Quick answer: If your AI system is a safety component of, or is itself, a device that already needs notified body sign-off under MDR 2017/745 or IVDR 2017/746, it's automatically high-risk under Article 6(1) of the EU AI Act. You have until August 2, 2027 (Article 113(b)) to be in conformity — a longer runway than the August 2, 2026 date that applies to most other high-risk AI. Article 43(3) lets you fold the AI Act assessment into the MDR/IVDR conformity assessment you already run, through the same notified body, if that body is designated to cover both.
That's the shape of it. The rest of this guide fills in the mechanics: what triggers the classification, what obligations follow, where your existing technical file already covers the new requirement and where it doesn't, and what to build first.
The AI Act (Regulation (EU) 2024/1689) doesn't create a parallel compliance universe next to your MDR or IVDR file. It bolts onto it. In some places it changes what your notified body has to review before you can put a CE mark on a device that uses machine learning.
Why Medical Device AI Gets Special Treatment
The AI Act sorts AI systems into four risk tiers: unacceptable (banned), high-risk, limited risk, and minimal risk. Most consumer AI never leaves the bottom two tiers. Medical devices work differently, and the reason is written directly into Article 6(1).
Article 6(1) sets a two-part test. First, the AI system is either a safety component of a product, or it is the product itself. Second, that product falls under the EU harmonization legislation listed in Annex I, and already requires third-party conformity assessment before it can go on the market. Medical devices satisfy both parts automatically: MDR 2017/745 and IVDR 2017/746 are both listed in Annex I.
So if your AI system is a safety component of a device, or is the device, and that device already needs notified body sign-off under MDR or IVDR, it lands in the AI Act's high-risk category by default. There's no separate risk-scoring exercise to argue your way out of it. The classification is structural, not judgment-based.
That's a different logic than the AI Act uses almost everywhere else. For most sectors, Annex III lists specific high-risk use cases — biometric ID, employment screening, credit scoring — and you check whether your system matches one. Medical devices skip that list entirely and inherit high-risk status from the underlying product regulation instead.
What Already Triggers Conformity Assessment Under MDR/IVDR
If you've been through an MDR submission, you already know Rule 11 of Annex VIII MDR governs software medical devices. It pushes almost anything that informs a diagnostic or therapeutic decision into Class IIa or higher, which means notified body involvement. IVDR has a comparable structure under Annex VIII, and most software that analyzes patient specimens or data for diagnostic purposes lands in Class B or above.
Practically, that means most AI-driven clinical decision support tools, diagnostic imaging algorithms, and IVD analysis software were already required to go through a notified body before the AI Act existed. What changes now is that the same product also has to satisfy the AI Act's high-risk requirements, layered on top of the MDR or IVDR technical file you already maintain.
The High-Risk Obligations You're Actually Taking On
Once a device is high-risk under Article 6(1), Chapter III of the AI Act applies. Articles 9 through 15 spell out six core obligations:
- Risk management system (Article 9) — a continuous process across the AI system's lifecycle, not a one-time assessment
- Data and data governance (Article 10) — training, validation, and testing data must be relevant, representative, and examined for bias
- Technical documentation (Article 11) — detailed enough for authorities to assess conformity, per Annex IV
- Record-keeping (Article 12) — automatic logging capabilities built into the system
- Transparency and instructions for use (Article 13) — deployers need to understand the system's capabilities and limitations
- Human oversight (Article 14) — measures that let a human operator understand, monitor, and intervene in the system's output
- Accuracy, robustness, and cybersecurity (Article 15) — appropriate levels maintained throughout the lifecycle
If you run an ISO 13485 quality management system, none of this is unfamiliar as a concept. Risk management, documentation, and post-market surveillance already carry weight in your MDR file. The AI Act asks you to extend those same disciplines to cover model behavior, training data provenance, and human oversight of algorithmic output specifically. It's additive work on a structure you already have, not a different philosophy.
Where MDR/IVDR and the AI Act Actually Meet
Article 43(3) of the AI Act keeps this from becoming two separate certification tracks running side by side. It allows your MDR or IVDR conformity assessment to also cover the AI Act's Chapter III requirements, carried out by the same notified body, as long as that body is designated to assess AI Act requirements too. Compliance with Chapter III can be demonstrated through your existing MDR or IVDR conformity assessment procedure, rather than through a duplicate standalone process.
