EU AI Act Conformity Assessment and Harmonised Standards

A high-risk AI system cannot be placed on the EU market until its provider has shown, through a defined conformity assessment, that it meets the AI Act's requirements, and most providers still do not know which route applies to them or what evidence a notified body will ask for. This guide explains how conformity assessment works under Regulation (EU) 2024/1689 as amended by the Digital Omnibus on AI (Regulation (EU) 2026/1744 of 8 July 2026): the internal control and notified body routes, harmonised standards and common specifications, certificates, the EU declaration of conformity, CE marking and registration. It also covers the standards being written by CEN-CENELEC Joint Technical Committee 21 and what to do while they are unfinished. Status was checked on 27 October 2026 against the European Commission's AI Act Service Desk and EUR-Lex. It is written for AI product owners, compliance leads and security architects at providers of high-risk systems. For which systems are high-risk in the first place, see the EU AI Act High-Risk AI Systems Guide.

Key takeaways

The dates moved. Under the Digital Omnibus on AI, high-risk requirements apply from 2 December 2027 for Annex III systems and 2 August 2028 for systems covered by Annex I product legislation.
Most Annex III systems self-assess. Annex III points 2 to 8 use internal control (Annex VI); only biometrics (point 1) can require a notified body, and only where harmonised standards or common specifications are not fully applied.
Harmonised standards are the shortcut to presumption of conformity. Standards cited in the Official Journal give a presumption of conformity to the extent they cover the requirements; until then, providers must justify their own technical solutions.
The standards are still arriving. CEN-CENELEC JTC 21 is drafting them; prEN 18286 on the quality management system entered public enquiry on 30 October 2025. Build to the Act's text and the drafts, and plan to re-map once standards are cited.
Conformity is not one-off. A substantial modification triggers reassessment, notified body certificates last at most four years for Annex III systems, and the declaration of conformity must be kept for 10 years.

What it is and who it applies to

Conformity assessment is the process by which a provider demonstrates that a high-risk AI system meets the requirements in Chapter III, Section 2 of the AI Act: risk management, data and data governance, technical documentation, record keeping, transparency, human oversight, and accuracy, robustness and cybersecurity. It sits with the provider (the organisation that develops the system, or has it developed, and places it on the market under its own name). Deployers do not run conformity assessment, but a deployer that substantially modifies a system or puts its name on it can become a provider. See the EU AI Act Provider vs Deployer Obligations Guide.

System type	Conformity assessment route (Article 43)	Notified body?
Annex III point 1 (biometrics), harmonised standards or common specifications fully applied	Provider chooses internal control (Annex VI) or QMS and technical documentation assessment (Annex VII)	Optional
Annex III point 1, standards or specifications absent, only partly applied or published with a restriction	Annex VII	Yes; where law enforcement, immigration or asylum authorities or EU bodies put the system into service, the market surveillance authority acts as the notified body
Annex III points 2 to 8 (critical infrastructure, education, employment, essential services, law enforcement, migration, justice and democratic processes)	Internal control (Annex VI)	No
AI systems covered by Annex I, Section A product legislation (for example medical devices)	The sectoral conformity assessment procedure, with AI Act Section 2 requirements, the Article 17 QMS and specified Annex VII points folded in	The sectoral notified body, where the sectoral law requires one
System covered by Annex I, Section A that also falls under an Annex III use case	The Annex I sectoral procedure	As the sectoral law requires

The Digital Omnibus also changed the notified body rules: under amended Article 43(3), bodies notified under Annex I Section A legislation must apply for designation under the AI Act by 28 January 2028, with a single application procedure for bodies seeking designation under both regimes. It also amends Annex I and adds transitional arrangements for notified bodies; check the consolidated text on EUR-Lex for the detail.

Structure and key requirements

Harmonised standards (Article 40)

Presumption of conformity: a high-risk system that conforms to harmonised standards whose references are published in the Official Journal under Regulation (EU) No 1025/2012 is presumed to meet the Section 2 requirements, to the extent the standards cover them.
Standardisation request: the Commission's request to CEN and CENELEC covers ten areas: risk management, governance and quality of datasets, record keeping, transparency, human oversight, accuracy, robustness, cybersecurity, quality management and conformity assessment. Requests must also seek deliverables on energy and resource performance.
Joint compliance (new): the Omnibus added a duty on the Commission to request deliverables supporting joint compliance and presumption of conformity with both the AI Act and the relevant Annex I product legislation, so manufacturers face one set of standards, not two.

