Five gaps between the AI Act high-risk criteria and our MDR QMS

A SaMD manufacturer pinged me last week asking whether the AI Act's high-risk classification overlapped with their existing MDR Class IIa SaMD. The short answer is yes. The longer answer is that the overlap is messier than the Commission guidance suggests, and the gaps between the two frameworks are showing up in our QMS workflows in ways I did not anticipate.

I have spent the last few weeks walking through Article 6 of the AI Act (the high-risk classification rules), Article 17 (the QMS requirement for providers of high-risk AI systems), and the relevant Annexes against our existing MDR Technical File structure. Here is what is landing for me in practice.

The overlap the Commission promised

The AI Act's recitals and the joint Commission guidance are clear: where a medical device is also a high-risk AI system under Annex I, the manufacturer follows the MDR conformity assessment pathway, and the AI Act's horizontal requirements are presumed to be satisfied to the extent the MDR framework covers them. Granted, the Q&A has evolved across drafts and the latest version is more specific than the early ones, but that operative phrase — "to the extent" — is doing a lot of work.

In practice this means: if your MDR Technical File already addresses risk management, data governance, transparency, human oversight, and post-market monitoring, the AI Act does not ask you to duplicate it. But the AI Act's expectations in some of these areas go further than the MDR baseline, and that is where the gaps open.

Where the gaps actually are

I am working through five areas where I do not yet have a clean mapping. If your team has a clean mapping, I would genuinely like to hear about it.

• Risk management under Article 9 vs ISO 14971. ISO 14971 is risk-based and lifecycle-wide. Article 9 of the AI Act is more prescriptive about specific risk categories (fundamental rights, bias, environmental, safety) and about risk acceptance criteria. The MDR framework does not naturally produce a "fundamental rights" risk register. We are now adding one as a separate artefact, cross-referenced from our ISO 14971 risk management files.
• Data governance under Article 10. The MDR asks for representative data and clinical evidence under Annex XIV, but the AI Act's data governance requirements are more granular — data quality criteria, bias detection, data provenance, annotation procedures. That granularity is meaningful. We are seeing questions in our notified body pre-submission correspondence that we could not previously have answered from the Technical File alone.
• Transparency under Article 13 and provider obligations under Article 16. Instructions for use under the MDR are well understood. The AI Act adds a layer: instructions must be concise, complete, correct, and clear, and for high-risk systems the provider must also give deployers information about the system's characteristics, capabilities, and limitations. The IFU is not the right home for all of this. Some of it belongs in deployer-facing documentation that did not exist before.
• Human oversight under Article 14. The MDR's "intended purpose" and "user" construct is narrower than the AI Act's human oversight model, which requires the system to be effectively overseen by natural persons during the period of use. That implies specific design choices (kill switches, explanation interfaces, override mechanisms) the MDR does not require for most devices. For SaMD this is not a gap, it is a new category of evidence.
• Post-market monitoring under Article 72 vs MDR PMS/PSUR. The AI Act has its own post-market monitoring system expectations, including active monitoring of serious incidents and reporting timelines that do not line up cleanly with the MDR's vigilance and PSUR cadence. The two regimes are not in conflict, but they are not the same. We are maintaining a side-by-side incident log to satisfy both until the delegated act clarifies the alignment.

The notified body question nobody can answer yet

Here is the part I find most uncomfortable. The Commission guidance is clear that the MDR notified body handles the AI Act conformity assessment via the MDR pathway. In practice this means your notified body — wherever you sit in the EU — is being asked to assess against AI Act requirements that did not exist when most of their auditors were trained. The audit checklists I have seen so far are inconsistent across bodies. The same SaMD can get materially different questions from different auditors. To be fair, the situation is genuinely new and the training is catching up. But I am telling my clients to budget for longer audit cycles and more document churn through 2025 and 2026.

What I am actually doing in the QMS

Practically, I have stopped trying to fold the AI Act into our existing MDR-only QMS. I am treating it as a parallel framework with explicit cross-references, which is roughly the framing I found in the qmsWrapper writeup on EU AI Act for medical device companies. The pitch that lands for me is not "AI does your compliance" but the connected-workflow point: when you open a change in your QMS, the system should already know which AI Act articles are in scope for that device, and it should prompt you to update the relevant cross-references. That is process automation in the boring, auditable sense.

The features that matter for my work are the ones that reduce cross-referencing errors: a change that touches a SaMD's risk management file should automatically flag the AI Act Article 9 mapping. A CAPA that touches data quality should prompt the Article 10 review. Native workflow integration between change control, CAPA, and risk is not glamorous, but it is what stops the gap between frameworks from becoming a finding.

Where I am still uncertain

I am not yet confident that post-market monitoring obligations under Article 72 can be fully discharged through the MDR PSUR and vigilance system. The Commission has hinted that the two will be harmonised through a delegated act, but the delegated act is not yet final. Until it is, I am double-reporting.

The other thing I do not have a clear answer on is the interaction with the Swiss MRA. Switzerland is not in the AI Act's regulatory perimeter, but devices placed on the EU market from Swiss-domiciled manufacturers are. I am mildly offended, as ever, that this is the regime I have to work in.

For those of you working through the same mapping — how are you handling the Article 9 risk register alongside ISO 14971? Running them as parallel documents, or have you found a way to integrate them into a single risk file that satisfies both frameworks without duplicating every line?

Disclosure: I work on qmsWrapper.