The most common confusion: it's not Annex III, it's Annex I
It's common to assume that any AI system used in a healthcare setting falls under Annex III of the AI Act, the section that lists sector-specific high-risk use cases (employment, education, law enforcement, welfare benefits, and so on). But that's not the route that applies to an assisted-diagnosis, triage, or clinical monitoring system that's part of a medical device.
The correct route is Article 6(1) together with Annex I. An AI system is classified as high-risk through this route when two conditions are both met: (a) the AI system is, or is a safety component of, a product covered by the EU harmonisation legislation listed in Annex I, and (b) that product is required to undergo third-party conformity assessment before it can be placed on the market.
Annex I, Section A, expressly includes point 11 — Regulation (EU) 2017/745 on medical devices, known as the MDR — and point 12 — Regulation (EU) 2017/746 on in vitro diagnostic medical devices, known as the IVDR. If your AI software is a safety component of a medical device or in vitro diagnostic product subject to MDR or IVDR, this is the route that applies to you, not Annex III.
Annex III, point 5(a), covers something different: AI systems used to assess a natural person's eligibility for essential public benefits and services — including public healthcare — meaning deciding who has the right to access a service or benefit. It does not cover the clinical act of diagnosing, triaging, or monitoring a patient with a medical device. Confusing the two routes means assessing your system against the wrong framework, and potentially missing obligations — like those under MDR/IVDR — that do apply.
One conformity assessment, not two
The most important practical consequence of falling under Annex I rather than Annex III shows up in Article 43(3): the manufacturer does not go through two separate conformity assessment procedures. Instead, it follows the conformity assessment procedure already required by MDR or IVDR, integrating into that same process the AI Act requirements set out in Chapter III, Section 2 (technical documentation, risk management, data quality, human oversight, registration and traceability, among others).
This also affects who evaluates the product. The same notified body already designated under MDR or IVDR can also assess the AI Act requirements, provided it demonstrates the additional competence needed. Notified bodies have until January 28, 2028 to apply for that extended designation with their national authority.
No additional third-party assessment is required simply because a product incorporates an AI component: if MDR or IVDR did not require notified-body involvement for your product class, the AI Act doesn't add one on its own.
When it's high-risk in practice
In practical terms, your system falls under this high-risk route when the product — the medical device or in vitro diagnostic product your AI is a safety component of — is subject to third-party conformity assessment under MDR or IVDR. That typically includes MDR Class IIa, IIb and III devices — clinical decision-support software, AI-assisted image diagnosis, vital-sign monitoring — and most IVDR Class B, C and D products.
One important difference from Annex III systems: there is no equivalent here of the "narrow procedural task" exception in Article 6(3), which lets some Annex III systems fall outside high-risk classification when their role is merely ancillary. Under the Article 6(1)/Annex I route, if both conditions are met — a component of an Annex I product, and that product subject to third-party assessment — the high-risk classification is automatic, with no room for that exception.
What obligations apply
Once classified as high-risk through this route, the relevant AI Act obligations fold into the technical file already required by MDR/IVDR, as covered above. The main ones are:
| Obligation | What it requires |
|---|---|
| Technical documentation | Integrated into the technical file already required by MDR/IVDR, rather than a separate dossier just for the AI component. |
| Human oversight (Art. 14) | Clinical staff using the system must be able to correctly interpret its output, detect automation bias — trusting the AI's result without questioning it — and have the genuine ability to decide not to use it or to interrupt the system. |
| Data governance and bias examination (Art. 10) | Training, validation and testing datasets must be examined for possible biases that could lead to discrimination or affect people's health and safety. |
The Article 10 bias examination deserves a paragraph of its own in a healthcare context, because here the risk isn't abstract. There is peer-reviewed scientific literature documenting disparities in clinical AI outcomes across demographic and racial groups — in both diagnostic accuracy and resource allocation or prioritization of care. A system trained predominantly on data from one population subgroup can perform worse, or in a systematically skewed way, for patients outside that subgroup. That's why subgroup bias testing isn't a generic compliance exercise in this context — it's also, directly, a matter of clinical safety.
The timeline: later than the rest, not earlier
This is the point that surprises people most familiar with the AI Act's general timeline. Article 6(1)/Annex I systems — including medical devices — have a date of application of August 2, 2028. That's later, not earlier, than the date of application for Annex III high-risk systems, which is December 2, 2027, following the delay introduced by the 2026 Digital Omnibus.
The reason isn't that the healthcare sector is a lower regulatory priority — it's the opposite. Lawmakers recognized that these products were already regulated under MDR and IVDR before the AI Act existed, and wanted to give notified bodies enough time to obtain the extended designation that lets them also assess AI requirements — hence the application deadline of January 2028 mentioned above. Applying the obligation before sufficient assessment capacity existed would have been counterproductive.
Worth keeping straight so you don't plan around the wrong date: if your product is a medical device with AI, your reference date is August 2028, not December 2027.
Practical steps
Step 1 — Confirm whether your product is already MDR/IVDR Class IIa or above
If your AI-enabled device already has, or needs, a CE mark under MDR or IVDR in a class requiring notified-body assessment, you start from the assumption that the correct route is Article 6(1) and Annex I, not Annex III.
Step 2 — Contact your notified body about the extended designation
Ask whether your notified body under MDR/IVDR is applying for — or has already applied for — the extended designation to also assess AI Act requirements, ahead of the January 2028 deadline.
Step 3 — Prepare bias testing by demographic subgroup
Design the evaluation of your training and validation datasets broken down by demographic or racial subgroup, before it becomes something an audit or inspection asks for.
Step 4 — Document clinical human oversight
Put in writing how clinical staff interpret the system's output, how the risk of automation bias is mitigated, and under what conditions they can choose not to follow the AI's recommendation or interrupt the system.
Frequently asked questions
Our assisted-diagnosis software already has a Class IIa CE mark — do we need two separate assessments (MDR and AI Act)?
No. Article 43(3) integrates the AI Act's requirements into the same conformity assessment procedure already required under MDR/IVDR, with the same notified body if it obtains the extended designation.
Does the AI Act classify our system under the public-health Annex III?
Not necessarily. If it's a safety component of a medical device subject to third-party conformity assessment under MDR/IVDR, the correct route is Article 6(1) plus Annex I, not Annex III, which covers eligibility for essential public benefits, not clinical diagnosis or triage itself.
Since when do we need to comply if our device is already on the market?
The date of application for Article 6(1)/Annex I systems is August 2, 2028, following the delay introduced by the 2026 Digital Omnibus — it was previously August 2027. It's worth confirming this date against an official source before locking in internal deadlines.
Is there any "grace period" for products already regulated under MDR?
Article 43(3) avoids duplicated procedures, and the Digital Omnibus introduces a mechanism for the Commission to limit AI Act requirements where MDR/IVDR already impose equivalent obligations — but this doesn't exempt you from compliance, it only avoids duplicating paperwork.