Operational Glossary

AI agent

An AI system capable of acting semi-autonomously on an environment (searching for information, drafting documents, sending emails, booking resources) from a general instruction. Different from a simple conversational assistant: an agent chains actions together without step-by-step supervision. The challenge for organisations is to define the boundaries of authorisation and human oversight.

Apertus

The first large-scale, multilingual, open-source language model developed in Switzerland. Launched on 2 September 2025 by EPFL, ETH Zurich and the Swiss National Supercomputing Centre (CSCS) as part of the Swiss AI Initiative. Trained on 15,000 billion tokens covering more than 1,000 languages, including Swiss German and Romansh, Apertus is available in two sizes (8 billion and 70 billion parameters) and complies with the requirements of the revised Federal Act on Data Protection (nFADP) and the European AI regulation. Its full documentation (architecture, weights, training data, methodology) is freely accessible, which sets it apart from partially open models.

Basculement compétitif

A term specific to the essay. It designates the reversal of the balance of power between qualified local players and offshore or generic SaaS competitors, made possible by well-orchestrated AI. The central argument: a local Swiss team, equipped with the right tools, can now produce under economic conditions it had not enjoyed since the industrialisation of services.

Bisse cognitif

The essay's central metaphor. By analogy with the bisses — the irrigation channels that capture water where it is abundant (glacier, torrent) to carry it where it is scarce (vineyards, meadows) — generative AI captures the cognitive expertise accumulated in the major urban hubs and makes it available in the valleys. The bisse cognitif does not create expertise: it redistributes it.

Capital humain qualifié résident

A key term of the essay. It designates the population of qualified working people who live and work year-round in the canton, as opposed to seasonal second-home residents or commuters. The essay holds this to be the central variable of Valais's economic trajectory over the next ten years: everything else (tourism, institutions, heritage, energy) only makes sense if this population is sufficient in size, engaged, and properly equipped.

Cloud souverain / SGC

The Swiss Government Cloud is Switzerland's state cloud infrastructure, approved by the Federal Chambers at the end of 2024 (budget: ~320 million francs). It will provide federal, cantonal and communal administrations with hosting compliant with Swiss law. The private sector is excluded from it. First functionalities will be available in early 2026, with full deployment planned for 2032.

Corpus

A structured set of textual, audio or visual data on which an AI model is trained or fine-tuned. A corpus of Walliserdeutsch, Valais oenological data, or alpage records is a strategic asset: it enables the production of specialised models that global players will have no interest in developing.

Dilution du capital cognitif

A term specific to the essay. It designates the process by which AI reduces the relative scarcity of the cognitive capital concentrated in urban hubs, without eliminating it. Capital remains denser in metropolitan centres, but exclusive access to it is gradually disappearing for intermediate uses. It is this process that opens the territorial window.

Fine-tuning

The process of refining a large general model on a specialised corpus to adapt it to a specific domain. Example: fine-tuning a general language model on rulings from the Valais Cantonal Court to produce a legal assistant adapted to Swiss cantonal law. There is no need to build a model from scratch: this is economically accessible at the scale of a regional institution.

Grammaire de gouvernance

A term specific to the essay. It designates the coherent set of rules, practices and structures that make it possible to manage collectively what cannot be managed alone. The Valais grammar of the commons (bourgeoisies, consortages, bisses) is a governance grammar tested over seven centuries, transposable, the essay argues, to today's cognitive commons.

Hallucination

The phenomenon by which an AI model produces a false answer with a high degree of apparent confidence: a non-existent citation, an invented statistic, incorrect legal reasoning. A critical risk in professions where precision carries consequences (law, medicine, taxation). Human oversight, particularly by senior practitioners, is the only reliable safeguard.

Hyperscaler

A very large-scale cloud operator: Amazon Web Services (AWS), Microsoft Azure, Google Cloud. These players invest tens of billions annually in their infrastructure. No European territory, not even at EU scale, has managed to build a competitive generalist alternative. The correct strategy for a canton is complementarity on specific uses, since substitution is impossible.

Modèle de fondation

A large AI model trained on massive quantities of data (hundreds of billions of parameters) that serves as the basis for specialised applications. GPT-4, Claude, Gemini, and Apertus (Swiss) are foundation models. A decision-maker does not need to understand their inner workings, but must understand that their organisation depends on these models and that choosing a provider is a matter of sovereignty.

nLPD

The revised Federal Act on Data Protection (nFADP), which came into force on 1 September 2023. It governs, among other things, the processing of personal data by AI tools, data transfers abroad, and the obligations of data processors (Data Processing Agreements). It applies to any Swiss organisation using AI tools, including tools from foreign providers.

Prompt / prompt engineering

A prompt is the instruction given to an AI model to obtain a response. Prompt engineering is the art of formulating these instructions in a way that produces superior-quality results. Contrary to intuition, this skill is less a matter of technical computing know-how than of clarity of thought, precision of framing, and domain knowledge. This is why seniors who know their field well formulate better prompts than young people who know the tools well.

RAG (Retrieval-Augmented Generation)

A technique that allows an AI model to respond by first searching a defined documentary base (your firm's archives, your internal documentation, your regulatory corpus) before generating its answer. It greatly reduces the risk of hallucination and makes it possible to build specialised assistants without training a model from scratch.

Senior qui orchestre

A term specific to the essay. It designates the human profile that determines the full effect of the basculement compétitif: an experienced professional capable of architecting a problem, orchestrating the production of an AI chain, and validating the final quality. Different from an AI specialist or a young graduate skilled with the tools: it is accumulated professional experience that makes the difference, not technological mastery.

Souveraineté pratique

A term specific to the essay. It designates digital sovereignty as it is actually built — not in large infrastructures, but in the everyday choices of economic actors: which tools, which data, under which jurisdiction. Example: a fiduciary firm that chooses an AI assistant compliant with Swiss law over a more powerful American tool is exercising practical sovereignty.

Stratégie de complément / stratégie de substitution

A central distinction from chapter 11. A substitution strategy aims to replace the hyperscalers on their own generalist ground; it is doomed to fail (as Gaia-X did). A complement strategy targets specific uses that the hyperscalers do not serve: medical data under Swiss law, regional linguistic corpora, needs for alpine winter resilience. This is the only viable strategy at cantonal scale.