AI for SMEs: getting started – first use cases and a 90-day plan

What "getting started with AI" really means for an SME

Getting started with AI is not a large IT project, nor a question of company size. For most Swiss SMEs it begins exactly where staff already lose time on recurring, language-based tasks: emails, quotes, minutes, translations, research. Today's generative AI is above all a tool that produces text, drafts and summaries – fast, but not error-free. That is precisely why the right entry point is small and controlled, not big and risky.

A realistic ambition for the first few months is not to "transform the company", but to complete two or three tasks per person per week noticeably faster and at the same or better quality. Starting this way builds experience, trust and reliable numbers – the foundation for everything that follows.

Low-risk starting points: where to begin without big exposure

Not every use case is equally suited to the beginning. Good candidates are tasks where a human reviews the result anyway, where mistakes are not immediately costly, and where no sensitive personal data is involved. The following criteria help with selection:

• Human in the loop: a result is always checked by a qualified person before it leaves the company.
• Low cost of error: a failed draft costs minutes, not customers or compliance breaches.
• No sensitive data: start without health, applicant or financial data and without confidential contracts.
• High frequency: tasks that recur daily or weekly deliver measurable value quickly.
• Easily judged output: you can immediately tell whether the result is good – for example a summary or translation.

First realistic use cases by function

The strongest entry-level use cases look similar across industries, because the same routine language tasks arise everywhere. The following examples are illustrative and deliberately low-threshold – they can usually be implemented with tools you already have (such as a business-licensed AI assistant):

• Communication: draft emails and quotes, suggested replies to enquiries, translations across DE/EN/FR/IT – always with a human final check.
• Knowledge & documents: summarise long documents, minutes or meeting notes and turn them into action items.
• Marketing & content: first drafts for blog posts, product copy or social media posts, refined internally.
• Sales & support: answer common customer questions in a structured way, maintain FAQs, tidy up call notes.
• Light analysis: review tables and feedback, surface patterns and open questions – as preparation, not as a substitute for decisions.

The 90-day plan: from curiosity to proven value

A clear time frame stops an AI start from fizzling out into vague experimentation. Three 30-day phases work well: first understand and set rules, then test in a tight pilot, then decide and roll out cleanly.

• Days 1–30 – Foundations: pick two or three use cases, write a short usage policy (what may go in, what may not), choose business-grade access with data-protection guarantees, name 3–5 pilot users, and measure the baseline (time spent today).
• Days 31–60 – Pilot: test in daily work, collect good prompts and templates, hold a short weekly exchange (what works, what doesn't), document results. The goal is learning, not perfection.
• Days 61–90 – Decide & roll out: compare time saved and quality against the baseline, lock in successful use cases as standard, run a short team training, and define responsibilities and the next two use cases.

Data protection and the Swiss angle: revFADP from day one

Swiss SMEs fall under the revised Data Protection Act (revFADP, also nFADP). Among other things it requires you to know which personal data you process, to keep processing proportionate and transparent, and to govern any outsourcing to service providers by contract. When personal data is transferred abroad – the norm with cloud AI – you need an adequate level of protection or appropriate safeguards. The FDPIC (EDÖB) is the competent supervisory authority.

In practice: use business AI offerings where your inputs are not used for training, avoid entering sensitive data at the start, and record what is allowed in your usage policy. These notes are not legal advice – for sensitive processing, a professional review is worthwhile.

Judging cost and ROI correctly – without number games

For entry-level use cases, per-user licence costs are usually modest. The costs that matter lie elsewhere: onboarding time, training, building good templates, and quality assurance. Calculate the benefit honestly: minutes saved per task times frequency times number of people – minus the time for review and rework. A use case that reliably saves a few hours a week is worth more than ten impressive demos nobody uses.

Common mistakes when starting with AI – and how to avoid them

• Starting too big: an enterprise AI strategy before the first real test. Better: a small pilot, real numbers, then scale.
• Blindly trusting output: AI invents plausible-sounding errors. Without a human final check, that gets expensive.
• Data protection as an afterthought: typing sensitive data into a consumer tool and asking questions later. Rules belong at the start.
• No measurement: without a baseline you cannot prove value – and the pilot evaporates into gut feeling.
• Forgetting the team: rolling out a tool without explaining it. A short training and shared templates decide adoption.
• Trying to automate everything at once: assistive AI is enough to start. Fully automated flows need more maturity and control.