Artificial intelligence in Swiss healthcare: what practice holds

Note revised on 25 May 2026. Article originally published in November 2025 — full rewrite.

The Swiss healthcare sector occupies a singular position in the public conversation on AI. At once a site of very high expectations — accelerated diagnosis, personalised medicine, transformed research — and a terrain where the slightest algorithmic error has consequences that few other sectors bear, it calls for an analytical discipline that resists both symmetrical temptations of enthusiasm and mistrust.

This note sets out what observable practice in 2026 allows us to assert with a certain firmness, and what it leaves open. It serves as an opening note before Cahier MCVA n°3 dedicated to this sector, to be published in the first quarter of 2027.

Three domains where practice is established

Three AI usage domains in Swiss healthcare have left the experimental stage and settled into the routine practice of institutions and companies in the sector.

Assisted medical-imaging analysis is probably the most mature domain. In radiology, dermatology and ophthalmology, diagnostic-aid systems driven by deep-learning models are tested, integrated or deployed in some hospital and clinical contexts. Their status is that of an automated second opinion: they flag elements for the practitioner to examine, without substituting themselves for clinical decision. The available scientific literature documents high performance on certain precise detection tasks. This performance holds on the measured tasks, under the conditions of the studies, and does not extrapolate mechanically to all clinical practice.

Administrative management constitutes the second domain where use is now routine. Transcription of consultations into structured reports, attribution of diagnostic codes, optimisation of hospital schedules and resources: these repetitive, high-volume, low-creativity tasks are absorbed by systems driven by generative models without direct clinical risk. The time gain released for the physician and care staff is the most solid economic argument in favour of adoption.

Decision support in clinical research constitutes the third established domain, mainly in the pharmaceutical industry and in academic research centres. Large-scale analysis of the scientific literature, identification of candidates for clinical trials on the basis of anonymised patient files, acceleration of molecule screening: these uses operate at a level where the occasional error is absorbed by the downstream scientific validation process.

The regulatory framework that structures practice

AI in healthcare in Switzerland operates under a dense regulatory framework that combines several sources of obligations.

The Federal Act on Data Protection (FADP), whose revised version came into force on 1 September 2023, classifies health data as sensitive data subject to a reinforced level of protection^[1]. Any use of AI models processing patient data requires either explicit consent, or documented irreversible anonymisation, or a specific legal basis. The FADP frames automated individual decisions and imposes obligations of information in the cases concerned.

The European Regulation on Artificial Intelligence (EU AI Act), in progressive application since 2024, classifies medical devices integrating AI as high-risk systems, subject to reinforced requirements of technical documentation, conformity assessment, risk management and post-market surveillance^[2]. Swiss companies exporting to the European Union or addressing a European market must integrate this framework into their development cycle.

Swissmedic, the Swiss regulatory authority for therapeutic products, is progressively harmonising its requirements with international standards, while preserving the specificities of the Swiss healthcare system. Medical devices integrating AI must satisfy the applicable conformity procedures before being placed on the market; Swissmedic intervenes in particular in market surveillance^[3].

The outcome of these three regulatory layers is an environment more demanding than it was five years ago, but also more predictable. Actors who anticipate compliance from the framing produce devices that can actually be deployed. Those who treat compliance as an end-of-project obstacle encounter substantial remediation costs.

The question of data sovereignty

AI in healthcare raises a strategic question specific to Swiss institutions: where is patient data stored, and under which law? Systems driven by generative models hosted on US clouds may raise questions of applicable law and extraterritorial access to data, which can come into tension with Swiss data-protection requirements.

Several operational responses are available. Hosting with Swiss-residency providers, which have multiplied since 2020, constitutes the first. Deployment of open-weight models on controlled infrastructure, now possible for reasonably sized models reaching useful performance, constitutes the second. Federated learning, which trains models without centralising the data — each institution keeps its data locally, only model parameters circulate — constitutes the third, still more marginal in practice but actively explored by several Swiss university hospitals.

None of these responses is universally superior. Each has costs, performance limitations, specific technical requirements. The choice falls under a qualification of risk by use case and by type of data processed.

Three points of vigilance that do not resolve by technique

Beyond the established uses, three points of vigilance run through the practice of AI in Swiss healthcare, and none resolves by a purely technical improvement of the model.

Algorithmic bias continues to affect models trained on historical data that under-represent certain populations. This limitation has been known and documented for several years. Its correction goes through the diversification of training data, systematic validation on sub-populations, and transparency on the scope of validity of a deployed model.

Medical liability in the event of algorithmic error remains legally under construction. The majority position in Swiss law maintains that the physician remains the final decision-maker and retains clinical responsibility. AI is a support tool, not an autonomous decision-maker. This position is comfortable for the experienced practitioner who knows when to follow and when to ignore an algorithmic recommendation; it weighs on the practitioner who lacks the time or expertise required to exercise that judgement.

Patient trust in medical decisions partly driven by a model is a subject that exceeds regulatory compliance. It is built on transparency — information to the patient when AI intervenes in their care pathway — and on the overall quality of the care relationship, which remains irreducible to any algorithmic improvement.

What this note does not claim to do

This note does not exhaust the subject. Cahier MCVA n°3, under preparation for the first quarter of 2027, will treat in greater depth the strategic implications for hospitals, health insurers, the pharmaceutical industry and medtech SMEs. It will rely on a dedicated production of observations and on coordination with an identified healthcare-sector co-author.

The practice of AI in Swiss healthcare is not measured solely against the promises it holds in the laboratory or the risks it stirs in comment threads. It is measured against the rigour of the regulatory frameworks it respects, the solidity of the deployment conditions it establishes, and the patience with which it settles into institutions where error is costly.

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] European Parliament, EU AI Act: first regulation on artificial intelligence. www.europarl.europa.eu/topics/en/article/20230601STO93804/eu-ai-act-first-regulation-on-artificial-intelligen [↩]

[3] Swissmedic, Overview of medical devices. www.swissmedic.ch/swissmedic/en/home/medical-devices/overview-medical-devices.html [↩]

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