AI and Responsible Digital: What Switzerland Can Deliver

Working with AI while staying responsible is, above all, a matter of right-sizing. Generative AI does consume energy — the International Energy Agency documents it — but most of what a user organization can actually control lies in its own choices: the model deployed for each task, how requests are phrased, hosting, measurement. A Swiss SME starts with a real advantage, thanks to a largely decarbonized electricity mix and a legal framework that already pushes toward traceability. What remains is putting it to use.

Published in December 2025, revised in May 2026 — this version is a full rewrite.

Where does AI's footprint actually concentrate?

The environmental footprint of a generative AI system breaks down into three components, and the distinction matters because each calls for different decisions.

Training large models weighs heaviest in absolute terms: thousands of accelerators running for weeks in data centers built for that density. This burden falls on model providers, not on you. Your only lever here: favor providers who document their footprint.

Inference — every request sent to a model — is the component that weighs most over time, because it operates at massive, continuous volumes. Depending on the model, the length of the response, and the infrastructure, a generative query can consume more than a standard web search; the International Energy Agency documents this demand in its report on AI^[1]. This is where your direct lever lies, because inference depends almost entirely on your usage choices.

Hardware infrastructure closes the picture: data centers, cooling, water consumption, component extraction. This component depends largely on your choice of host — and Switzerland holds particular advantages here, which I return to below.

The key point: in 2026, the question for a user organization is no longer whether generative AI consumes energy — it does — but whether each use is sized to its task. Sufficiency is decided request by request, not in statements of intent.

Five levers within your reach, ranked by impact

On the inference side, every organization has five concrete levers available, from the most effective to the most fine-grained:

• Size the model to the task — a compact model is enough for classification, sorting, translation, or summarization, at a consumption several orders of magnitude below top-tier models;
• Structure your requests — a clear request with a specified response format produces a shorter answer: every token saved counts;
• Cache recurring answers — no need to query the model again for the same question asked a hundred times;
• Choose your host — the energy mix varies widely by geography, and Switzerland ranks among the most favorable;
• Measure and report — without measurement there is no improvement, and no demonstrable practice.

The first lever is the one I apply most literally. FiscalDoc, the app that classifies my tax documents, runs on a compact open-source model installed on my Mac: no call to a giant remote model for a classification task that the local machine handles perfectly well. Same logic behind the redesign of mcva.ch — a lightweight static site, with no script calling a model on every visit. Sufficiency starts with not invoking a model when a script will do, and not invoking a top-tier model when a compact one suffices. It's exactly the sizing logic that structures the solutions I build for SMEs.

Why Switzerland starts with a head start

Three structural advantages make responsible use more accessible here than elsewhere.

First, the electricity mix: dominated by hydropower and complemented by nuclear, it shows a carbon intensity per kilowatt-hour markedly lower than that of neighbors dependent on fossil fuels. For an equal computing load, processing run on Swiss infrastructure emits less.

Then the climate: temperate, it allows passive cooling of data centers for a good part of the year, whereas hot climates require air conditioning that accounts for a substantial share of total consumption.

Finally, the legal framework: the Federal Act on Data Protection^[2] imposes transparency on data processing that indirectly pushes organizations to trace the flows feeding their models. Combined with growing interest in data sovereignty, this discipline steers Swiss companies toward local hosting — which also happens to be the most carbon-favorable choice.

Sufficiency and compliance: the same move

These two requirements, often handled by different people within a company, converge in practice. Mapping the data flows that feed your AI tools serves DPA compliance while also revealing the computing volumes behind the footprint. Favoring local hosting satisfies sovereignty requirements for sensitive data and benefits from the Swiss energy mix. Informing your users about automated processing meets legal obligations and feeds the overall transparency of the approach.

The trade-off between subscribing to an external provider and a custom tool hosted locally touches both dimensions directly — I covered it in detail in SaaS versus custom-built in Switzerland. A word of honesty here: local custom-building isn't systematically the right choice. For occasional, non-sensitive use, a well-sized shared subscription can remain the most sufficient answer. The sorting happens flow by flow, not by principle.

Responsibility becomes a commercial criterion

The environmental argument is leaving the realm of pure ethics. Tenders from major contracting authorities — the public sector, regulated industries — now build social and environmental responsibility criteria into supplier evaluation: a structured approach stops being a nice-to-have and becomes an eligibility condition in certain markets. On the consumer side, several studies converge on the finding that environmental practices weigh on brand perception; the gap between stated intent and purchasing behavior remains documented, but the pressure is real, especially among younger segments.

The real trade-off, then, isn't between performance and responsibility. It's between opportunistic adoption — stacking subscriptions without qualifying the footprint — and disciplined adoption, which frames usage, chooses models, measures, and documents. The latter costs a bit more upfront. It produces an approach that can be defended in front of a client, a tender committee, or a regulator, and it fits naturally into the roadmap I detail in What should a Swiss SME do about AI in 2026? — whose tooling component runs through a flow-by-flow sort of your usages.

Key takeaways

— Your environmental lever is inference: sizing each use — compact model, structured requests, caching — outweighs any statement of intent. — Switzerland offers three structural advantages: a decarbonized electricity mix, passive cooling, and a data protection framework that pushes toward traceability and local hosting. — Sufficiency and compliance are handled by the same move: mapping flows, choosing hosting, measuring, and documenting.

FAQ

Is generative AI compatible with a responsible digital approach? Yes, provided it's properly sized. The footprint of a professional use depends on the model chosen, how requests are phrased, the use of caching, and the host. Adoption framed around these levers stays consistent with sufficiency commitments — opportunistic adoption does not.

Is a small model really enough for professional use? For classification, sorting, summarization, or everyday translation, very often yes — at a consumption several orders of magnitude lower. My app FiscalDoc runs entirely on a compact, local model. Large models still earn their place for complex reasoning and demanding writing.

Does hosting in Switzerland really change the footprint? It reduces it significantly, thanks to an electricity mix dominated by hydropower and a climate that allows passive cooling for part of the year. It simultaneously meets sovereignty expectations for sensitive data — two benefits for a single choice.

Where to start without a dedicated budget? By mapping your current AI usage, including undeclared uses, then by applying two simple rules: the lightest model capable of doing the task, and no sensitive data sent to an unqualified service. Measurement comes next. These first steps cost nothing but framing time.

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

Sources

[1] International Energy Agency (IEA), Energy and AI — Energy demand from AI. www.iea.org/reports/energy-and-ai/energy-demand-from-ai [↩]

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

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