Training teams in artificial intelligence: what installed practice calls for

Note revised on 25 May 2026. Article originally published in February 2026 — full rewrite.

Generative artificial intelligence is now accessible to every employee of a company through numerous consumer and professional tools. This accessibility has triggered a spontaneous adoption movement. It has also revealed a growing gap between organisations that deploy licences without support and those that structure a skills progression for their teams. This gap is not played out in technology. It is played out in the quality of the training arrangement.

This note sets out the observable mechanics of this difference, identifies the phases of a serious skills progression, and qualifies what Swiss SMEs gain by clarifying in their approach.

Why non-training produces a cost no one sees

In a company acquiring licences for model-driven tools without structured support, three phenomena settle in within a few months.

The observed adoption rate remains low. On deployments tracked in the firm's practice, a minority share of employees — typically those who had already explored these tools personally — use the licences regularly. The majority try once or twice, do not identify how to integrate the tool into their work, and return to their previous practices. The ROI of acquired licences remains theoretical.

The parallel use of unframed tools settles in underground. Employees who have identified operational value in model-driven tools, and who do not have a clear in-house framework of use, bypass company policies by using the free or personal versions of these tools. This practice — often referred to as shadow AI — produces real compliance risks: undocumented transfer of personal or confidential data to third-party services, unframed use in contexts where the company carries legal responsibility.

Confusion settles in over the relevance of tools. Employees who have had disappointing experiences — imprecise responses, undetected hallucinations, productions to be laboriously corrected — conclude that the tool is not useful, whereas it was the usage that was not appropriate. This conclusion hardens resistance to change, sometimes durably.

None of these phenomena is easily measurable in the company's financial statements. All reveal themselves in the medium term, when competitors that have structured their adoption display productivity the untrained organisation cannot reach.

Three employee profiles to recognise before training

Observable practice distinguishes three employee profiles in the face of model-driven tools. This distinction is useful at the initial framing of a training programme because a uniform approach for everyone regularly fails.

The already autonomous employees constitute a minority. They have explored the tools personally, identified productive use cases, and now use these tools in their daily work, sometimes without the management knowing. Their role in training is to act as operational relays among their peers, provided this contribution is recognised by the organisation.

The curious employees constitute the central share. They have heard of the tools, perhaps tested once or twice, but they have not identified how to integrate them into their daily tasks. This profile benefits most directly from structured training that shows, on their own cases, how to move from occasional experimentation to regular use.

The reluctant employees constitute the share that calls for the most attention. Their reluctance is not irrational — it often rests on fears about employment, the quality of results produced, or the meaning of work to come. Treating these fears seriously, rather than sweeping them away with efficiency arguments, is constitutive of training that reaches the whole team and not only those already inclined.

The mapping of these profiles, through a brief and anonymous questionnaire, generally constitutes the first step of a serious programme.

Three phases of a serious skills progression

Beyond the initial mapping, a structured skills progression deploys in three successive phases.

Awareness first. This phase sets the common framework: what the model-driven tools do and do not do, how they work in broad lines, what the typical operational use cases are for the company's profession, what the good practices of confidentiality and compliance are. This phase does not require technical expertise from participants. It is conducted collectively, in two or three short sessions, and it aligns the team on a shared vision.

Guided practice next. This phase is the most structuring. It consists of having employees practise the tools on real cases of the company, in a framework where they can try, fail, adjust, without risk to their operational results. Workshops by profession — one format for marketing functions, one for finance, one for technical functions — allow a targeted skills progression. Practice teaches what theory cannot transmit: prompt formulation, result verification, integration into the existing workflow.

Progressive autonomy completes the sequence. This third phase consists of letting employees themselves identify use cases in their own tasks, organising peer sharing sessions to diffuse practices that work, building an internal prompt library that capitalises on proven approaches. This phase is not conducted in classical training mode — it is conducted in community-of-practice mode, over a longer duration.