That's the useful part. The less convenient part: not every notified body currently designated under MDR/IVDR has built out the technical competence to assess AI Act requirements yet. Confirm your notified body's scope of designation covers the AI Act before you assume an integrated pathway is available to you.
| Requirement | MDR 2017/745 / IVDR 2017/746 | EU AI Act (Regulation 2024/1689) |
|---|---|---|
| Basis for high-risk status | Risk classification rules (Annex VIII) | Article 6(1): safety component of, or itself, an Annex I product requiring third-party assessment |
| Core process document | Technical file / clinical evaluation | Technical documentation per Annex IV |
| Ongoing obligation | Post-market surveillance (Article 83 MDR) | Post-market monitoring (Article 72 AI Act) |
| Human oversight requirement | Instructions for use, labeling | Article 14: dedicated human oversight measures |
| Conformity assessment body | Notified body | Same notified body, if AI Act-designated (Article 43(3)) |
| Applies from | In force since May 2021 (MDR) / May 2022 (IVDR) | August 2, 2027 for Article 6(1) product cases |
Mapping Your MDR/IVDR File Against Annex IV
This is the part most guides skip: which section of your existing technical file actually satisfies which Annex IV requirement, and which requirement has no home in MDR at all. Here's the crosswalk.
| MDR/IVDR Technical File Section | AI Act Annex IV Item | Status |
|---|---|---|
| Device description & specification (Annex II MDR, §1) | Annex IV(1): general description of the AI system | Overlaps fully — reuse as-is |
| Design & manufacturing information (Annex II MDR, §3) | Annex IV(2)(a)–(c): development methods, design specs, architecture | Overlaps mostly — extend with algorithm-level detail |
| GSPR checklist (Annex I MDR) | Annex IV(7): standards applied to meet requirements | Overlaps mostly — reuse checklist mapping |
| Risk management file, ISO 14971 (Annex II MDR, §5) | Annex IV(5): risk management system per Article 9 | Overlaps mostly — extend to cover model drift and bias as risks, not just device failure modes |
| Verification, validation, clinical evaluation (Annex II MDR, §6) | Annex IV(2)(g) and (4): validation procedures, performance metrics | Overlaps mostly — add accuracy/robustness metrics required under Article 15 |
| PMS plan / PSUR (Annex III MDR) | Annex IV(9): post-market monitoring system | Overlaps structurally — integrate under Article 72(3) |
| No direct equivalent | Annex IV(2)(d): training/validation/testing data provenance, labeling, curation | Gap — build new under Article 10 |
| No direct equivalent | Annex IV(3): human oversight design, monitoring, and control specifications | Gap — build new under Article 14; IFU labeling alone doesn't satisfy this |
Two rows carry no MDR equivalent at all: the data governance file and the human oversight design file. Everything else in Annex IV is an extension of documentation you already produce, not a blank page.
The Timeline That Actually Matters
The AI Act entered into force on August 1, 2024, but it doesn't apply all at once — Article 113 staggers application by chapter. Prohibited AI practices under Chapter II became applicable February 2, 2025. Obligations for general-purpose AI models, governance structures, and penalty provisions became applicable August 2, 2025. The general rule for the rest of the regulation is 24 months after entry into force: August 2, 2026.
Medical devices get more time. Article 113(b) extends the deadline for high-risk systems captured under Article 6(1) — the product-safety-component category that includes MDR and IVDR devices — to 36 months after entry into force. That's August 2, 2027. If your device is high-risk purely because it's a product under Annex I legislation, that's your deadline, not the August 2026 date that applies to Annex III categories.
Treat August 2027 as a date to be finished by, not a date to start planning around. Notified body capacity for AI Act assessments will be a real constraint as it approaches, and folding AI Act documentation into an MDR technical file you're already maintaining is far easier done incrementally than backfilled in one push. For how these dates interact with AI Act obligations outside the medical device space, see our EU AI Act compliance timeline guide.
