Under the EU AI Act, most AI medical devices count as “high-risk” because they already need a notified body under the MDR or IVDR. The law currently applies those obligations to them from 2 August 2027. Both EU co-legislators back delaying this to August 2028, but as of 1 May 2026 no deal has been agreed.
Consider a European start-up with a CE-marked AI tool that reads chest X-rays. It is a Class IIa device under the Medical Device Regulation, with a notified body, a quality system and a clinical evaluation. Its founders could be forgiven for assuming the EU AI Act changes little for them.
They would be half right. The AI Act does not replace the MDR. It sits on top of it. For AI medical devices, the practical result is a product that is high-risk twice over, under two regulations written by different people with different words for overlapping ideas.
My view is that the double regime matters less for what it adds on paper than for what it moves. It shifts attention from the device at the moment of certification to the data behind it and the people using it afterward. It also puts real legal duties on hospitals, which under the MDR were mostly customers.
This article is general information, not legal advice. It reflects the law and the legislative state of play as of 1 May 2026, which is moving fast.
When does the EU AI Act apply to healthcare?
The AI Act, Regulation (EU) 2024/1689, entered into force on 1 August 2024. Its obligations arrive in stages:
- 2 February 2025: bans on prohibited AI practices and the AI literacy duty for providers and deployers.
- 2 August 2025: obligations for general-purpose AI models, and the governance and penalties framework.
- 2 August 2026: most remaining rules, including obligations for the “stand-alone” high-risk systems listed in Annex III and transparency duties such as telling people they are dealing with a chatbot.
- 2 August 2027: obligations for high-risk AI that is a product, or a safety component of a product, covered by the EU laws listed in Annex I. The MDR and IVDR are on that list.
That last date is the one most medical AI makers care about, and it is now in play. In November 2025 the European Commission proposed a “Digital Omnibus on AI” that would delay high-risk obligations and link them to the availability of harmonised standards. The Council adopted its position on 13 March 2026. Parliament adopted its position in plenary on 26 March. Both backed fixed dates: 2 December 2027 for Annex III systems and 2 August 2028 for AI embedded in Annex I products such as medical devices.
The delay is not law yet. On 28 April 2026, a trilogue between the Parliament, Council and Commission ended without agreement after about twelve hours. According to accounts from law firms following the talks, the sticking point was not the delay itself. It was the architecture for Annex I products, including whether sectors such as machinery and medical devices should stay under combined AI Act and sectoral assessment or move to a lighter track. Until an amending regulation is published, the dates above are the dates.
Why is medical AI “high-risk” under the AI Act?
Article 6(1) makes an AI system high-risk if two conditions are met. First, it is a product, or a safety component of a product, covered by the EU legislation in Annex I. Second, that product must undergo third-party conformity assessment under that legislation.
For medical devices, this maps neatly onto the MDR and IVDR. The joint guidance from the Medical Device Coordination Group and the AI Board, published in June 2025 as MDCG 2025-6 / AIB 2025-1, confirms the logic. If medical device AI needs a notified body under the MDR or IVDR, it is high-risk under the AI Act. The guidance also makes clear that being high-risk under the AI Act does not change the device’s MDR risk class.
Because MDR Rule 11 puts most diagnostic and therapeutic software into Class IIa or higher, most medical AI needs a notified body. That makes it high-risk. Class I software that is self-certified generally is not caught by Article 6(1), though it could still be caught by other parts of the Act.
Healthcare can also appear in Annex III, the list of stand-alone high-risk uses. It includes AI used to evaluate and classify emergency calls or triage patients in emergency healthcare, AI used by public authorities to assess eligibility for essential public services including healthcare, and AI used for risk assessment and pricing in life and health insurance. These can be high-risk even if they are not medical devices at all.
What does the AI Act add on top of the MDR?
Many high-risk requirements, from risk management to post-market monitoring, look familiar to device makers.
The additions sit in places the MDR treats more lightly:
- Data governance (Article 10): training, validation and test data must be relevant, sufficiently representative and, as far as possible, free of errors and complete for the intended purpose, with examination for possible biases.
- Record-keeping (Article 12): the system must technically allow automatic logging of events over its lifetime.
- Transparency to deployers (Article 13): instructions must let users interpret outputs and use the system appropriately, including known limits on accuracy.
- Human oversight (Article 14): the system must be designed so that people can oversee it, understand its limits, avoid automation bias and override or stop it.
Under the MDR a manufacturer could treat data curation as an internal engineering matter. Under the AI Act it becomes an auditable legal obligation. The questions our earlier piece raised about bias in clinical algorithms stop being ethics questions and become compliance questions.
What must hospitals do under the EU AI Act?
Under the MDR, a hospital that buys a CE-marked device is mostly a user. Under the AI Act, a hospital that uses a high-risk AI system in its work is a deployer, with duties of its own under Article 26. Deployers must:
- use the system in line with the provider’s instructions;
- assign human oversight to people who have the competence, training and authority to exercise it;
- make sure input data under their control is relevant and sufficiently representative for the intended purpose;
- monitor operation and inform the provider, and in serious cases the authorities, about risks or serious incidents;
- keep automatically generated logs under their control for at least six months, unless other law says otherwise;
- inform workers and their representatives before putting a high-risk system into use at work.
