EU Study on the Deployment of AI in Healthcare: Challenges, Accelerators, and Future Considerations
TLDR
Commissioned by the European Commission's Directorate-General for Health and Food Safety and conducted by PwC EU Services and Open Evidence, this study examines the current state of AI deployment in European healthcare across the full breadth of the topic: what is working, what is not, and what structural interventions would most effectively accelerate responsible adoption. Drawing on a systematic review of over 14,000 articles, stakeholder surveys across five groups, interviews, workshops, and four country case studies, it maps four challenge domains, corresponding accelerators, and five future considerations for EU-level action. Against a backdrop of a projected 4.1 million healthcare worker shortfall by 2030, the study frames AI not as an optional enhancement but as a structural necessity for European health systems.
A Mixed-Methods Portrait of AI in European Healthcare
Studies of AI in healthcare tend to be either narrowly technical, focusing on the performance of specific algorithms, or broadly normative, outlining what responsible AI should look like. This study is neither. Commissioned to inform European Commission policy, it attempts to describe the deployment landscape as it actually exists: where AI is being used, where it is stalling, what is driving adoption, and what is preventing it. That empirical ambition required a substantial methodological effort.
The research team screened 14,407 academic articles, reviewing 173 in full depth and a further 64 for specific supplementary questions. They surveyed patients, healthcare professionals, hospital managers, AI developers, and regulatory experts. They conducted interviews and workshops with stakeholders across EU member states. They examined four case studies in depth: radiology AI in a major hospital network, AI-assisted triage in emergency medicine, predictive analytics for chronic disease management, and AI in mental health service delivery. The result is a portrait of AI deployment that is recognisably European in its institutional complexity, regulatory density, and variation across member state contexts.
The manuscript was completed in June 2025, and the study carries a DOI of 10.2875/2169577. It is one of the most comprehensive assessments of AI in European healthcare produced to date and represents a significant evidence base for the policy discussions that will shape EU-level action in the coming years.
Four Challenge Domains
Technological and Data Challenges
The most frequently cited barriers to AI deployment in European healthcare are data-related. Health data in Europe is fragmented across institutional silos, stored in incompatible formats, and governed by legal frameworks that, while essential, create significant practical obstacles to the data sharing that AI development and validation require. Training datasets that are too small or too homogeneous produce models that do not generalise; models trained on data from one country or one hospital system often perform poorly when deployed in different contexts. Interoperability standards exist in principle but are inconsistently implemented in practice. The European Health Data Space, the EU's flagship initiative for creating a coherent framework for health data access and reuse, was still in implementation at the time of the study and had not yet resolved these challenges.
Infrastructure challenges compound the data challenges. Many healthcare facilities across the EU, particularly in lower-income member states, lack the computing infrastructure, the connectivity, and the digital systems integration that AI deployment requires. A sophisticated diagnostic AI is of limited value in a hospital where radiology images are stored in a different system from the electronic health record, and both are inaccessible from the clinical workstation where decisions are made.
Legal and Regulatory Challenges
European healthcare AI operates within one of the most complex regulatory landscapes in the world. The AI Act, the Medical Device Regulation and In Vitro Diagnostic Regulation, the Product Liability Directive, the Health Technology Assessment Regulation, and the European Health Data Space regulation all interact with each other and with national-level health regulation in ways that are not always coherent. Developers struggle to navigate overlapping requirements; healthcare providers are uncertain about their liability when using AI tools; and regulators themselves are in the process of working out how the various frameworks apply to AI systems that do not fit neatly into existing regulatory categories.
The study found that regulatory uncertainty was a significant deterrent to investment, particularly for smaller developers who lack the resources to maintain specialist regulatory expertise across multiple frameworks. The MDR and IVDR, which came into full application between 2021 and 2022, created significant compliance burdens for existing device manufacturers and introduced delays that affected the availability of AI-enabled devices in European markets.
Organisational and Business Challenges
Healthcare organisations face structural barriers to AI adoption that go beyond data and regulation. Procurement processes designed for physical equipment are poorly suited to software-as-a-service AI tools that update continuously and whose value depends on ongoing vendor relationships rather than a one-time purchase. Financial models that fund healthcare through activity-based payments create perverse incentives around AI adoption: efficiency gains that reduce activity may reduce revenue for providers, even when they deliver better patient outcomes. Change management is consistently underinvested: institutions acquire AI tools without adequately preparing the clinical and administrative staff who will use them, and without redesigning the workflows that the tools are intended to improve.
