The EU AI Act in Healthcare: Deployer Obligations After the Digital Omnibus

Yes, and the answer has become more layered than it was. Regulation (EU) 2024/1689 entered into force on 1 August 2024 with a staged application timetable. Its general date of application was 2 August 2026. Days before that date, Regulation (EU) 2026/1744, the Digital Omnibus on AI, entered into force on 27 July 2026 and rewrote several of the deadlines.

What applies today, as of August 2026: the prohibited practices in Article 5, the AI literacy obligation in Article 4 as rewritten, the general purpose AI model obligations, the governance and penalty provisions, and the Article 50 transparency obligations. What has been pushed out: the substantive high-risk requirements and obligations in Chapter III Sections 1, 2 and 3, which now apply from 2 December 2027 for Annex III systems and 2 August 2028 for Annex I systems.

One detail is easy to miss and worth knowing. The Omnibus deferred Chapter III Sections 1, 2 and 3 with the express exception of Article 6(5), which is the Commission's duty to publish guidelines on the practical implementation of the high-risk classification rules together with a list of practical examples. That obligation was not deferred, which means the interpretive material should arrive well ahead of the compliance dates. If you are trying to classify a borderline system, watching for those guidelines is more productive than guessing.

The timetable below reflects Article 113 as amended by Regulation (EU) 2026/1744. Older summaries circulating online still show 2 August 2026 for Annex III high-risk obligations. That date has moved.

The grace period in Article 111(2) also received a useful clarification. It applies where the type and model of AI system has already been placed on the market, so if at least one unit was lawfully placed before the relevant date, other units of the same type and model can continue, without additional obligations, for as long as the design remains unchanged. Any significant change to the design triggers full compliance.

Some of it, by two quite different routes, and the distinction determines your deadline.

Route one, Article 6(1) and Annex I. An AI system is high-risk where it is intended to be used as a safety component of a product, or is itself a product, covered by the Union harmonisation legislation listed in Annex I, and where that product must undergo third-party conformity assessment. Annex I Section A includes the Medical Device Regulation (EU) 2017/745 at point 11 and the In Vitro Diagnostic Regulation (EU) 2017/746 at point 12. In practice this catches AI inside CE marked medical devices above the self-certification classes. Obligations apply from 2 August 2028.

Route two, Article 6(2) and Annex III. Certain standalone systems are high-risk by listing. The healthcare relevant entries sit in Annex III point 5: systems used by or on behalf of public authorities to evaluate eligibility for essential public assistance benefits and services including healthcare services, and to grant, reduce, revoke or reclaim them; systems used for risk assessment and pricing in life and health insurance; and systems that evaluate and classify emergency calls or dispatch or prioritise emergency first response, expressly including emergency healthcare patient triage systems. Annex III point 4 on employment also catches recruitment, promotion, termination, task allocation and performance monitoring tools, which is where many health systems first meet the regulation. Obligations apply from 2 December 2027.

Article 6(3) provides a derogation from Annex III classification where a system does not pose a significant risk of harm, and it applies where the system performs a narrow procedural task, improves the result of a previously completed human activity, detects decision-making patterns or deviations without replacing or influencing a previously completed human assessment without proper human review, or performs a preparatory task. A system that performs profiling of natural persons is always high-risk regardless. A provider relying on the derogation must document its assessment before placing the system on the market and is subject to the registration obligation in Article 49(2), so ask to see that documentation rather than accepting the conclusion.

Most healthcare organisations are deployers. Article 26 is therefore the operative provision, and it applies to high-risk systems from the deferred dates above. It requires:

• Use in accordance with the instructions for use, with appropriate technical and organisational measures to ensure that happens.
• Human oversight assigned to named natural persons who have the necessary competence, training and authority, as well as the necessary support. Authority and support are the words most often ignored. A clinician who can technically override an output but has no time and no institutional backing to do so is not oversight.
• Input data control where the deployer exercises control over it, ensuring it is relevant and sufficiently representative for the intended purpose.
• Monitoring in operation, informing the provider or distributor and the market surveillance authority and suspending use where the system appears to present a risk, and immediately reporting serious incidents.
• Retention of automatically generated logs under the deployer's control for a period appropriate to the intended purpose and at least six months.
• Informing workers and their representatives before putting a high-risk system into service in the workplace.
• Informing natural persons that they are subject to the use of a high-risk Annex III system where it makes or assists in making decisions about them.
• Using the provider's Article 13 information to support a data protection impact assessment under Article 35 of the GDPR.

Article 27 adds a fundamental rights impact assessment before first use, for deployers that are bodies governed by public law or private entities providing public services, and for deployers of Annex III points 5(b) and (c). A private hospital delivering publicly funded care should assume this can reach it and take advice rather than assume it cannot. Where a data protection impact assessment already covers an element, the fundamental rights assessment complements it rather than duplicating it.

They are in scope of the regulation, and generally not high-risk. That is a different thing from being unregulated, and the distinction trips people up.

