Brussels, 17 November 2026. The Apply AI Summit closed the first European AI Innovation Month and marked one year of the Apply AI Strategy, hosted by Executive Vice-President Henna Virkkunen1. The healthcare session, « Scaling AI in Healthcare: From Prevention to Clinical Practice », was moderated by Merle Becker, with four clinicians and research coordinators from radiology and cardiology: Marie-Pierre Revel, Regina Beets-Tan, Luis Martí-Bonmatí and Thomas Lüscher2.
One question was put to all four speakers.
Why do so many AI tools that work technically still stay in pilots?
Four points came out of the discussion that followed. Each has a consequence for a company building and selling a regulated device into European health systems.
1. Imaging and detection: already in routine use
AI is in daily use in screening, imaging and cardiology. It reads faster than a human can, it catches findings that get missed at the end of a long reading session, and it absorbs volumes that screening programmes cannot staff. The clinicians using it described its limits themselves rather than discovering them afterwards.
For a manufacturer, this part is no longer a differentiator. Good performance figures get a product onto a shortlist, and the decision is taken on everything that comes after.
2. Prevention, and the question of access
The case made for prevention was economic as much as clinical. An AI that identifies risk in someone who does not yet know they are at risk brings that person into care earlier, and earlier contact costs less at every subsequent step. Cardiovascular prevention was the example, with AI placed at the centre of a European plan and risk scores published free for any doctor to use.
The limiting factor is access. A model that only operates once a patient is already in a specialist pathway intervenes after most of the cost has been committed.
So what: consider whether the product is designed for the point of first contact or for the point of diagnosis. The two have different buyers, different evidence requirements and different workflows to integrate into.
3. Scaling across Europe involves national specialisation
One question in the session was whether a tool trained on data from Valencia is valid for a patient in Stockholm. Populations differ, imaging protocols differ, scanner fleets differ, and so do clinical cultures and national screening programmes. The European networks presented are designed to connect centres that are advanced with centres that are starting, while leaving those differences in place.
For a manufacturer, each population and each protocol claimed has to be substantiated. Each claim widens the intended purpose and the technical file attached to it. In this market, scaling tends to mean a core product with qualified national variants rather than a single configuration deployed across twenty-seven health systems.
So what: decide early which variants are catalogued options covered by the technical file and which are specials. The second kind carries its own evidence and its own maintenance for a decade.
4. Data, anonymisation and sovereignty
Shared, high-quality European imaging data was described as the basis for trustworthy AI. The data largely cannot leave the hospital, which is why the infrastructures are federated rather than centralised. Anonymisation then has to be strong enough for security requirements and European data sovereignty, and selective enough that what remains is still clinically useful. The two requirements work against each other, and the balance is settled case by case.
There is also a timing question. Cancer Image Europe, the project building the federated cancer imaging platform, runs to 31 December 20263. Infrastructure built inside a funded project needs an owner once the funding ends, and a product that depends on it is exposed to that transition.
So what: if a shared infrastructure sits anywhere in your validation chain, document what happens to your evidence if its governance changes.
Responsibility stays with the clinician
The summit was framed around putting AI first, with the qualification that this does not put people second. In a clinical setting the qualification is concrete. The model produces an output, the clinician takes the decision, and the responsibility for that decision does not move. Nobody in the session argued otherwise.
For a deployment, that means one question has to be answered before anything is installed: who is accountable when the output is wrong, and does the person using it know the answer?
The same problem, from product development
I spent four years running the digital transformation of product development for a regulated medical device manufacturer, in oncology treatment equipment. None of it involved clinical AI. The bottleneck was the same.
Tools that performed well were not adopted. Evidence existed but could not be produced on request. Each market variant added a file to maintain for a decade. Four practices had more effect than any choice of tooling.
| Practice | What it looked like | Why it mattered |
|---|---|---|
| One named owner per artefact | Requirements, risks and verification on a single platform, a person accountable for each, not a team | An interface between two teams has no owner unless one is assigned |
| Traceability usable in minutes | Requirement to implementation to test, retrievable on request | It answers what has to be revalidated when one value changes, which is what an auditor or a payer asks |
| Adoption counted, not assumed | 80 systems engineers and compliance specialists moved deliberately, one at a time | The tooling took months. The adoption took two years |
| Process mapped to mandatory deliverables | Each activity tied to the regulatory output it produces, across 1,600 people | Evidence is produced by the work rather than reconstructed afterwards |
The programme ran at €5M against a 10% internal efficiency target, and most of the effort went into traceability and adoption rather than into the tools.
A European network can share data, pool expertise and connect a hospital that is starting out with one that is ahead. What it cannot do is decide, inside a company, who owns the production decision for a given tool. In most organisations that decision is unassigned: research owns the model, regulatory owns the file, commercial owns the customer, and the pilot stays in the space between them.
Four questions for a medtech executive this week
Which of your pilots has a named owner for the production decision, and does that person hold the budget for the evidence, the integration and the adoption?
Does your product reach the patient at first contact or only once they are already in a specialist pathway?
On how many populations, protocols and equipment fleets do you claim validity, and what substantiates each claim?
Who is accountable when the output is wrong, and does the clinician using it know the answer?
Notes
- European Commission, Apply AI Summit, 17 November 2026, The EGG Brussels and online, marking one year of the Apply AI Strategy and closing the first European AI Innovation Month, with Executive Vice-President Henna Virkkunen as special host: digital-strategy.ec.europa.eu. ↩
- Apply AI Summit agenda, 17 November 2026, session « Scaling AI in Healthcare: From Prevention to Clinical Practice ». Marie-Pierre Revel is professor of radiology at Université Paris Cité, head of radiology at Cochin, AP-HP, and first Vice-President of the European Society of Radiology. Regina Beets-Tan is former President of the European Society of Radiology, scientific director of the European Institute for Biomedical Imaging Research and coordinator of COMPASS-AI. Luis Martí-Bonmatí heads medical imaging at La Fe University Hospital in Valencia and is scientific coordinator of Cancer Image Europe. Thomas Lüscher is director of research, education and development at the Royal Brompton and Harefield NHS Foundation Trust and Imperial College London, and immediate past President of the European Society of Cardiology. Discussion points in this article are taken from the session. ↩
- EUCAIM, the project building Cancer Image Europe, runs from 1 January 2023 to 31 December 2026, with €17.79M of European funding and 76 partners: eibir.org. ↩
About the author. Bernard Lecuivre runs Coeos Transformation in Brussels. He led the digital transformation of product development for a regulated medical device manufacturer, a €5M programme against a 10% internal efficiency target, covering requirements, risk and verification artefacts for 80 systems engineers and compliance specialists, and a process map tied to mandatory regulatory deliverables across 1,600 people. Certified Prince2 Practitioner, Agile Project Management, Professional Scrum Product Owner, ITIL and OCEB.
For a review of a pilot that has to reach production: coeostransformation.eu | LinkedIn
