Introducing EN 18286: The AI Act's Quality Management Standard

In brief

On 14 October 2026 the AI Act acquired its first harmonised standard. Commission Implementing Decision (EU) 2026/2274 published the reference of EN 18286:2026, Artificial intelligence – Quality management system for EU AI Act regulatory purposes, in the Official Journal.¹² From that day, a provider of a high-risk AI system whose quality management system (QMS) conforms to the standard is presumed to meet the Article 17 requirements the standard covers.²

• What it is. A European standard drafted by CEN-CENELEC Joint Technical Committee 21 (JTC 21) at the Commission's request. It turns Article 17 into auditable QMS requirements organised around the AI system lifecycle.¹²
• What it covers. Article 17(1), including points (a) to (m), and the first sentence of Article 11(1), parts of it only on conditions set out in the standard's Annex ZA. Nothing more. Risk management, data governance, logging, transparency, human oversight, accuracy, robustness and cybersecurity still have no cited standard.³⁴
• What it is not. A substitute for ISO/IEC 42001. The Commission describes EN 18286 as designed to be layered on top of ISO/IEC 42001, ISO 9001 or a sectoral QMS such as ISO 13485, not to replace them.²
• Why now. After the Digital Omnibus, Annex III high-risk obligations apply from 2 December 2027 and Annex I obligations from 2 August 2028.⁵ The QMS is the longest-lead item on that path, and today it is the only one with a cited harmonised standard.

This is the first of three posts. The second maps how the AI Act, EN 18286 and ISO/IEC 42001 relate to each other. The third sets out what a Chief Compliance Officer should now be doing about all three.

Why the AI Act runs on harmonised standards

The AI Act is built on the EU's New Legislative Framework for product regulation. The Regulation sets legal requirements; European standardisation organisations, working to a Commission request, write the technical means of meeting them. Article 40(1) supplies the incentive: a high-risk AI system that conforms to a harmonised standard whose reference is published in the Official Journal is presumed to conform with the requirements that standard covers, and only to that extent.⁶

Three consequences follow, and each shapes how EN 18286 should be read.

• Publication is not citation. A standard adopted and sold by CEN-CENELEC is simply a European standard. The presumption attaches only once the Commission publishes its reference in the Official Journal.⁶³ EN 18286 sat in that gap from July to October 2026.
• The presumption is bounded. It reaches only the provisions mapped in the standard's Annex ZA, and it lasts only while the Official Journal keeps the reference.⁴
• Use remains voluntary. A provider may build its QMS another way. It then has no presumption to rely on and must show, on its own evidence, that the QMS meets Article 17.⁷ That is a separate point from Article 17(1)(e), which concerns the Section 2 requirements: wherever harmonised standards are not applied in full, the QMS must record the means used instead.

One textual wrinkle is worth flagging for counsel. Article 40(1) speaks of presumed conformity with the requirements in Section 2 of Chapter III (Articles 8 to 15). Article 17 sits in Section 3. The Commission's decision and its accompanying communication nonetheless frame EN 18286's presumption around Article 17, and Article 17(4) itself contemplates harmonised standards being relevant to QMS obligations.²⁷ EN 18286 itself mirrors the distinction: it reserves the term "essential requirements" for the Section 2 requirements and treats Article 17 as one of the wider regulatory requirements its QMS supports.⁴ We expect market surveillance authorities and notified bodies to follow the Commission's reading. The point would matter mainly if a presumption were ever contested.

How the Commission brought it into play

EN 18286 exists because the Commission asked for it, told the drafters what it had to achieve, and then decided whether the result was good enough to cite. From first request to citation took almost three and a half years.

Sources: ³, ⁴, ⁸, ⁹, ¹⁰, ¹¹, ¹².

The Commission used three levers along the way.

It framed the request. The 2023 request predated the final text of the Act and set an April 2025 deadline. The June 2025 replacement re-based the work on Regulation (EU) 2024/1689 and moved the deadline for deliverables to August 2026.⁹¹² Throughout, the AI Office sat in the working groups and assessed drafts against the request as they developed.¹² According to JTC 21's own reporting, the amended request expires on 28 February 2027.¹⁰

It set the content brief. The Commission's Joint Research Centre told standard-setters what would be different about AI Act standards. They must address risks to health, safety and fundamental rights, rather than the organisation's own objectives. They must be oriented to AI systems and products, complementing the organisation-centric view of existing international standards. And they must lead to verifiable outcomes.¹³ The Commission also checks each draft for eligibility as a harmonised standard before it goes to public enquiry.¹⁰ That brief explains why JTC 21 wrote a new QMS standard rather than simply adopting ISO/IEC 42001, a point the second post in this series develops.

