Many small-business AI strategies start with a purchase. Licences for everyone first, the question of what for comes later.

That is the expensive order. Not because the tools are bad, but because nobody can tell any more whether they help. If you never decided which task should get faster, all you can measure afterwards is an impression.

An AI strategy for a small firm is not a forty-page document. It fits on one page and answers four questions: which task, measured how, with which data, using which tool. In exactly that order.

Where Swiss SMEs stand today

Use is already widespread – often ahead of the rules for it.

  • 41 % of the Swiss labour force used generative AI at work in the three months before they were surveyed (Federal Statistical Office, ICT Omnibus 2025, published April 2026).
  • 35 % of SMEs have deliberately built AI into their processes, 39 % are trying it, 26 % use none (AXA/Sotomo, SME labour market study 2026, 336 firms with five or more employees).
  • Only a third of SMEs that use AI have a rule on which company data may be entered into AI tools (AXA/Sotomo, SME labour market study 2025).

So the gap is rarely access. It is the question of what AI is used for and who sets the limits. That is what strategy means here.

Step 1: The task first, the tool second

A good first task has three properties:

  1. It repeats – at least ten times a week: a customer enquiry, a quote, meeting minutes.
  2. It is clearly bounded – everyone knows when it is done.
  3. Its result can be checked – someone can tell whether it is right.

Why tasks like these? Because that is where the effect has been measured most cleanly. At one software company, 5,172 support agents with AI assistance resolved 15 % more issues per hour on average (Brynjolfsson, Li, Raymond, Quarterly Journal of Economics, 2025).

The frontier you cannot see

AI is not good or bad for whole professions, only for individual tasks. In an experiment with 758 consultants, those using AI inside that frontier became faster and better. On a task outside it, they were right less often with AI than without (Dell'Acqua et al., Organization Science, 2026).

So the question is never "Does AI help in sales?" It is "Does AI help draft this quote, and how will we notice when it gets it wrong?"

Which workflows typically qualify is broken down in 12 processes you can automate today.

Step 2: Measure first, or you are measuring a feeling

This is the step most often skipped – and the most expensive one to skip.

In a randomised experiment with 16 experienced developers and 246 tasks, the developers estimated afterwards that they had needed 20 % less time with AI. Measured, they needed 19 % more (METR, 2025). The estimate was not slightly off. It was on the wrong side of zero.

For your firm that means:

  • Time each case to "done and checked", not to the first draft.
  • Split the cases for two weeks, alternating with and without AI – not by preference.
  • Record rework: what had to be corrected, and how long it took.

You will not have a study at the end, but you will have your first measured KPI. It is worth more than any projection from a sales meeting. A template for that list is in the guide AI in Swiss SMEs.

Step 3: A data rule in one sentence

Before a team works with AI, it needs one sentence that states which data never goes into an AI tool. For example: "Our customers' personal data only goes into tools whose provider we have a contract with."

This is not a formality:

  • The Federal Data Protection and Information Commissioner states that the Data Protection Act is directly applicable to AI-supported data processing (FDPIC, 2023). No separate AI law is needed for that.
  • 74 % of employees in Switzerland who use AI at work have relied on an AI output without checking it. 52 % have used AI against their organisation's policies (KPMG and University of Melbourne, 2025).

The second figure is the point: a rule on its own is not enough – people have to know it and live by it. Send the sentence to the team the same day.

Step 4: The tool you may already own

Only now does the tool question come up, and it is often smaller than expected.

  • Microsoft 365 already includes automation in many licences. Where the included licence stops is covered in Power Automate for SMEs.
  • A language model (LLM) for drafts, summaries and text. Where it pays off and where it does not: 20 ChatGPT use cases.
  • A website chatbot answers routine questions – but only if you are prepared to stand behind its answers (see below).

Free tools, with a caveat

Many tools have a free tier, and for a first test it is often enough. Check the terms of use, though, for whether your inputs may be used for training, and stick to your data rule from Step 3. A private account is a tool with no contract with you.

Whoever deploys a chatbot should expect to answer for what it says. The Civil Resolution Tribunal of British Columbia (Canada) ordered Air Canada to pay damages because its chatbot gave a customer wrong information (Moffatt v. Air Canada, 2024).

Step 5: Bring the team along, newcomers first

Technology rarely fails because of the technology. It fails because nobody knows what is expected of them.

The research offers a surprising pointer here: the least experienced gain the most. In the same support study, output of less experienced agents rose by 30 %, while for the most experienced there was hardly any measurable effect (Brynjolfsson, Li, Raymond, 2025).

Two practical rules follow:

  • Start where people are being trained in, not with the specialists.
  • Make checking part of the job: an AI output is a draft until a person has checked it.

What an AI strategy is not

Three misunderstandings cost the most:

  • A licence for everyone. Without a task and a measurement you are buying access, not impact.
  • A reference to "95 % of AI projects fail". The figure comes from a report marked "Preliminary Findings", based on 52 interviews, 153 survey responses and a review of publicly known projects – with no control group (MIT NANDA, 2025). It works neither as a warning nor as an argument.
  • A project with an end date. Tools change; what sits outside the frontier today may sit inside it a year from now. That is why the measurement list from Step 2 has a shelf life.

What getting started costs

Less than most people expect – as long as the order is right. The steps up to your first measured figure mainly cost time: pick a task, measure for two weeks, write down a rule.

It gets expensive when the tool comes before the task. A European Investment Bank working paper on big-data analytics and AI shows that it takes more than the licence: productivity gains came mainly in medium-sized and large firms, and were larger where firms also invested in software, data and training (Aldasoro et al., EIB, 2026).

If you would rather not do this alone: in our AI consulting we take exactly these steps with you – potential analysis, use cases, measurement, implementation plan.

Conclusion

An AI strategy for a small firm is a sequence, not a budget: task, measurement, data rule, tool, team. Start that way and after two weeks you have a first measured figure and a rule the team knows. Start with the licence and after two weeks you have an invoice.

Where to read on

Frequently asked questions (FAQ)

How do I start using AI in my business?

With one task that repeats at least ten times a week and whose result someone can check. For two weeks, alternate with and without AI and time each case to "done and checked". Only then is it worth asking which tool fits.

Do I need in-house specialists to get started with AI?

No. Getting started does not take a data science team. It takes someone who knows the task, measures the time and checks the results. Outside expertise pays off mainly when choosing use cases and connecting AI to existing systems.

Are there free AI tools for businesses?

Yes. Many tools have a free tier that is enough for a first test. Check the terms of use for whether inputs may be used for training, and do not enter personal data as long as you have no contract with the provider.

What are the risks of getting started with AI?

The three most common are unchecked outputs, company data in tools without a contract, and chatbot answers you have to stand behind. A data rule and a duty to check limit all three.

How do I find the right AI use cases?

Look for where your firm does the most repeatable manual work: enquiries, quotes, minutes, filing. That is where the effect is best documented. Tasks with hard-to-check results or that rarely come up are not a good place to start.