Deploying AI in your business without messing it up takes less technology than people think and more method: a single well-chosen use case, data whose origin you know, a short pilot, an honest before-and-after measurement, a team trained on its own tasks and a few written rules. This is the position we defend with SMEs in French-speaking Switzerland. It is better to start small, on ground you can measure, than to buy a licence for everyone and hope the uses will follow.
Many business leaders still wonder whether to “get on board or wait”. In most companies, though, the question has already been settled, often without management knowing it: employees are using AI. What remains to be decided is whether you manage these uses and whether you can measure what they bring you.
What the 2025 and 2026 surveys show
Let us start with the figures, because they paint a picture quite different from the prevailing talk. According to the ICT Omnibus survey of the Federal Statistical Office, conducted from April to July 2025 among 3’000 households and individuals, 43% of the population aged 15 to 88 had used a generative AI application by spring 2025. In addition, 31% of the population use it for work purposes. In other words, a good share of your teams has already picked up the habit of these tools, with or without your agreement.
On the business side, the labour market study carried out by the Sotomo institute for AXA, published in October 2025, surveyed 300 SMEs in German-speaking and French-speaking Switzerland in March 2025. The share of SMEs that integrate AI into their work processes rose there from 22% to 34% in one year. Uses remain modest, with translation (52%) and correspondence (47%) in the lead. Above all, only a third of the SMEs that use AI have set themselves clear rules on data protection, and this share falls to 23% in companies of 5 to 9 employees.
EY Switzerland's AI Survey 2026, published in May 2026 on the basis of 604 people working in companies in Switzerland, completes the picture. 89% of respondents already use AI in their daily professional life. On the other hand, 31% place their company in a pilot or proof-of-concept phase, and 14% in a company that has launched no concrete initiative. The obstacles cited are telling: data quality and silos (20%), security and data protection (19%), lack of qualified staff (18%).
At global level, the 2026 edition of McKinsey's The State of AI survey, published on 25 August 2026, sums up the gap well. It is based on 1’719 respondents in 97 countries, surveyed from 4 May to 8 June 2026. Nearly nine respondents in ten report regular use of AI in at least one function, and eight in ten say it has improved their individual productivity. Yet only 37% attribute to AI any effect, even a modest one, on their organisation's operating result (EBIT), a share that has not moved in a year. This panel includes many large organisations: 36% of respondents work in organisations with more than one billion dollars in revenue. It should therefore be read as a trend, not as a mirror of SMEs in French-speaking Switzerland.
These four sources say the same thing. Individual use is massive, company method is lagging behind, and it is this lag that a well-framed first project starts to catch up.

Choosing the first use case for deploying AI in a business
The first decision is also the most decisive. A good first use case is not the most spectacular one, it is the one that meets several conditions at once. The task comes up often, at least every week, which makes it possible to observe the effect without waiting months. It is measurable, because you already know how long it takes or how many errors it produces. It relies on documents you control. A mistake by the tool remains recoverable, because an employee reviews before anything goes to a client. Finally, at first, it does not touch sensitive personal data.
Conversely, we advise against starting with anything that decides in place of a human: sorting job applications, granting credit, setting a price client by client. These uses raise heavier legal questions, which we address below, and a mistake there is costly.
Two typical scenarios from French-speaking Switzerland
Take, as a typical scenario, an accounting firm of ten people in Fribourg. Its staff answer the same client questions every week about filing deadlines or the documents to supply. A reasonable first use case is to have AI prepare a draft reply from a base of answers validated by the partners. The employee rephrases the question without the client's data, reviews the draft, completes it and sends it themselves. Nothing goes out without them.
Another typical scenario: an estate agency in Lausanne receives many viewing requests every week through property portals and its own website. AI can summarise each request, link it to the property concerned and suggest a standard reply, while the agent stays in control of which viewings go ahead and of the personal contact. Here, prospects' personal data is in circulation, which means choosing a tool whose hosting and data use are clearly stated. In both cases, we set no quantified gain in advance: measuring it is precisely the job of the pilot.
If you are looking for other ideas, we have detailed seven use cases for automating repetitive tasks that lend themselves well to a first trial.
Take stock of your data before choosing the tool
AI is only as good as the documents you give it. The first obstacle cited by the EY survey, data quality and silos, is no coincidence. A base of outdated answers produces outdated drafts, and a customer file full of duplicates simply produces duplicates faster.
In practice, before the pilot, we sort the information involved into four categories. Public information covers your website, your general terms and conditions or your published prices. Internal information covers procedures and document templates. Next comes confidential information, such as contracts or financial figures. Finally, there is personal data, meaning anything that can identify a customer, an employee or a prospect. The first two categories can generally be used from the start of the pilot. The last two require you to know precisely where the tool hosts the data, who has access to it and whether it is used to train the model.
These are, incidentally, the questions that the Federal Data Protection and Information Commissioner (FDPIC) recommends asking before any use. It advises finding out about the manufacturer, knowing who has access to the data and whether it is passed on to third parties, and checking whether it is possible to object to its use for training. Its simplest recommendation is also the most useful for an SME: remove personal data before entering a request, and give up the service if that is not possible.
Run a short pilot, with a measurement before and after
A pilot is not a demonstration. It is an experiment that starts from a precise question, covers a limited scope and ends on a date set in advance. We frame it over a few weeks, with two or three volunteers rather than the whole team, and on a single process.
The point most projects forget is the starting measurement. Before switching anything on, you record for a week or two the time spent per file, the number of corrections or reminders, and the response time to the customer. Without this baseline, the end of the pilot comes down to an impression. And impressions are almost always favourable when you chose the tool yourself.
