Study on the deployment of AI in healthcare — European Commission (PwC and Open Evidence), 2025
Despite the promise and availability AI-based tools in the market, their deployment in clinical practice is slow.
— the report’s abstract (p. 2)
TL;DR
A 241-page study for the Commission’s health directorate on how far AI is deployed in EU health care, what blocks it, and what the EU could do. It combines a literature review, a market analysis of AI medical devices, five stakeholder surveys, interviews, workshops and four case studies.
- The study warns against its own kind of evidence. “Surveys may lead to overly optimistic estimations of AI deployment in healthcare since participants are usually more familiar with these technologies than the average healthcare professional” (p. 41).
- Claimed use and evidenced use differ. Of EU-based clinicians answering, 63% (29 of 46) said they use AI tools in clinical practice; only 20 of the 46 (43%) gave evidence of actual use of AI medical devices. Some named ChatGPT (p. 42).
- The hospital survey is small. 35 hospital representatives answered: 20 pilot an AI solution, 19 have bought and deployed a commercial one, 11 run an in-house one, 2 have adopted none (p. 42). Among adopters, “only a minority track specific performance or efficiency metrics” (p. 156).
- Deployment lags research and approvals. Despite rising R&D and device approvals, “there is a large disconnect between the amount of research and development on AI medical devices and their adoption in clinical practice” (p. 40). The study’s scenarios for deployment in clinical practice range from 5% to 48%, with 27% in between (p. 46).
- The 94% in the report is about radiology’s potential. 94% of 32 hospital representatives name radiology as the area with the most transformative potential (Annex 4, p. 153). No figure in the report says 94% of providers use or plan AI.
Key data
| Measure | Figure | Base | Page |
|---|---|---|---|
| Survey invitations / responses, all groups | 1,224 / 238 | five stakeholder surveys | 10 |
| … hospital representatives | 307 invited / 35 responded | CIOs, CEOs, AI officers, hospital associations | 9–10 |
| … healthcare professionals (HCPs) | 242 / 83 | 51 answered the AI-use question | 10, 42 |
| Countries with most responses | Netherlands 27, Latvia 21, Spain 15 | all groups | 10 |
| HCPs claiming AI use, EU-based | 63% (29 of 46) | HCPs with “good knowledge of AI usage” | 42 |
| HCPs giving evidence of actual AI medical-device use | 43% (20 of 46) | same 46 | 42 |
| Hospitals piloting / deployed commercial / deployed in-house / none | 20 / 19 / 11 / 2 | 35 hospital representatives | 42, 156 |
| Health-sector firms claiming at least one AI tool (EC enterprise survey 2020) | 47% | literature | 41 |
| CE-marked AI radiology products (Radiology Health AI Register, Oct 2024) | 214 | market analysis | 43 |
| New AI radiology products CE-marked, Jan 2021 – Jun 2024 | 50 (78% diagnostic) | market analysis | 44–45 |
| Deployment scenarios for clinical practice: baseline / average / best case | 5% / 27% / 48% | authors’ assumptions | 46 |
| Unclear liability a significant challenge | 40% of hospital reps (10 of 25); 43% of HCPs (20 of 47) | survey | 68 |
| Low digital health literacy a significant challenge | 58% of hospital reps; 43% of HCPs | survey | 89 |
| Radiology as the area with the most transformative potential | 94% | 32 hospital representatives | 153 |
| Admin support tools (e.g. EHR management) as most transformative | over 87% | 32 hospital representatives | 153 |
Method and definitions
- Who. Written by PwC EU Services EEIG and Open Evidence for DG SANTE, Unit C.1 Digital Health; study period January 2024 – May 2025; manuscript completed June 2025. The cover says the views are the authors’, not the Commission’s.
- Surveys. Five targeted surveys on EUSurvey (patients, HCPs, hospital representatives, AI developers, regulatory experts), launched 10 June 2024 with translations from 3 July 2024 (p. 10). The closing date is printed as “25th September 2025”, which must mean 2024. Stakeholders were picked for “exposure to deployment of AI” (p. 16).
- “Deployment” means AI/ML-enabled medical devices in clinical practice. The study measures it three ways: literature estimates (job postings, surveys), registers of approved and CE-marked devices, and its own surveys.
- Scenarios (p. 46) are assumptions, not measurements: the 5% baseline follows Goldfarb et al.’s US job-posting estimate; the 48% best case follows the share of radiologists in a European Society of Radiology survey who claimed to use AI.
Scope and reliability
- No representative hospital figure. The hospital survey has 35 respondents, chosen for AI exposure; 32 answered the transformative-potential question. The authors call their own survey results potentially “biased estimations” (p. 42) and “positively biased” (p. 42).
- Self-report is the problem the study names. Respondents may count rule-based decision support or general tools such as ChatGPT as AI (pp. 41–42, 46).
- A consultancy study for a regulator. It reports stakeholder views on the AI Act, MDR, EHDS and the Product Liability Directive without a full regulatory assessment (p. 16).
- Partly read. The four case studies and the full monitoring framework were not read closely.
Dynamic-capabilities reading
digital-sensing/digital-scouting: an EU-wide scan of where AI in health care stands, from research and device approvals through to use in clinical practice.contextual/internal-barriers: the organisational and social challenges it lists (lack of end-user involvement, no added-value assessment, no AI strategy, lack of trust, job-security fears) are barriers inside hospitals, alongside the external regulatory ones.
How it connects
- enterprise-ai-adoption — the clearest case in the corpus of a source discounting its own survey: claimed use 63%, evidenced use 43%.
- ai-maturity-measurement-comparison — why survey-based adoption and maturity figures for health care read high.
- responsible-ai — liability, transparency and post-deployment monitoring as conditions for deployment.
- Adoption Monitor — shows survey adoption figures diverging by population and question; this study gives a health-care example of a survey figure that falls when evidence is asked for.
- AI Index 2026 — its medicine chapter finds 5% of 500+ clinical AI studies use real clinical data; this study finds deployment trailing research and approvals.
- In this batch: Europe readiness survey measures the same region at country level, where legal uncertainty and liability also lead the barriers.
- ai-in-health-care: the concept page that sets the health-care sources in this wiki side by side.
What was actually ingested
The official PDF from the Publications Office (identical in text to the mirror at medicaldevices.ipn.pt cited in the trigger report). Read: the abstract, methodology, limitations, deployment chapter, conclusions and the Annex 4 synopsis of the surveys, plus keyword searches across the whole text for every figure on this page. The challenge chapters (pp. 47–99), future considerations and the case-study annexes were skimmed, not read.
Linked entities and concepts
- PwC — co-author (PwC EU Services EEIG).
- enterprise-ai-adoption, responsible-ai — concepts. ai-maturity-measurement-comparison — thread.
Dangling (single-source mention, deferred): European Commission / DG SANTE (commissioning body; no entity page), Open Evidence (co-author), Yiannos Tolias (DG SANTE contact; also a peer reviewer of the WHO/Europe report in this batch), Radiology Health AI Register (Radboud University Medical Center).