Common specifications (Article 41)

If European standardisation organisations reject or fail to deliver a requested standard on time, or deliver one that does not meet the request or does not sufficiently address fundamental rights, and no standard is expected within a reasonable period, the Commission may adopt common specifications by implementing act. Conformity to them also gives a presumption of conformity; providers that do not follow them must show equivalent technical solutions. Once a harmonised standard covering the same requirements is cited, the Commission repeals the overlapping common specifications. The AI Act Service Desk notes that Article 41 was amended by the Digital Omnibus and that its displayed text did not yet reflect the change when checked; read the consolidated text on EUR-Lex before relying on detail.

The JTC 21 standards

Deliverable (as named by CEN-CENELEC or the Commission)	AI Act anchor	Status when checked
prEN 18286 Artificial Intelligence, Quality Management System for EU AI Act Regulatory Purposes	Article 17	Entered public enquiry 30 October 2025 (Commission); final publication not yet reported on the pages checked
AI Trustworthiness Framework	Section 2 requirements as a whole	Under development (CEN-CENELEC)
AI Risk Management	Article 9	Under development
AI Conformity Assessment	Article 43 and Annexes VI and VII	Under development
Supplementary standards: datasets, bias, cybersecurity, robustness, logging, computer vision, natural language processing	Articles 10, 12, 15	Planned to follow (CEN-CENELEC)

JTC 21 was set up on 1 June 2021, works through five working groups and adopts ISO/IEC standards from SC 42 where they fit, developing European standards where they do not. Expect some deliverables to be adoptions or adaptations of ISO/IEC work. Drafts change between enquiry and publication: avoid building controls around clause numbers from a draft.

Identity and NHI requirements

The AI Act does not use the term non-human identity, but its requirements reach the credentials, service accounts and agent permissions around a high-risk system, and assessors will look for evidence on them.

Cybersecurity (Article 15): high-risk systems must be resilient against attempts by unauthorised third parties to alter their use, outputs or performance, including data poisoning, model poisoning and adversarial inputs. Access control on training pipelines, model registries and inference endpoints, and management of the keys and tokens that reach them, is the practical evidence. See the AI Model Risk Management Guide.
Record keeping (Article 12): automatic logging over the lifetime of the system. Logs are only useful if each actor, human or machine, has a distinct identity; shared service accounts defeat traceability.
Human oversight (Article 14): people assigned to oversight need the access and authority to intervene or stop the system, and that authority should be separate from the system's own runtime identity. See the AI Agent Kill Switch and Containment Guide.
Quality management system (Article 17): covers data management and change management; secrets and credential rotation for pipelines belong in it. See the AI Agent Credential Lifecycle and Rotation Guide.
Notified body access (Annex VII): notified bodies may access training, validation and testing datasets, including remotely through APIs, subject to security safeguards. Plan a scoped, time-limited, logged assessor identity rather than sharing an engineer's credentials.

Implementation steps

Classify and pick the route. Confirm whether each system is high-risk under Article 6, which Annex applies and which Article 43 route follows. Document any Article 6(3) "not high-risk" conclusion, which still requires registration under Article 49(2).
Stand up the QMS. Build the Article 17 quality management system first, using prEN 18286 as the working reference and ISO/IEC 42001 where you already run an AI management system.
Write the technical documentation as you build. Annex IV documentation, including any predetermined changes for systems that keep learning (Annex IV point 2(f)), so later updates are not substantial modifications.
Map requirements to standards and gaps. For each Section 2 requirement, record the harmonised standard, draft, common specification or own technical solution used, and why it is adequate.
Test. Run accuracy, robustness and cybersecurity testing, including adversarial testing; see the AI Red Teaming Guide. Keep test plans, data and results as evidence.
Assess. For internal control (Annex VI), have an independent internal reviewer check the QMS and documentation against the requirements. For Annex VII, choose a notified body early: capacity is limited and the Omnibus transitional rules affect which bodies can act.
Declare, mark, register. Draw up the EU declaration of conformity (Article 47), affix CE marking (Article 48) and register in the EU database (Article 49) before placing on the market.
Operate. Run post-market monitoring, track changes against the substantial modification threshold, and reassess when it is crossed.

Decision guidance

Internal control versus notified body for biometrics: once harmonised standards are cited and fully applied, internal control is cheaper and faster; a notified body certificate gives external assurance that some customers and insurers will ask for.
Standards now or own solutions: before standards are cited, own solutions are lawful but carry the burden of proof. Building to prEN drafts reduces rework later.
Small providers: one QMS across all systems, one documentation template and an external reviewer for the internal control check. Large providers: a central conformity function, a standards watch and a register of certificates and declarations.

