Choosing an AI consulting firm in Switzerland: what distinguishes a serious approach

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

The Swiss artificial-intelligence consulting market has become denser as generative AI has become a board-level topic. Structures of every kind have appeared: generalist strategy firms that have added an AI practice, relabelled digital agencies, specialised independents, technology integrators widening their offering, academic spin-offs. The diversity of supply is not a problem in itself. It becomes one when the executive looking for a partner has no framework to distinguish a serious approach from an opportunistic offering.

This note sets out the framework the firm uses to qualify an AI provider in the Swiss context. It does not establish a ranking; it proposes verifiable criteria.

The structural context Switzerland offers

Switzerland has a recognised academic and technical ecosystem in artificial intelligence. The federal institutes of technology and several research institutes active on Swiss territory maintain a sustained scientific output in machine learning, language processing and computer vision. Since 2024, a federal programme has been structuring national coordination between universities of applied sciences and economic actors.

This academic density feeds a fabric of spin-offs and specialised SMEs that will not disappear. For a Swiss company looking for an AI partner, the market for competent providers is real; there is no shortage. The decision problem is not scarcity, it is qualification.

The first criterion: what is the firm effectively selling?

A serious approach begins by stating what it does not sell. A firm that presents AI as a technological break to be negotiated through a chain of proofs of concept is selling a grammar of urgency that holds neither over the duration of a mandate nor in the feedback gathered after two years. A firm that announces a quantified return on investment before having measured the initial state is selling a promise, not a method.

A serious approach sells a logic of measurement. This means: a diagnosis that establishes the current situation of the company on the observed scope (whether visibility, productivity, automation or internal use of generative models), an explicit prioritisation of corrections, and a stable cadence of re-measurement over time.

The test is simple. In the first hour of conversation, does the firm propose a sale of services or a qualification framework for the situation? If it sells, it has not begun to understand. If it qualifies, it has the posture that will make the mandate possible.

The second criterion: FADP compliance built into the framing, not added at the end

The revised Federal Act on Data Protection, in force since 1 September 2023, imposes a demanding framework on processing of personal data, particularly when such processing involves international transfers or automated decisions. Generative models accessed via APIs hosted outside Switzerland raise specific questions^[1].

A serious firm integrates this dimension from the framing of the mandate. It does not treat the FADP as a checkbox at the end of a project. It proposes, according to the sensitivity of the data concerned, an architecture that can combine Swiss hosting for the most sensitive data, open-weight models deployable on controlled infrastructure for use cases where autonomy matters, and public APIs only for uses that touch no personal or strategic data.

The Swiss hosting ecosystem effectively allows this layered architecture. Several actors have offered Swiss-residency cloud infrastructure for years, and the major international providers have structured operational Swiss regions. The choice between these options is not an ideological preference, it is a qualification of risk by use case. A firm that cannot conduct this qualification should not conduct the mandate.

The third criterion: limited technological lock-in

The market for generative models evolves at a pace that renders obsolete any technological bet made at a given moment. A firm that anchors its practice to a single publisher weakens the companies it accompanies.

A serious approach maintains operational technological independence. This means the capacity to compare the available models on a given use case, the capacity to migrate a production system from one provider to another without complete rebuild, the capacity to use open-weight models when the context demands it.

The practical test: ask the firm, on a specific use case, why it recommends one model rather than another, and what would happen if a migration became necessary in six months. A reasoned answer signals real practice. An evasive answer signals a disguised single-vendor practice.

The fourth criterion: measurement of AI citability when it is in scope

For mandates that touch on business visibility in AI-augmented environments, an additional criterion has applied since May 2026: does the firm propose a codified measurement of citability, or does it confine itself to producing content while betting it will be cited?

The official Google doctrine published on 15 May 2026 on optimising for the search engine's generative features clarified that the central ranking systems govern these features^[2]. This clarification disqualifies approaches that present optimisation for AI as a discipline separate from SEO. It does not disqualify the measurement of citability itself, which remains a strategic object. But it invites clients to require providers to know how to measure what they promise to improve.

MCVA Consulting SA has stabilised the Score GEO™ for this purpose, set out in full in the Cahier MCVA n°1. Other methods exist and can legitimately coexist. The criterion is not to use this or that instrument, it is to use one, transparently and reproducibly.

The evaluation framework in practice

Taken together, these four criteria draw a simple qualification grid. The firm proposes a logic of measurement rather than a catalogue. It builds the FADP into the initial framing. It maintains operational technological independence. It knows how to measure what it promises to improve, or it honestly says it does not know.

This grid does not exhaust the subject. It says nothing about sector-specific expertise, the human quality of the relationship, effectively verifiable references, or operational reliability. But it filters the essential: the posture of method, as opposed to the commercial posture.

A Swiss company preparing an AI mandate will find here, if not certainty, at least an arbitration framework that holds up against marketing pressure.

Sources

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

[2] Google Search Central, Optimizing your website for generative AI features on Google Search, published 15 May 2026. developers.google.com/search/docs/fundamentals/ai-optimization-guide [↩]

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