Eight years of AI for Health: the barriers that keep returning to the agenda

In November 2025, the AI for Health summit held its eighth edition inside Adopt AI, a large AI event organised by the consultancy Artefact at the Grand Palais in Paris. Sébastien Marguerès, who leads AI for Health and directs public affairs for Adopt AI, sat down with Tech Anatomy on the show floor. Asked which subjects the summit has still not resolved after eight years, he did not hesitate: "Access to data. Every year, I feel we talk about it."

The answer is worth taking seriously because of where it comes from. A summit organiser sees the same question reframed by pharmaceutical companies, regulators, hospital staff and start-ups year after year, which makes the programme a rough record of what has and has not moved.

Three barriers, in the order he gave them

The first is access. Marguerès described a familiar caricature: industry blames regulators for closed data, slow notified bodies and the added complexity that AI brings to the European Medical Device Regulation, while regulators work within frameworks built to protect patients. In his account, the two sides "are not playing on the same field", and the summit's job is to put them on one.

The second is quantity, which he said is discussed less often than quality. Cancer research can draw on large patient populations. Rare diseases cannot, and models trained on them need data pooled across countries. "If we don't go looking for data with European or international partners, we cannot train the model," he said (translated from French).

The third is infrastructure. "Some hospitals want to adopt AI, but some departments are not even digitised," he noted, before moving on.

What has been put in place since

Several of the measures he listed are now law or in operation, and their timelines explain part of the frustration. The European Health Data Space regulation was adopted in 2025 and entered into force that March, but its provisions on secondary use, the reuse of health data for research and innovation, apply only from March 2029. France's Health Data Hub was set up as a single entry point to national data, which Marguerès credited while adding that "we can feel it is still blocked". He described France's national claims database, the SNDS, as "a jewel", while noting that it covers care and reimbursement records rather than genomic data.

On genomics, he pointed to a structural difference. In France, Article 16-10 of the civil code limits genetic testing to medical or scientific purposes, which rules out consumer DNA kits. The United Arab Emirates, by contrast, runs a national genome programme to which more than 800,000 citizens had contributed by April 2025, with samples linked to clinical records. Marguerès presented it as a source of inspiration for France, without going into how it could be transposed.

He also gave one example of borrowing that has already happened: France's PECAN scheme, which offers twelve months of early reimbursement to digital medical devices, drew on Germany's DiGA model.

Learning from other industries

Asked whether health can learn from other sectors, Marguerès cited the partnership between Airbus and the IUCT-Oncopole cancer centre in Toulouse, which he said had presented at AI for Health a few years earlier. He was careful to say he no longer remembered the project's details. The public record is more specific than his recollection: in September 2024, Airbus, the radiotherapy equipment maker Accuray and the Oncopole signed an agreement to use AI to predict performance problems in radiotherapy machines, so that breakdowns do not delay patients' treatment. It is predictive maintenance, a technique Airbus uses on aircraft fleets, applied to hospital equipment. His broader suggestion, that the same tools might one day predict cancer relapse, was framed as a possibility, not as something the project does.

Two readings of a recurring agenda

One reading is that the agenda repeats because the problem is structural. Health data in Europe is held by many institutions under strong protection rules, and the summit format, which places a start-up, a regulator and an academic on the same panel, exists because each speaks a different professional language. Marguerès described preparing a round table where a technologist, a regulator and a physician "were talking about exactly the same thing, with different words". On this reading, slow progress is the cost of building trust in health data, which he also called necessary.

A second reading is that the tools now exist and the delay is in implementation. The EHDS, the Health Data Hub, PECAN and the EU's TEF-Health testing facilities are all in place or scheduled. If the same complaints persist once secondary use applies in 2029, the explanation would have to lie elsewhere: in hospital infrastructure, in data quality or in the business models of the companies asking for access.

What the interview does not establish

This was a short conversation recorded at the event, and Marguerès spoke as its organiser, with an interest in the convening role he described. His account of what recurs is an impression from the programme, not a count of sessions. Some figures he mentioned, such as a breakfast for ambassadors with more than thirty countries represented, could not be checked against a published source and are left aside here. The question of whether AI in health has actually "stalled", as the episode title puts it, would need adoption data that the interview did not cover.

What to watch

Marguerès said the next edition, scheduled for 3 and 4 December 2026 at the Grand Palais, would focus on European cooperation, on data access through specific use cases, and on patients, including patients on stage rather than only discussed. Two signals would show whether the agenda is moving: whether data access still opens the programme once the EHDS secondary-use rules approach, and whether the patients invited are given the floor on how their data is used.

Listen to the episode

Pourquoi l'IA en Santé Stagne Depuis 8 Ans avec Sébastien Marguerès, AI for Health

Listen · 16 min