Evidence and assessment

Artefact	Source in the Act	Key rule
Quality management system documentation	Article 17; Annex VII point 3	Notified body checks it against Article 17 and surveils it with periodic audits
Technical documentation	Annex IV; Annex VII point 4	Notified body may request further evidence or run its own tests; access to model parameters only as a last resort, on reasoned request
Union technical documentation assessment certificate	Annex VII; Article 44	Valid for the stated period, at most five years (Annex I) or four years (Annex III); renewable after reassessment; can be suspended, withdrawn or restricted
EU declaration of conformity	Article 47; Annex V	Written per system (machine readable, physical or electronically signed), kept for 10 years, one combined declaration where other Union law also requires one
CE marking	Article 48	Affixed before placing on the market
EU database registration	Article 49	Before placing on the market; Annex III point 2 systems registered nationally; law enforcement and migration systems in a secure non-public section
Logs, test results, change records	Articles 12, 15, 43(4)	Show substantial modifications were identified and reassessed

Mapping to other frameworks

AI Act element	ISO/IEC 42001	NIST AI RMF	Security controls (NIST SP 800-53 Rev. 5)
Article 17 QMS	AI management system clauses 4 to 10	Govern	PM-9 Risk Management Strategy
Article 9 risk management	AI risk assessment and treatment	Map, Measure, Manage	RA-3 Risk Assessment
Article 12 record keeping	Operation and monitoring of AI systems	Measure	AU-2 Event Logging, AU-12 Audit Record Generation
Article 15 cybersecurity	Supported by ISO/IEC 27001 controls	Manage	AC-6 Least Privilege, IA-5 Authenticator Management, SI-7 Software, Firmware, and Information Integrity
Article 43 conformity assessment	Internal audit and certification audit (not a substitute)	Govern	CA-2 Control Assessments

ISO/IEC 42001 certification is useful evidence for an Article 17 QMS but is not a conformity assessment under the AI Act and gives no presumption of conformity unless a corresponding harmonised standard is cited. See the ISO/IEC 42001 Guide and the NIST AI RMF Guide.

Measuring readiness

Route clarity: percentage of in-scope systems with a documented Article 6 classification and Article 43 route.
Requirement coverage: percentage of Section 2 requirements per system mapped to a standard, draft, common specification or justified own solution.
Documentation currency: days between a model or data change and the matching Annex IV update.
Change control: number of changes assessed against the substantial modification threshold, and how many triggered reassessment.
Identity hygiene: percentage of pipeline and inference service accounts that are unique, owned and rotated.

Common pitfalls

Working to the old dates. Plans built on 2 August 2026 for high-risk obligations are out of date; equally, do not treat the extra time as a reason to stop.
Assuming a notified body is always needed. Most Annex III systems use internal control; budget for the evidence, not the certificate.
Assuming ISO/IEC 42001 is enough. It supports the QMS but does not by itself demonstrate conformity with Section 2.
Treating drafts as final. prEN clause numbers and content can change before citation.
Missing substantial modifications. Retraining and model swaps without a predetermined change plan can require reassessment.
Shared service accounts in the pipeline. Article 12 logs that say "svc-ml" for every action cannot show who changed the model.

Practitioner checklist

Confirm the high-risk classification and the Article 43 route for each system.
Record the applicable date (2 December 2027 or 2 August 2028) per system.
Build the Article 17 QMS and use prEN 18286 as a working reference.
Map every Section 2 requirement to a standard, draft or justified own solution.
Write predetermined change plans into the Annex IV documentation.
Give pipelines, registries and inference endpoints unique, owned machine identities.
Plan a scoped, logged access route for notified body dataset access.
Prepare the declaration of conformity, CE marking and Article 49 registration steps.
Track JTC 21 deliverables and the Official Journal for cited standards.
Recheck the consolidated text on EUR-Lex before each milestone.

Standards and references

Related NHI Mgmt Group resources: EU AI Act High-Risk AI Systems Guide · EU AI Act Compliance Guide · EU AI Act Implementation Timeline Tracker · ISO/IEC 42001 Guide. This guide summarises the conformity assessment provisions of the EU AI Act as amended, as checked on 27 October 2026, and is not legal advice.

EU AI Act Conformity Assessment and Harmonised Standards Guide: routes, evidence and the JTC 21 standards