Social and Cultural Challenges
Trust is the social dimension that runs through every other challenge. Patients are uncertain about how AI is being used in their care and whether their data is being protected. Healthcare professionals are concerned about deskilling, about liability, and about the reliability of AI recommendations in situations where they cannot independently verify the reasoning. There is a broader cultural gap between the quantitative optimisation logic of AI systems and the relational, contextual character of clinical care, particularly in settings like primary care, mental health, and palliative medicine where the therapeutic relationship itself is part of the treatment.
The study found that patient and clinician trust was strongly correlated with transparency: where AI tools were clearly explained, where their limitations were disclosed, and where human oversight was visible and meaningful, acceptance was substantially higher. The implication is that the social challenge is not fundamentally a communication problem or a resistance-to-change problem; it is a governance problem. Trust follows credible oversight.
Five Future Considerations for EU Action
1. Common Data Standards Across Member States
The single most impactful structural intervention the EU could make for AI in healthcare would be the effective implementation of common health data standards across member states. This means going beyond the policy commitment of the European Health Data Space to the practical work of ensuring that HL7-FHIR and other interoperability standards are actually implemented in clinical systems, that data sharing agreements reflect realistic operational requirements, and that the governance frameworks for cross-border data access work in practice rather than on paper.
2. Centres of Excellence for AI in Health
The study recommends the establishment of EU-level centres of excellence that can support member states with AI evaluation, implementation expertise, and knowledge sharing. Smaller member states in particular lack the institutional capacity to independently assess AI tools, negotiate with large technology vendors, or build the regulatory expertise required to navigate the compliance landscape. Centres of excellence could serve as shared infrastructure, reducing the duplication of effort that currently characterises European AI in health and accelerating the spread of what actually works.
3. Consolidated Funding Mechanisms
European funding for health AI is currently distributed across multiple programmes, Horizon Europe, the Digital Europe Programme, the European Regional Development Fund, and others, with limited coordination between them. This fragmentation creates inefficiencies, gaps in coverage, and incentives for duplicative research rather than implementation at scale. The study calls for a consolidated funding approach that explicitly supports the transition from research and pilots to sustained deployment, the phase at which European health AI consistently stalls.
4. Local Performance Testing Infrastructure
AI tools that perform well in their development environment may not perform equivalently in different clinical contexts. The study recommends investment in infrastructure that allows healthcare providers to test AI tools against their own patient populations before full deployment, including regulatory sandboxes and access to anonymised local data for validation purposes. This would reduce the risk of adverse outcomes from deploying unvalidated tools and build the evidence base that procurement decisions currently lack.
5. An EU AI Catalogue for Healthcare
A publicly accessible, regularly updated catalogue of AI tools authorised for healthcare use in the EU, with standardised information about their intended use, performance characteristics, regulatory status, and evidence base, would substantially improve the quality of procurement decisions across the healthcare system. The study found that decision-makers frequently lack the information needed to compare tools or to assess whether a specific AI product is appropriate for their context. A well-designed catalogue could address this directly.
The Demographic Imperative
The study's opening framing deserves emphasis. Europe faces a projected shortfall of 4.1 million healthcare workers by 2030. That figure is not a forecast of a possible future; it reflects demographic trends already in motion, the ageing of the existing healthcare workforce, the ageing of the patient population, and the mismatch between healthcare demand and supply that will intensify regardless of technology choices. AI cannot solve this problem on its own, but a Europe that fails to deploy AI effectively in healthcare will face that shortfall without the tools that could most significantly augment the capacity of its existing workforce.
That framing changes the character of the policy question. The barriers to AI deployment in European healthcare are real and documented in this study in granular detail. But the cost of failing to overcome them is also real. The study treats accelerating AI deployment not as a choice between caution and ambition but as a risk management question: the risks of poorly governed AI deployment must be weighed against the risks of insufficient deployment in the face of a structural workforce crisis. Getting the governance right is not an alternative to moving at pace; it is the precondition for doing so sustainably.