Some public-sector deployers of Annex III systems must also carry out a fundamental rights impact assessment before first use. And since February 2025, every provider and deployer has had to take measures to ensure a sufficient level of AI literacy among staff who operate AI systems. The Omnibus positions would soften that duty. For now it stands.
A hospital can also become a provider. If it builds its own AI tool, substantially modifies a purchased one or puts its name on it, it takes on provider obligations.
A worked example: one sepsis tool, two regimes
Here is a hypothetical. A German university hospital buys a CE-marked AI tool that predicts sepsis in ward patients. The tool is a Class IIa device, so it is high-risk under the AI Act.
The vendor’s duties are the heavier set. Its technical file already covers MDR requirements. For the AI Act it must add documented data governance, showing that its training data, drawn mainly from hospitals in two countries, was examined for bias and is representative of its intended population. It must build logging into the product, write instructions that explain known performance limits and design the alert so a clinician can override it.
The hospital’s duties are lighter but real. It must name who oversees the tool and train them. Suppose the ward population is older than the population in the vendor’s data. The hospital controls the inputs, so it should check that the data it feeds the model fits the intended purpose. If the tool starts firing on patients who plainly are not septic, the hospital must tell the vendor. It must keep the logs for at least six months.
Now suppose the hospital’s data science team retrains the model on local data to reduce false alarms. That may count as a substantial modification. If so, the hospital becomes a provider, with the full set of provider duties, and the MDR questions about who is the manufacturer come back too. The most useful thing the hospital can do is decide, before anyone touches the model, whether it wants that role.
Is double regulation of medical AI a mistake?
The strongest case against the regime is cost and delay. Medical device makers in Europe spent the years since 2021 struggling through the MDR transition. Notified bodies were scarce, certificates expired and the EU extended transition deadlines to avoid device shortages. Adding a second layer before the first has settled risks a repeat. Small companies feel this most, and some may simply launch in the US first. Duplication is a fair worry too: harmonised standards for the AI Act were still unfinished as of this writing, which is exactly why both co-legislators support a delay.
I take this seriously, and I think a delay is sensible. But the case against the substance is weaker than the case against the timing. The MDR was written for devices, not for statistical models trained on patient data. It says little about representativeness, subgroup performance or how a model should behave when it meets people unlike those it was trained on. The AI Act’s data and oversight rules fill a genuine hole. If the price is a year or two more for standards and notified bodies to catch up, that is a reasonable trade. Dropping the substance would not be.
What this means for manufacturers, hospitals and policymakers
For manufacturers, plan for August 2027 while hoping for August 2028. Integrate AI Act requirements into your MDR quality system now rather than running a parallel project. Treat your training data documentation as if an auditor will read it, because one will. Our piece on why the device regulator you choose is really a business model applies doubly in Europe now.
For hospitals, build a register of AI systems in use and mark which are high-risk. Assign oversight roles by name. Check your log retention. Brief staff representatives. Decide who can modify a model, because modifying one can make you a provider.
For policymakers, the lesson from the MDR transition is capacity. Deadlines without enough notified bodies and finished standards produce extensions, not compliance. The same logic applies to data: the European Health Data Space Regulation, in force since March 2025, is meant to open up health data for secondary use, and it will shape what representative training data is even possible. We covered why ownership framing misleads in who owns patient data is the wrong question.
Frequently asked questions
Is all medical AI high-risk under the EU AI Act?
No. AI that is a medical device, or a safety component of one, is high-risk when it requires notified body assessment under the MDR or IVDR, which covers most diagnostic and therapeutic software. Class I self-certified software usually is not high-risk on that basis. Some non-device health uses, such as emergency triage and health insurance pricing, are listed separately in Annex III.
Does the AI Act replace the MDR for AI medical devices?
No. Both apply. The MDR still governs safety, performance, clinical evaluation and classification. The AI Act adds requirements on data governance, logging, transparency and human oversight. EU guidance from June 2025 encourages manufacturers to meet both through one quality system and one set of technical documentation, assessed by a notified body designated under both laws.
Has the EU delayed the AI Act for medical devices?
Not yet, as of 1 May 2026. The Commission proposed a delay in November 2025, and both the Council and Parliament adopted positions in March 2026 backing 2 August 2028 for AI in regulated products such as medical devices. A trilogue on 28 April 2026 ended without agreement. Until an amendment is adopted and published, the legal date remains 2 August 2027.
What are the penalties for breaking the EU AI Act?
Fines are tiered. Using prohibited AI practices can cost up to 35 million euros or 7 percent of worldwide annual turnover, whichever is higher. Breaching most other obligations, including those for high-risk systems, can cost up to 15 million euros or 3 percent. Smaller companies face the lower of the two amounts.
Do patient-facing chatbots in healthcare have to disclose they are AI?
Under Article 50, from 2 August 2026, providers must design AI systems that interact directly with people so users are told they are dealing with AI, unless it is obvious from context. This applies to symptom checkers and booking assistants whether or not they are high-risk.
The takeaway
The EU AI Act does not make medical AI safe by itself, and its timeline may yet slip by a year. What it does is move legal responsibility to places the MDR left soft: the data a model learned from, and the people who oversee it once it is switched on. Manufacturers will feel the paperwork. Hospitals will feel something newer, the discovery that using AI is now a regulated activity in its own right.
Is your organisation treating the AI Act as a vendor problem or its own? We’d like to hear how hospitals are preparing.
