Artificial intelligence in recruitment in Switzerland: what practice holds, what the framework restricts

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

Artificial intelligence applied to recruitment occupies a particular place in the current conversation on AI in the enterprise. Particular because it promises immediate productivity gains on tasks recognised as costly, and because it simultaneously exposes the employer to regulatory and ethical risks above the average of professional uses of AI. This dual characteristic calls for an analytical discipline that resists both technological enthusiasm and mistrust on principle.

This note sets out what the practice of AI in recruitment seriously allows in Switzerland in 2026, and what the legal and ethical framework strictly regulates. It serves as an opening note before Cahier MCVA n°4 dedicated to AI in HR functions of Swiss companies, scheduled for October 2026.

Three uses where practice is now established

Three uses of AI in recruitment are now sufficiently established to be observed in the practice of Swiss companies.

Assisted application sorting constitutes the first established use. For an attractive position, a recruiter in Switzerland may receive several hundred applications over a few weeks. Systems driven by language-processing models extract the structured elements of a CV — skills, experience, training — whatever the form of the document, and produce a relevance ranking against the sought profile. The time gain is tangible and documented by several sector studies. The status of the system is that of an assistant: it proposes an order, the recruiter arbitrates.

Semantic candidate-position matching constitutes the second use. More sophisticated than keyword sorting, it relies on the contextual understanding of models to recognise that a skill formulated differently in two CVs covers the same reality. This capacity opens use in the "position to candidates" direction — identifying in the internal pool the best profiles for a new opening — as well as in the "candidate to positions" direction — proposing to a candidate opportunities beyond the position they initially targeted.

Pre-qualification chatbots constitute the third use. Available at any hour, capable of conducting an exchange in several languages — a particularly useful capacity in a trilingual Swiss market — they conduct a structured preliminary interview and produce a pre-selection report. Their place is strictly upstream: they never decide, they filter and orient.

What the FADP strictly regulates

The Federal Act on Data Protection (FADP), in force in its revised version since 1 September 2023, rigorously frames the use of AI in recruitment processes^[1]. Five obligations structure practice.

Information to the candidate on the use of automated tools in the processing of their application must be treated as a transparency requirement. This information must be given upstream, in the privacy policy of the recruitment site or in the application form, in accessible terms.

The candidate's right of access to the data collected about them and, in certain cases, to the algorithmic evaluations produced concerning them, is inscribed in the law. An AI system that does not permit a response to an access request within the legal deadlines constitutes a direct operational risk.

The retention period of application files cannot be indefinite without a legal basis. Archiving practices must be documented and justified.

High-risk automated profiling calls for a data-protection impact assessment. This documented analysis must precede the deployment of a system producing individual decisions on the basis of automated processing.

Information on the criteria used in an automated decision that individually affects a candidate constitutes the fifth axis. The FADP frames automated individual decisions and imposes obligations of information in the cases concerned.

What the Swiss framework prohibits

Beyond the FADP, the Swiss framework prohibits unlawful discrimination in the employment relationship and at hiring. This prohibition applies fully to algorithm-assisted recruitment decisions, with no exception on the pretext of technological mediation.

This rule has a direct operational consequence: the employer remains legally responsible for the decisions produced by the algorithms it deploys, including when those decisions are discriminatory without explicit intent. The argument that "the algorithm decided" is not admissible. The employer's responsibility is full.

What the EU AI Act adds for the Swiss companies concerned

The European Regulation on Artificial Intelligence (EU AI Act), in progressive application since 2024, classifies recruitment AI systems among high-risk systems^[2]. This qualification entails several reinforced obligations for companies operating in or exporting to the European market: exhaustive technical documentation of the system, conformity assessment prior to deployment, effective human supervision of the decisions produced, transparency on the functioning of the system, continuous post-deployment surveillance.

For a Swiss company recruiting candidates residing in the European Union, or using a system marketed by a European publisher, these obligations are added to the FADP framework. The combination of the two frameworks builds a dense but predictable regulatory environment. Actors who anticipate compliance at the initial framing produce systems that can effectively be deployed. Those who treat it as an end-of-project obstacle encounter substantial remediation costs.

The risk that does not resolve by technique: algorithmic bias

Machine-learning models reproduce the biases present in their training data, and can amplify them. This risk is known, documented for several years, and constitutes the central ethical problem of AI in recruitment.

Three manifestations are regularly observed. Gender bias, which under-evaluates or over-evaluates certain profiles in sectors where historical recruiting was unbalanced. Age bias, which disadvantages certain age brackets on the basis of correlated but non-causal reasons. Linguistic bias, which penalises candidates whose native language is not that of the posting, regardless of their actual technical competence.

None of these manifestations resolves by a purely technical improvement of the model. Correction goes through an operational discipline that combines several levers. Diversification of training data, which reduces the correlations learned by default. Regular audit of results by demographic stratum, which identifies effective gaps and flags biases that have settled in. Effective human supervision of individual decisions, which maintains an arbitration point outside the automatic system. Transparency on the criteria used, which makes appeals and challenges possible.

This discipline is not an optional supplement. It is constitutive of a serious use of AI in recruitment.

Five operational principles that distinguish serious practices

For a Swiss company preparing or revising its use of AI in recruitment, five operational principles emerge from the observation of practices.

The final hiring decision remains human. AI pre-selects, ranks, recommends, signals; it does not decide. This boundary protects the company legally and preserves the quality of recruitment.

Audit of results by demographic stratum is conducted at a regular cadence. Quarterly at a minimum, monthly for high-volume structures. This audit identifies biases that may have settled in and triggers the necessary corrections.

Information to the candidate is explicit, in accessible terms. The site's privacy policy and the application form clearly signal automated processing, without using technical language that dilutes the message.

The recourse channel is documented. A candidate rejected by an automated system must know how to contest that decision, and the company must be able to respond within reasonable deadlines.

Training data is qualified before use. The generic data sets supplied by publishers do not necessarily match the reality of the Swiss labour market, nor the profiles the company seeks to recruit. Prior qualification and regular adjustment are required.

What this note does not claim to do

This note sets out an operational framework. It prescribes no particular tool — the market for AI-assisted recruitment solutions evolves too rapidly for a nominal recommendation to make sense within six months. Nor does it propose pricing — ranges vary considerably by recruitment volume, sector of activity, existing infrastructure, and the level of outsourcing retained.

Cahier MCVA n°4, scheduled for October 2026, will treat in greater depth the adoption of AI in HR functions of Swiss companies, integrating recruitment, training, talent management and strategic workforce planning. It will rely on a production of observations specific to the Swiss market and on coordination with identified HR practitioners.

AI in recruitment is neither a miraculous productivity promise nor a technological threat to HR functions. It is a tool whose value strictly depends on the rigour of the framework in which it is deployed.

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 [↩]

[2] Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence (EU AI Act). eur-lex.europa.eu/eli/reg/2024/1689/oj/eng [↩]

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