Training your team on AI: where to start?

Bringing a team to work with AI rarely starts where people think. Not by buying licenses, nor by picking a tool: by a training program. Three phases — shared awareness-building, guided practice on the company's real cases, progressive autonomy — and a clear usage framework make the difference between adoption that produces value and subscriptions that sit idle. In an SME, this program fits into a few weeks, not a multi-year plan.

Published February 2026, revised May 2026 — this version is a full rewrite.

Why licenses alone produce nothing

I spent fifteen years rolling out digital tools in large organizations — analytics platforms, collaboration spaces, content management systems. The pattern doesn't change from one technology to the next: without support, a license becomes a cost line, not a practice. With generative AI, three phenomena take hold within a few months at a company that equips without training.

Actual usage stays token. Those who were already exploring the tools on their own keep going; the rest try once or twice, don't see how to fold the tool into their tasks, and go back to their old habits. The return on the licenses stays theoretical.

Workarounds set in quietly. Employees who've spotted value but have no internal framework turn to free or personal versions of the tools — what the market calls shadow AI. The risks are concrete: personal or confidential data transferred to third-party services with no trace, in contexts where the company carries legal liability under the Federal Act on Data Protection^[1].

And false conclusions harden. Someone who got an imprecise answer or an undetected hallucination concludes that "the tool doesn't work," when it was really the use case that wasn't appropriate. That conclusion is very hard to correct after the fact.

Hence the point worth remembering: in 2026, the gap between two companies equipped with the same AI tools doesn't come from the tools — it comes from the training program. A license without support produces occasional use; a structured program produces a skill.

Who to train, and in what order?

Identical training for everyone regularly fails, because a team brings together three very different profiles when facing these tools.

The self-starters make up a minority. They already use AI in their work, sometimes without leadership knowing. Well recognized, they become valuable relays to their peers; ignored, they fuel shadow AI.

The curious make up the bulk of the workforce. They've tried once or twice, without finding the bridge to their daily tasks. This is the profile that benefits most from structured training, provided it works on their own cases — not on generic examples.

The reluctant need the most attention, and they deserve it. Their reluctance rests on real concerns: employment, quality of results, meaning of the work. Brushing them aside with efficiency arguments guarantees adoption in appearance only. Listening to them also gives you reliable information on the program's weaknesses.

A short, anonymous questionnaire is enough to map these profiles. That's the first serious step — before any tool choice.

The three phases of upskilling that sticks

Awareness-building lays the shared foundation: what models do and don't do, typical use cases for the profession, confidentiality rules. Two or three short, collective sessions, with no technical prerequisite. The goal is a shared vocabulary, not expertise.

Guided practice is the phase that decides everything. Employees practice on the company's real cases — one workshop for administrative functions, one for sales, one for technical staff — in a setting where you can try, get it wrong and start again without consequence. Framing requests, verifying answers, fitting the tool into the existing workflow: all of this is learned by doing. I don't teach anything I don't practice myself — it was by building FiscalDoc, my local tax-filing application, that I understood what a model can do and where it fails, and it's the same discipline of documented practice that produced the Score GEO™ method. Theory alone would have taught me none of it.

Progressive autonomy closes the sequence, over a longer stretch: employees identify new use cases themselves, share what works among peers, and feed an internal prompt library. You move from training mode to community-of-practice mode.

Four levels for placing everyone

To steer progress, four levels suffice, applicable to any role:

• Discovery — understanding what a language model is and telling apart relevant uses from risky ones;
• Use — practicing regularly: effective requests, verifying answers, respecting confidentiality rules;
• Integration — combining the tools into one's workflows with an observable time saving;
• Innovation — identifying new use cases and training peers.

The "use" level is the minimum goal for most of the team. The "innovation" level will only involve one or two people — and that contribution deserves to be explicitly recognized, because it's what keeps the program alive after the training ends.

Resistance is information, not a fault

Fear for one's job comes first. It's not irrational given the prevailing discourse. The honest answer lies in the distinction I develop in What skills does AI not replace?: the tools absorb tasks, and the time freed up must be reinvested in activities the employee identifies themselves — otherwise the fear was justified.

Doubt about quality comes next. It's fair: models produce imperfect results, sometimes wrong ones. The remedy is to teach systematic verification from the very first session, rather than promising a reliability the tools don't guarantee.

Lack of time comes last. "I don't have time to learn" rarely expresses a scheduling problem; it's a problem of priority and recognition. Training happens on work time, its gains get measured, and the initial investment is treated as legitimate work.

What changes in the Swiss context

Three local specifics deserve to be built into the program. Multilingualism: operating in two, three or four languages raises concrete questions — what language to phrase a request in, how to maintain editorial quality in each — that generic training ignores. Compliance: the data protection component isn't an afterthought; which data can be submitted to which tool, how to anonymize, how to document — this gets taught from the awareness-building stage onward. The economic fabric: most Swiss SMEs have neither a training department nor dedicated IT support, which calls for short, pragmatic formats, followed by ongoing support that doesn't stop when the workshop ends.

This observation shapes the way I work: training isn't an off-the-shelf product, it's built on the company's actual workflows — the same workflows that determine what deserves custom software and what stays in SaaS. Training a team on poorly chosen tools amounts to cementing a mistake; that's why I treat both questions together in the solutions I offer SMEs. And to place training within the broader sequence of initiatives — before what, after what — the starting point remains What should a Swiss SME do about AI in 2026?.

Key takeaways

— Licenses without training produce three predictable effects: token usage, free tools used on the sly, false conclusions about relevance. — Upskilling that sticks follows three phases — awareness-building, guided practice on real cases, progressive autonomy — after mapping out the profiles. — In Switzerland, the program builds in multilingualism, the data protection component, and formats suited to SMEs without a training department by default.

FAQ

How long does it take to train an SME team on AI? For a team of five to fifty people, awareness-building and guided practice generally fit into a few weeks, through short sessions built into work time. Real autonomy, meanwhile, builds up over several months of supported practice. Be wary of promises of transformation in a single day.

Should consumer tools be banned while training is underway? An outright ban fuels workarounds: employees who find value in them will turn to their personal accounts, beyond any oversight. It's better to quickly set a minimal framework — which tools, which data, which checks — then train within that framework.

Who should lead training in a small company? The owner-manager carries the framework and the priority; an already self-sufficient employee can act as an internal relay; external input is worth bringing in to structure the workshops and the compliance component. What doesn't work: fully delegating the topic to IT or to a provider without leadership involvement.

How do you measure whether the training worked? Through usage and outcome indicators defined before you start: share of the team using the tools each week, specific tasks where time savings are observed, confidentiality incidents avoided thanks to the framework. If nothing is measured, the program will remain a questionable expense come the next budget round.

And in your team? The AI Usage Diagnostic: sixty minutes to lay out your actual workflows, identify what deserves custom software, what stays in SaaS, and what needs no AI at all. Book a diagnostic

Sources

[1] Federal Act on Data Protection (FADP), revision of September 25, 2020, entered into force on September 1, 2023. www.fedlex.admin.ch/eli/cc/2022/491/fr [↩]

Jérôme Deshaie is CEO and founder of MCVA Consulting SA, an AI-augmented agency based in Valais. Fifteen years serving major international brands, now working directly with Swiss SMEs. Background.