An ambient documentation tool, a patient intake agent, an AI phone agent or scheduling automation is not listed in Annex III and is not a safety component of a regulated device, so Chapter III does not attach. Three obligations still do.

Article 4, AI literacy. Applicable since 2 February 2025 and rewritten by the Omnibus, it requires providers and deployers to take measures to support the development of AI literacy among staff and others operating AI systems on their behalf, proportionate to their knowledge, experience and context. The Omnibus text is explicit that this does not require guaranteeing any specific level of literacy in any individual, which makes it a training and documentation obligation rather than an assessment one. It is the easiest thing on this page to evidence and the most commonly forgotten. Our staff training programmes are built around exactly this evidence problem.

Article 50, transparency. Applicable from 2 August 2026. Systems interacting directly with people must make clear they are AI, and providers of systems generating synthetic audio, image, video or text must mark that output in a machine readable way. A patient facing phone or chat agent sits squarely in the first limb. Systems placed on the market before 2 August 2026 have until 2 December 2026 for the Article 50(2) marking obligation.

Article 5, prohibitions. Rarely relevant to administrative agents, but worth a single check against emotion recognition in the workplace and any inference about protected characteristics, both of which occasionally appear in workforce analytics without anyone having decided to buy them.

Four things, in order of how often they surprise people.

Extraterritorial reach is real but narrower than the headlines. Article 2 extends the regulation to providers and deployers established in a third country where the output produced by the AI system is used in the Union. A US health system treating a patient who happens to be an EU citizen, in the United States, is not producing output used in the Union. A US organisation running a telehealth service into a Member State, or operating an EU subsidiary, or supplying software used by EU clinicians, is a different case entirely. Draw the line on where the output is used, not on patient nationality.

The GDPR runs in parallel and usually bites first. For most US organisations touching EU patients, data protection is the immediate obligation and the AI Act is the later one. Article 26(9) explicitly points deployers back to their Article 35 GDPR impact assessment duty. Do not let an AI Act project displace a data protection project that is already overdue.

Medical device regulation is separate. A tool cleared by FDA is not thereby CE marked, and the EU conformity assessment route under the Medical Device Regulation is its own exercise. The AI Act sits on top of that, not instead of it, and the new Article 2(13) is the mechanism intended to stop the two regimes duplicating each other for Annex I systems.

The deferral is not a reprieve for the parts that already apply. AI literacy and Article 50 transparency are live obligations now. They are also the cheapest to satisfy, which makes leaving them undone a poor trade.

Enforcement is national. Each Member State designates market surveillance authorities, and the Commission's AI Office plays a central role for general purpose AI models. The Digital Omnibus strengthened that central role. Penalty ceilings sit at up to EUR 35 000 000 or 7 percent of total worldwide annual turnover for prohibited practices, and up to EUR 15 000 000 or 3 percent for most other infringements, whichever is higher in each case.

Two practical observations. First, the deferral of Chapter III means the obligations that could be enforced against a healthcare deployer through 2026 and most of 2027 are the transparency, literacy and prohibition provisions, not the high-risk ones. Second, the fines are ceilings applied through national procedures, and early enforcement across new EU digital regulation has generally started with information requests and corrective orders rather than headline penalties. Plan for the information request. It arrives asking which systems you operate, what they do and who oversees them, and an organisation that cannot answer that quickly has a problem independent of any fine.

That is the same answer we give about HIPAA enforcement in the United States, and for the same reason. The register is the deliverable.

Sequenced against the amended dates.

1. Now: determine whether you are in scope at all. Is output from any AI system you provide or deploy used in the Union. If the honest answer is no, document that conclusion once and revisit it when you open an EU service line. Do not spend a year on a regulation that does not reach you.
2. Now: build the inventory, classifying each system as prohibited, Annex I high-risk, Annex III high-risk, subject only to Article 50, or none of these. Record the reasoning, including any reliance on the Article 6(3) derogation and the provider's documented assessment behind it.
3. Now: satisfy Article 4. Role appropriate training for everyone operating an AI system, with a record of who received what and when. This obligation has been live since February 2025.
4. Now: satisfy Article 50. Any patient facing agent must disclose it is AI. Confirm marking of synthetic output with your vendor, and note the 2 December 2026 backstop for systems placed on the market before 2 August 2026.
5. Through 2027: build the Article 26 operating model for anything you have classified as Annex III high-risk. Named oversight owners with real authority, six month log retention, an incident route to the provider and the authority, and worker notification where the system is used in the workplace. Add an Article 27 fundamental rights impact assessment if you are a public body or a private entity providing public services.
6. By 2 December 2027 and 2 August 2028 respectively, have Annex III and Annex I systems operating under that model, and watch for the Commission delegated acts under Article 2(13), due by 2 August 2027, which will shape how the medical device overlap is resolved.

Steps two and five are where the work is, and both produce artefacts that are useful well beyond the EU. A single AI register with owners, classifications and oversight arrangements answers questions from a market surveillance authority, an OCR investigator and your own board with the same document. Building that register once, so it serves all three, is the work in our AI governance and compliance engagement.