Four competence levels to structure

Beyond the temporal phases, the structuring of durable competence rests on four progressive levels, applicable to any profession.

The discovery level corresponds to understanding the principles: what a language model is, how it produces its responses in broad lines, what distinguishes relevant uses from risky ones. An employee at this level can explain in simple terms what the tool does, without necessarily using it daily.

The usage level corresponds to regular practice: formulating effective prompts, evaluating the quality of responses, respecting confidentiality rules, integrating the tool into two or three weekly tasks. This level constitutes the minimal objective of serious training for the majority of employees.

The integration level corresponds to the sophistication of practice: combination of several tools, complex chain-of-thought prompts, integration into existing workflows, measurable time gain. This level is reached by employees who invest in practice beyond the minimum, and it distinguishes efficient users from occasional ones.

The innovation level completes the hierarchy: identification of new use cases, contribution to the internal prompt library, peer training, participation in the evolution of tools used in the organisation. This level concerns a minority of employees, and it deserves to be recognised and valued as a skill in its own right.

Treating resistance as a subject, not as an obstacle

Resistances to adoption are frequent, legitimate, and carry information about the quality of the training arrangement. Three classical resistances deserve structured treatment.

Fear for employment first. This fear is not irrational in a context where public discourse oscillates between enthusiasm and catastrophism on the impact of generative models. Productive treatment consists of factually explaining what the tools absorb — tasks, not professions — and showing concretely how the time freed on these tasks can be reinvested in higher-value-added activities the employee identifies themselves.

Doubt about the quality of results next. This resistance is fair: generative models effectively produce imperfect, sometimes false results, which one must know how to evaluate. The treatment consists of integrating systematic verification practice from initial training, rather than promising reliability the tools do not guarantee.

Declared lack of time completes the list. "I do not have time to learn" is rarely the real reason; it is more often the expression of a prioritisation in tension with other demands. Productive treatment consists of integrating training into work time, measuring the operational gains that result, and explicitly recognising the initial investment as legitimate work, not a personal expense by the employee.

The specificities of the Swiss context

Three characteristics of the Swiss market modify the approach to AI training.

Structural multilingualism first. Swiss companies often operate in two to four languages, and training must cover the specificities of multilingual practice: when to prompt in the source language, when to prompt in the target language, how to manage the cultural nuances that models do not carry by default. This dimension is not anecdotal in a country where editorial quality in each language is expected.

The regulatory framework structured by the Federal Act on Data Protection^[1] next. Training must integrate a substantial compliance component: which data can be submitted to a model-driven tool, which tools are approved by the organisation, how to anonymise sensitive information, how to document processing. This legal dimension distinguishes organisations that adopt the tools seriously from those exposing themselves to risks that the evolution of the law will progressively reveal.

The fabric dominated by SMEs completes the list. A substantial share of Swiss companies does not have an internal training department, nor an IT department dedicated to supporting employees on digital tools. This structure calls for external arrangements adapted to size — short and pragmatic programmes, mixed in-person and remote format, prolonged post-training support — that differ from programmes designed for large structures.

The discipline that distinguishes the serious

Training a team in artificial intelligence is not a question of enthusiasm or technical expertise of the management. It is a question of operational discipline. Companies that succeed in their adoption pose four clear trade-offs upstream: who takes charge of the support, over what duration, with what measurable objectives, with what governance framework for usage.

This discipline does not require a substantial budget to produce its effects. It requires sustained attention from the management, explicit recognition of the time investment by employees, and patience over the duration required for effective team autonomy. This patience distinguishes, once again, organisations consolidating durable competence from those stacking licences without deriving the available value.

Artificial intelligence does not transform an organisation by its mere presence in the tools. It transforms it through the quality of practice the employees make of it, and this practice is built in the long time of training and support.

Sources

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

Jérôme Deshaie is CEO of MCVA Consulting SA, a Swiss firm specialising in strategic consulting on artificial intelligence, based in Valais.