During the pilot, we also track what costs money: review time, which is often underestimated, the tool's errors and the cases where everything had to be redone. We also record usage fees. In the McKinsey 2026 survey, around 20% of respondents say that the operating costs of AI, including pay-per-use billing of requests, have already slowed its use in their organisation. An SME is well advised to know this cost item before scaling up.
The budget, for its part, depends mainly on the tool's licences or subscription, the internal time devoted to the pilot and to training, and any integration with your existing software. We do not give a generic range, because drafting a reply in an office tool you already pay for and connecting to your management software are not comparable. This costing is done case by case, on the real scope.
Decide on the evidence: extend, adjust or stop
On the planned date, the decision is based on the measurements, not on enthusiasm. Extend, adjust the scope or stop: stopping is as legitimate an outcome as the others. A pilot that shows the tool does not save time once review is counted has done its job, since it spared you from rolling it out to twenty people.
If the pilot is conclusive, the temptation is to add AI everywhere. The McKinsey survey suggests another path. It singles out a small group of high-performing organisations, around 6% of respondents, that attribute at least 5% of their operating profit to AI. Nearly three quarters of them say they have fundamentally redesigned their processes because of their use of AI, compared with only a quarter of other respondents. The lesson for an SME can be put in a few words: the second project is often to rethink the process around the tool, rather than to add yet another tool.
Train the team on its own tasks
General training on “AI” leaves little trace. We prefer short training built on the team's real tasks: writing a clear request, providing the right reference document, reviewing an answer and spotting an error that is plausible but wrong. The FDPIC itself points out that these tools can sometimes generate erroneous information. Review is therefore a skill in its own right, not a formality.
The challenge is also to bring existing uses out of the shadows. According to the EY survey, 29% of respondents can use external AI tools through private accounts, and for 8% this is even their only access to AI. We speak of shadow AI when these uses escape any rule. Banning them is rarely enough. It is better to offer an approved tool and explain what can be put into it and what must never be pasted into it.
Governance and legal framework: the nLPD first, European law if you target the EU
For a Swiss company, the reference is the Federal Act on Data Protection, revised and in force since 1 September 2023 (nLPD). The FDPIC reminded us in May 2025 that it applies directly to AI-based data processing, even in the absence of a specific law. It requires transparency about the purpose, operation and data sources. It allows data subjects to object to automated processing and to require that an automated individual decision be reviewed by a human. Where the risk is high, a data protection impact assessment is mandatory. The Confederation's SME portal also recalls the duty to inform whenever personal data is collected, as well as the prompt notification of the FDPIC in the event of a data security breach.
The European GDPR applies in addition only in one specific case: when you process the data of people who are in the Union, because you offer them goods or services or monitor their behaviour. An SME in Geneva that sells in neighbouring France is concerned for that clientele. A joinery firm in Valais that works only in Switzerland is not, in principle.
Where the Swiss project and the European AI Act stand
Switzerland does not yet have its own AI law. On 12 February 2025, the Federal Council decided to ratify the Council of Europe Framework Convention on Artificial Intelligence, which Switzerland signed in March 2025, and to adapt Swiss law accordingly. On 28 October 2026, it reviewed the progress of the work. The Federal Office of Justice is preparing the binding measures, which will cover transparency, data protection, non-discrimination, oversight and risk assessment. Their consultation is planned for spring 2027. In parallel, OFCOM is coordinating voluntary commitments by sector, and a guide for associations and companies wishing to adopt them has been presented to the Federal Council.
The European AI Act, which entered into force on 1 August 2024, applies in stages. The prohibited practices and the AI literacy rules have applied since 2 February 2025, those on general-purpose AI models since 2 August 2025, and the text became applicable on 2 August 2026, subject to exceptions. For high-risk systems, the so-called “digital omnibus on AI” revision, in force since 27 July 2026, has pushed back the deadlines. They fall on 2 December 2027 for areas such as employment, education or biometrics, and on 2 August 2028 for AI built into regulated products.
This regulation concerns a Swiss company as soon as it places an AI system on the European market, or when the outputs produced by its AI system are used in the Union. A company in French-speaking Switzerland that used an AI tool to screen applications for its branch in Lyon, for example, should have its situation checked before any deployment.
A one-page usage charter
For an SME, governance often fits on a single page signed by management. It lists the authorised tools and the type of account to use. It specifies the data that must never be entered, such as personal data of customers or employees, contracts or unpublished figures. It sets the rule of human review before anything is sent externally, how to inform customers or candidates when an AI is involved, and the name of the person to ask a question. Note that, according to the AXA study, only a third of SMEs that use AI have set clear rules on data protection. A page of this kind already puts you among that third.
How we support a first AI project
In a first project, our role is to help you choose the use case, prepare the data, frame the pilot and its measurement, then train the team and draft the charter. The details of this approach are on our AI and automation page. If the pilot shows that the use case does not hold up, we will tell you, and that will be a useful decision rather than a failure.
One last word of caution: the main risk is not starting too late, it is starting without knowing what you are measuring. If you want to examine your own case, the first conversation is there precisely for that.
Sources
- McKinsey: The State of AI in 2026, global survey
- Federal Statistical Office: press release on the 2025 ICT Omnibus survey
- AXA: 2025 study on the labour market and AI in SMEs
- EY Switzerland: AI Survey 2026, press release
- FDPIC: the current data protection law is directly applicable to AI
- FDPIC: using artificial intelligence in everyday life
- Confederation SME portal: new data protection law
- Federal Council: progress on AI regulation (28 October 2026)
- European Commission: regulatory framework for AI and application timeline
- EUR-Lex: Regulation (EU) 2024/1689 on artificial intelligence
- EUR-Lex: Regulation (EU) 2016/679, GDPR




