: rerun them with your own volumes and costs.
Three indicators are enough to steer the first few months:
A high rework rate is not a failure: it signals that the instructions, the context supplied to the model or the scope of the task need revisiting. On that point, instruction quality matters more than model choice, which is precisely what prompt engineering addresses.
Start with the process, never with the tool. The most common mistake is to pick a platform and then look for something to do with it. The reverse order delivers faster and cheaper results.
Step 1: map the friction points, not the technologies. Half a day with three or four department heads is enough to list the tasks that recur, that annoy people and that can be measured. Keep the ones whose volume you already know.
Step 2: assess real maturity. State of your data, quality of internal documentation, team comfort level, confidentiality constraints. A serious AI audit mostly produces a list of prerequisites, not a list of tools. Your AI maturity level dictates how fast you can move, and it has little to do with company size.
Step 3: run a single pilot over six to eight weeks. One process, one team, one numeric indicator defined before you start. A pilot with no written stopping criterion never stops and proves nothing.
Step 4: train the people concerned. Not a general awareness talk, but a workshop built on the company's own cases and documents. It has also been a regulatory obligation since February 2025, as covered in our article on Article 4 of the AI Act. Budget is covered in our guide to AI training costs.
Step 5: write the rules before scaling. What may be entered into an AI tool, what may not, who approves what, how errors get reported. An internal AI policy template is more useful than a statement of principles.
This sequence takes about three months. Going faster is possible, usually at the expense of step 5, which is the one companies end up regretting.
Three families of tools coexist, and most SMEs only need the first two.
General purpose assistants (ChatGPT, Claude, Copilot, Gemini) cover writing, analysis and summarisation. Business plans typically run between 20 and 35 euros per user per month. What you buy in the business version is not model power, which is roughly identical to the consumer version, but the contractual frame: no reuse of your content for training, account management, logging. That is the subject of our article on ChatGPT and AI Act compliance.
AI features built into the business software you already run: CRM, accounting, support, HR. Often the best entry point, because the data is already there and adoption does not require changing habits. The cost is frequently a subscription add-on rather than a project.
Custom builds, from a simple automated workflow to an agent connected to internal systems. No-code and automation have cut the entry ticket here considerably, as detailed in our no-code guide. This family only makes sense after a conclusive pilot on the first two.
A budget order of magnitude for a 30-person SME getting serious: 250 to 700 euros a month in subscriptions for the people actually concerned, plus an upfront investment in training and scoping. The line item most often forgotten is neither of those: it is internal management time, roughly two to four days a month in the early stages.
Two frameworks apply at once, and neither exempts you from the other: the GDPR on personal data, and the European AI regulation on the systems themselves.
Regulation EU 2024/1689, known as the AI Act, applies in stages. The most immediate obligation for a typical SME is Article 4: ensuring a sufficient level of AI literacy among the people using AI on the company's behalf. It became applicable on 2 February 2025 and covers deployers, not just system providers. The detail and the following deadlines are set out in the full AI Act timeline and in our analysis of penalties.
In practice, four deliverables cover the essentials for an SME that does not operate a high-risk system:
On the GDPR side, the decisive questions are legal basis and where processing takes place. An assistant receiving employee or customer data constitutes processing, with the obligations that follow: records, information notices, retention periods. The comparison between the two frameworks is developed in our AI Act and GDPR article.
The dominant cause of failure is not technical: it is the absence of any decision about what to do with the time saved. The other five follow from there.
These mistakes share one trait: they are organisational, not technological. That is good news, because an SME can act on all of them without specialist technical skills.
Two national schemes fund support directly in France, and sector training funds cover the training itself.
Under the Osez l'IA plan, Bpifrance Conseil operates two schemes on behalf of the French state, co-funded by the France 2030 programme:
The Directorate General for Enterprise also publishes a catalogue of 88 AI solution providers suited to SMEs and mid-caps, compiled with Hub France IA, alongside a regional network of AI ambassadors.
For training itself, funding runs through the sector skills operator your company belongs to. Conditions, ceilings and application mechanics are detailed in our OPCO funding guide. Nonprofits follow a similar logic, described in our dedicated article, and organisations under 10 employees in the one on micro-businesses.
Do you need a data scientist to deploy AI in an SME? No, in the vast majority of cases. The use cases that pay off in an SME rely on existing tools plus scoping, training and measurement work. A technical profile becomes useful once you connect AI to internal databases or build custom automations.
How long before you see a measurable result? Allow six to eight weeks for a properly scoped pilot, provided the indicator was defined before launch. Deployments that show nothing after three months almost always suffer from too broad a scope or too little usage, not from the wrong tool.
Does my data end up with the vendor? It depends on the plan. Business versions of the main assistants contractually exclude reuse of your content for model training, which free versions do not guarantee. Check that point, the processing location and the retention period before using any real data.
Will AI cut jobs in my company? In an SME, the observed effect is on the mix of tasks rather than on headcount: repetitive work shrinks, review and relationship work grows. The question to settle upfront is the same one as for the ROI calculation: what the freed hours are for.
Is AI training really mandatory? Article 4 of Regulation EU 2024/1689 requires companies deploying AI systems to ensure a sufficient level of literacy among the people involved. It prescribes no format and no number of hours, but it assumes you can produce evidence. The subject is developed in our article on training certificates.
Which use case should I start with if I have no idea? Take the most repetitive written task in your busiest department. It is almost always standardised document writing or inbound request handling. Both combine high volume, low risk and simple measurement.
What if my teams already use AI without approval? This is the most common situation and it is not a disaster, provided you address it quickly. Inventory existing usage without looking to punish anyone, set the input rules, open business accounts. Informal usage often reveals the company's best use cases.
GrowthPerf is a Qualiopi-certified French training provider specialising in AI, no-code and automation for SMEs and nonprofits. Our approach follows the sequence described above, in that order: scoping the use cases, training the teams concerned on your own documents, then putting usage rules and tracking indicators in place.
In practice that takes the form of a one-day acculturation programme to set common ground, described on the AI acculturation for business page, or a two-day operational programme built around your processes, detailed on the operational AI for SMEs page. Both formats are eligible for French OPCO funding.
If you are not sure where to start, the simplest route is a 30-minute conversation to identify your two or three priority use cases and check what is fundable in your situation. Book your free 30-minute AI audit.
To go further, two useful reads depending on your current priority: the complete guide to AI training for business if your issue is team skills, and the AI Act guide for SMEs if your issue is compliance.
| A minimum volume and a defined reading grid |
| Meeting preparation and note summarisation | Sales | Low | Clean capture of the conversation and participant consent |
| Assisted development and internal automation | Technical, ops | High | Someone able to review what is produced |