AI implementation in Swiss SMEs: the realistic path

Few topics are discussed as often and concluded as rarely in Swiss SMEs as artificial intelligence. There are workshops, trial accounts, a few enthusiastic individuals — and six months later nothing has changed. This article describes why that happens and what an approach looks like that reaches a production use case in roughly eight weeks.

1. Why AI projects stall at the pilot stage

The most common reason is not technical but organisational: there is no use case, only an interest. "We should do something with AI" is not a project brief. Without a concrete work step whose current cost is known, there is no yardstick against which a result could be measured — and without a yardstick every trial ends in a matter of taste.

Three further patterns show up regularly:

• The tool comes before the process. A licence is procured before anyone has decided which task it should handle. The organisation then looks for a problem that fits the solution.
• Data protection stays unresolved. Because nobody wants to decide which data a model may see, everyone works with invented examples as a precaution. Those always work — and therefore prove nothing.
• Nobody owns it. AI becomes everybody's task, which in practice means nobody's as soon as daily business gets busy.

Notably, none of these reasons has anything to do with the capability of the models. Technology is rarely the bottleneck.

2. What AI does reliably in an SME — and what it does not

For planning purposes it helps to distinguish two kinds of task. Reliable are tasks where a model transforms existing information: extracting fields from a PDF, summarising a long text, drafting on the basis of templates and data, answering a question from stored documents, translating or shifting the tone of a text.

Unreliable are tasks where a model is meant to produce information it does not have: exact prices without a stored price list, legal assessments, calculations with many intermediate steps, statements about your own stock levels or appointments without a connection to the relevant systems.

As a rule of thumb: anything a new employee could do with only the existing documentation works well. Anything requiring system access or years of experience works only with the corresponding integration.

3. Five use cases with a good effort-to-benefit ratio

These five appear across industries in Swiss SMEs and can be put into production comparatively quickly:

What all five share is that a human reviews the result before it takes effect. That is exactly what makes them suitable entry points: the damage from an error is limited while the benefit is immediately noticeable.

4. Data protection under revFADP: what is actually required

The revised Swiss Federal Act on Data Protection has been in force since 1 September 2023. It does not prohibit the use of AI but it does demand clarity. For an SME undertaking that means five points in practice:

1. Define data categories. Which data may be entered into a model — and which explicitly may not? That list belongs on one page, not in a concept document.
2. Check provider and location. Where is the data processed, are inputs used for training, is there a business offering with the corresponding assurances?
3. Regulate processing. Wherever personal data is processed, a data processing agreement with the provider is required.
4. Extend the record of processing activities. The new purpose is documented where your other processing is already recorded.
5. Inform and train employees. A short policy creates more certainty than a ban that gets circumvented anyway.

Where particularly sensitive data is involved — health, salary or application data — European providers or locally hosted models are worth considering. The latter are now capable enough for many tasks and never leave the company.

5. The eight-week roadmap

The following sequence has proved realistic — provided there is a responsible person inside the company with some time available.

Weeks 1–2: collect and rate use cases

Conversations with the people who do the work. Every candidate is rated on four criteria: frequency, current time spent, data availability and cost of errors. The outcome is a short list — and a decision for exactly one use case.

Weeks 2–3: measure the baseline and settle data protection

Two things run in parallel: current metrics are recorded so a comparison becomes possible later, and the data protection questions are decided. Both are unspectacular, and both determine the credibility of the result later on.

Weeks 3–5: build and test with real cases

The solution is set up and tested with real cases — deliberately including the awkward ones: the badly scanned invoice, the enquiry with three questions in one sentence, the document in the wrong format. Testing happens against the baseline, not against a feeling.

Weeks 6–8: rollout, training, adjustment

Introduction across the team, short instructions, clear review rules. After two to four weeks in operation the setup is adjusted. Only then does the discussion about the second use case begin — and that one goes considerably faster, because access, rules and experience already exist.

6. Cost and business case

Licence and usage costs for AI tools are usually the smaller item in an SME. The larger share goes to analysis, integration and enablement — the part that determines whether the solution is still in use after six months. Building the business case on the licence alone underestimates the undertaking and overestimates the saving.

The economics can be estimated in advance: frequency of the process per month, time per run, internal hourly cost. A process that occurs 250 times a month and takes four minutes each time consumes around 17 working hours monthly. If half of that is saved, the order of magnitude is clear — and the discussion becomes factual.

7. The five most common mistakes

• Too many use cases at once. Five half-finished attempts produce less benefit than one in production.
• No baseline. Without a before-measurement every assessment stays an opinion.
• Testing with showcase examples. Edge cases decide everyday viability, not the model case.
• Data protection left to the end. Reviewing only after the build risks having to switch the result off again.
• No ownership. Without a named person the undertaking peters out at the first crunch in daily business.

Conclusion

AI in an SME is less a technology question than a question of sequence. Anyone who starts with a clearly defined use case, knows the baseline, settles data protection upfront and names a responsible person will reach a result within roughly eight weeks that stays in the business. Anyone who starts with the tool will have gathered experience after six months — but rarely something that runs.