Practical Steps for Medical Device Manufacturers
Confirm your device is actually in scope. Not every piece of software touching a medical device is high-risk under the AI Act. The trigger is specifically that the AI system is a safety component of, or is itself, a product requiring third-party conformity assessment under MDR or IVDR. A Class I self-certified device without AI as a safety component generally sits outside Article 6(1), though it may still owe transparency obligations under Article 50 depending on what the AI does.
Use the crosswalk above to scope the real gap. Most of your Annex IV documentation already exists inside your MDR technical file under a different heading. The work that doesn't already exist is the data governance file and the human oversight design file — start there.
Check your notified body's AI Act designation status. Ask directly whether they're positioned to conduct an integrated assessment under Article 43(3), or whether you'll need a separate AI Act assessment from a different body. This affects both your timeline and your cost planning.
Build the data governance file first. Article 10's requirements on training data relevance, representativeness, and bias examination are the piece of Annex IV with the least overlap to existing MDR documentation. If your model was trained on a dataset you can't fully characterize, that's the finding you want to surface now, not during a 2027 assessment.
Treat human oversight as a design requirement, not a labeling exercise. Article 14 wants oversight measures built into the system, not just described in the instructions for use. That's a different conversation with your engineering team than the one MDR labeling requirements usually trigger.
Manufacturers building AI governance programs from scratch often benefit from a structured gap assessment before touching the technical documentation itself. Our healthcare and pharma practice page covers how this work fits into the broader regulated-industry AI governance picture.
A Word on ISO 42001
ISO/IEC 42001:2023, the AI management system standard, isn't a legal requirement under the AI Act, but it maps closely onto the Article 9–15 obligations, particularly risk management and continuous monitoring. For a manufacturer that already runs ISO 13485, adding an ISO 42001-aligned AI management system gives you a structured way to demonstrate the ongoing governance the AI Act expects, rather than treating each article as a standalone checklist item. It's worth evaluating alongside your MDR renewal cycle rather than as a separate project.
FAQ
What penalties apply if a medical device manufacturer misses these obligations? Article 99 of the AI Act sets tiered fines: up to €35 million or 7% of global annual turnover (whichever is higher) for violations of prohibited practices under Article 5, and up to €15 million or 3% of global annual turnover for non-compliance with other obligations, which includes the Article 9–15 requirements that apply to high-risk medical device AI. A separate, lower tier applies for supplying incorrect information to authorities.
Does the AI Act require a separate post-market monitoring system from MDR's post-market surveillance? No. Article 72(3) allows a provider already subject to post-market surveillance under other Union legislation, including MDR or IVDR, to integrate the AI Act's post-market monitoring plan into that existing system rather than standing up a second one. It's the same integration logic Article 43(3) applies to conformity assessment, applied to the ongoing monitoring obligation instead.
Do AI-enabled devices that don't trigger Article 6(1) have any AI Act obligations at all? Some do. Article 50 imposes transparency obligations on certain AI systems regardless of risk tier — systems that interact directly with people, generate synthetic content, or perform emotion recognition, for example, need to disclose that fact to users. A Class I device without AI as a safety component may sit outside Article 6(1) and still owe an Article 50 disclosure if the AI does one of those things.
What exactly has to go in the Article 10 data governance file that MDR doesn't already require? Annex IV(2)(d) asks for a description of the training, validation, and testing data sets: their provenance, scope and characteristics, how they were labeled, and what data cleaning or curation methodology was applied. MDR's technical file requires clinical and performance data to support safety and performance claims, but it doesn't ask for this level of detail on the data used to train and tune the model itself. That gap is usually the largest new documentation task in the whole exercise.
How does AI Act incident reporting compare to MDR vigilance reporting? Article 73 of the AI Act requires providers to report serious incidents involving a high-risk AI system to the market surveillance authority, generally within 15 days of becoming aware of a causal link. MDR Article 87 sets a comparable serious-incident reporting obligation, also within 15 days, running to national competent authorities. The timelines are close enough that a single incident-reporting procedure can often serve both, but confirm with your notified body whether the two reports can be combined or need to be filed separately.
Last updated: 2026-08-18
Jared Clark
Principal Consultant, Certify Consulting
Jared Clark is the founder of Certify Consulting, helping organizations achieve and maintain compliance with international standards and regulatory requirements.