Skip to content

Conformity Assessment Procedures for High-Risk AI (Article 43)

10 min read · updated August 11, 2026

The question every provider of a high-risk AI system asks first is whether an outside body has to sign it off. For most of Annex III the answer is no, and the exception is narrower and more conditional than the reputation of the Act suggests.

The two routes

Article 43 of Regulation (EU) 2024/1689 sends a high-risk AI system down one of two procedures, set out in the annexes:

  • Annex VI — conformity assessment based on internal control. The provider verifies that its quality management system complies with Article 17, examines the technical documentation against the Chapter III Section 2 requirements, and verifies the design and development process and post-market monitoring are consistent with that documentation. No third party is involved. The provider then draws up the EU declaration of conformity under Article 47 and affixes CE marking under Article 48.
  • Annex VII — assessment of the quality management system and of the technical documentation, with a notified body. An accredited third party assesses the QMS, approves it, carries out surveillance, and separately examines the technical documentation. See the notified bodies page.

Both end in the same place — declaration of conformity, CE marking, registration in the EU database under Article 49 — which is why the routes are easy to conflate. The difference is entirely who does the verifying. The output side is on the CE marking page.

Which route applies to a specific system depends on its Annex III classification and on the standards applied, and getting that classification wrong is the expensive error here. This page describes the routing rule and is not legal advice; take advice on the classification of your own system.

Which Annex III category takes which route

The rule splits Annex III into point 1 and everything else.

Annex III point 1 — biometrics. Article 43(1) gives a conditional choice. Where the provider has applied harmonised standards referred to in Article 40, or where applicable the common specifications referred to in Article 41, in demonstrating compliance with the Chapter III Section 2 requirements, it may choose between the Annex VI internal control route and the Annex VII notified body route. Where those harmonised standards have not been applied, or have been applied only in part, or where such harmonised standards do not exist and common specifications are not available, the provider shall follow Annex VII — the notified body route. The category is described on the Annex III biometrics page.

Annex III points 2 to 8. Article 43(2) is unconditional: providers shall follow the conformity assessment procedure based on internal control as referred to in Annex VI, which does not provide for the involvement of a notified body. That covers critical infrastructure, education, employment, essential private and public services, law enforcement, migration and border control, and the administration of justice and democratic processes. All of it is self-assessed.

This surprises people, and it is worth stating the design plainly rather than editorialising: the Act's assurance for those categories comes from the substance of the Chapter III requirements, the quality management system under Article 17, registration in the public database, post-market monitoring and market surveillance powers — not from ex ante third-party certification. Whether that is the right balance is a live policy argument; whether it is what the Act says is not.

Article 43(6) gives the Commission power to adopt delegated acts amending paragraphs 1 and 2 to subject Annex III points 2 to 8 systems to the Annex VII procedure, where internal control proves insufficient. So the allocation above is stable but not fixed, which is one reason this page carries a revisit marker.

Products already regulated under Annex I

A different rule applies to high-risk systems that are safety components of products, or are themselves products, covered by the Union harmonisation legislation listed in Annex I Section A — machinery, medical devices, in vitro diagnostics, lifts, toys, radio equipment and the rest.

Article 43(3) requires the provider to follow the conformity assessment procedure required under those legal acts, rather than a separate AI Act procedure. The AI requirements are folded into the existing assessment: the Chapter III Section 2 requirements apply and the notified body already involved under the sectoral legislation is entitled to control the AI system's conformity with them. Certain Annex VII requirements apply within that process.

The practical effect is that a manufacturer of an AI-enabled medical device does not get a second CE process; it gets an expanded one, run by a notified body that must now be competent on the AI requirements too. That competence bottleneck is real, and it is the reason the Annex I interface is worth watching more closely than the Annex III one: the constraint is the supply of assessors rather than the content of the rule.

Why harmonised standards decide this

Article 40 provides that high-risk AI systems in conformity with harmonised standards, or parts thereof, whose references have been published in the Official Journal, are presumed to be in conformity with the Chapter III Section 2 requirements to the extent those standards cover them. Article 41 lets the Commission adopt common specifications where the standardisation process fails or is delayed.

Those two articles are what make Article 43(1) operable, and they are also the weak point in the machinery. The Commission issued a standardisation request to CEN and CENELEC, and the work sits with the joint technical committee JTC 21. At the time of writing, harmonised standards covering the Chapter III requirements had not been published in the Official Journal, which has a direct consequence for Annex III point 1 providers: without them, Article 43(1) routes you to Annex VII by default. The status of that work is tracked on the JTC 21 page.

Note also that conformity with a harmonised standard is a presumption rather than a safe harbour in the absolute sense: it is rebuttable, and it only reaches the requirements the standard actually covers. Providers that treat a certificate as the end of the analysis are over-reading it.

Substantial modification and re-assessment

Article 43(4) requires a high-risk AI system that has already been through a conformity assessment to undergo a new one in the event of a substantial modification, regardless of whether the modified system is intended to be further distributed or continues to be used by the current deployer.

It then carves out the case that would otherwise make continuously learning systems unmanageable: for systems that continue to learn after being placed on the market or put into service, changes to the system and its performance that have been pre-determined by the provider at the moment of the initial conformity assessment, and are part of the information contained in the technical documentation referred to in point 2(f) of Annex IV, do not constitute a substantial modification.

That is a strong incentive to specify the intended change envelope up front, in the technical documentation, rather than to discover afterwards that a retraining cadence has triggered re-assessment. The same idea appears in medical device regulation as a predetermined change control plan, which is a useful analogue for how much specificity the envelope needs to have. What counts as substantial in the first place is the harder question and is dealt with on the substantial modification page; Article 3(23) defines it, and its application to model updates is not settled.

On timing, both dates in this area moved and the gap between the two routes widened. Regulation (EU) 2024/1689 as adopted set 2 August 2026 for Annex III systems under Article 113 and 2 August 2027 for high-risk systems falling under Article 6(1), the Annex I product route. Regulation (EU) 2026/1744 postponed the first to 2 December 2027 and the second to 2 August 2028. Article 111(2) contains transitional rules for systems placed on the market before those dates.

For a provider choosing a route, the extra time is doing something specific: the Annex III point 1 choice between Annex VI and Annex VII turns on whether harmonised standards exist, and a later application date is more room for CEN-CENELEC JTC 21 to deliver them. It does not change the rule. If the standards are still unpublished when the date arrives, Article 43(1) still routes biometrics providers to a notified body.

The postponements stated above come from Regulation (EU) 2026/1744, the digital omnibus on AI, published in the Official Journal on 24 July 2026 and in force from 27 July 2026. Nothing else on this page is attributed to it: the two-route structure, the Annex III allocation and the Article 43(4) substantial-modification rule are stated from Regulation (EU) 2024/1689 as adopted and were not re-verified against the amending text. The status of harmonised standards is stated as at the time of writing. Before deciding a route, check the consolidated AI Act on EUR-Lex and the Official Journal listing of harmonised standards.