It decided on citation. Once CEN-CENELEC published the standard, the Commission assessed it against the request and the Act before publishing the reference.⁸ Implementing Decision (EU) 2026/2274 is short: six recitals, two articles and a one-line annex listing EN 18286:2026 without restriction. Its recitals record that the standard answers point 9 of the 2025 request, on quality management systems including post-market monitoring, and that it satisfies the requirements it aims to cover.¹⁴ The decision entered into force on publication on 14 October 2026, which was World Standards Day.¹³

What EN 18286 requires

EN 18286 specifies requirements for the QMS of a provider of AI systems, with providers of high-risk AI systems as its primary audience, in any sector.⁴¹¹ One QMS may cover several AI systems, and the standard expressly contemplates providers established outside the EU that place systems on the EU market.⁴ It is not written for general-purpose AI model providers, and it addresses deployers only where they become providers under Article 25, for example by substantially modifying a high-risk system.⁴

Two design choices set it apart from familiar management system standards.

• Quality means regulatory compliance. The standard defines quality as, at a minimum, compliance with the AI Act and with the law protecting fundamental rights, health and safety.⁴ Customer satisfaction and business objectives are not the yardstick.
• The AI system lifecycle, not the organisation, is the organising unit. The Commission describes the standard as built around the lifecycle rather than the organisation.²

The requirements sit in clauses 4 to 10.

Source: ⁴.

Four informative annexes complete it. Annex A covers consultation of interested parties on risks to fundamental rights, health and safety. Annexes B and C map clauses to EN ISO 9001:2015 and EN ISO/IEC 42001:2026. Annex ZA is the bridge to the AI Act.⁴ Two absences stand out. There is no continual improvement clause and no internal audit clause: audit results appear only as a recommended input to management review, and improvement runs through management review and the planning of changes.⁴ Consultation with affected persons is likewise recommended rather than required.⁴

Several requirements are more specific than readers of ISO management system standards will expect:

• test plans must be reproducible, recording such things as the compute environment, library versions, random seeds, hyperparameters and dataset versions (clause 8.4.2);
• each AI system, and each version that undergoes conformity assessment, needs a unique reference linked to its instructions for use, technical documentation and datasets (8.7);
• the provider must document why each post-market modification is or is not substantial; a substantial one produces a new AI system that needs a new conformity assessment (9.4);
• serious incidents must be reported within fixed time limits, and deployers need a route to report incidents to the provider and to suspend use (9.6).

Source: ⁴.

How far the presumption reaches

Annex ZA's Table ZA.1 maps clauses to Article 17(1), points (a) to (m), and to the first sentence of Article 11(1), the duty to draw up technical documentation before placing on the market and keep it up to date. Four rows carry conditions or caveats. The technical documentation row applies only if the documentation contains every element listed in Annex IV. The point (g) row applies only if the risk management system complies with Article 9. The row for the documentation duty in Article 17(1) applies only if all of points (a) to (m) are documented. And the table records that applying the standard does not by itself fulfil every regulatory requirement; each must still be examined, applied and verified. Article 17(2) to (4) are marked as not covered.⁴

Sources: ⁴, ¹⁴.

The Article 9 row deserves emphasis. Table ZA.1 makes the presumption for point (g) conditional on a risk management system that complies with Article 9, and EN 18286's own risk management clause simply requires such a system and points to the draft EN 18228 for how to build it.⁴¹⁴ Until a risk management standard is cited, the point (g) presumption is only as strong as the provider's own Article 9 evidence. Providers who read the citation as covering their whole Section 2 compliance will be disappointed.

Where EN 18286 sits in the timetable

EN 18286 is one of ten deliverables under the request, and at the end of October 2026 it is the only one cited. The rest of JTC 21's programme has struggled at public enquiry: in summer 2026 the drafts on risk management, cybersecurity and logging were all voted down.¹⁰

Sources: ¹⁰ (status as at mid-September 2026), ³.

Three timing points follow.

The obligations moved; the presumption did not. The Digital Omnibus, Regulation (EU) 2026/1744, deferred the high-risk rules to 2 December 2027 for Annex III systems and 2 August 2028 for Annex I systems. The stated reason was that harmonised standards and notified bodies were not ready.⁵ The Commission had originally proposed linking application to the availability of standards and other support tools.⁸ Article 40 was not deferred, so EN 18286 supports a presumption now, more than a year before Article 17 binds anyone.¹⁴

Section 2 standards may not all be cited by December 2027. The amended request expires on 28 February 2027.¹⁰ Providers should plan to demonstrate Articles 9 to 15 through their own documented means under Article 17(1)(e), or through common specifications if the Commission adopts them under Article 41.

For biometrics, the gap has a direct cost. A provider of an Annex III point 1 biometric system may choose internal control only if it has applied harmonised standards or common specifications for Section 2. Where none exist, Article 43(1) routes it to notified-body assessment under Annex VII.¹⁵ EN 18286 alone does not change that.

What providers should do with it now

The citation turns EN 18286 from a reference document into the default yardstick for Article 17. Whoever audits a QMS for a high-risk system, whether the provider itself under Annex VI or a notified body under Annex VII, will now read it against EN 18286.¹⁴

1. Confirm that you are a provider of a high-risk system. EN 18286 is a provider's standard. Deployers come into it only if they rebrand, substantially modify or repurpose a system under Article 25.⁴ The third post in this series covers role mapping.
2. Gap-assess what you already run. Annexes B and C of EN 18286 map its clauses to EN ISO 9001:2015 and EN ISO/IEC 42001:2026. They map clauses, not requirements: correspondence is not equivalence.⁴
3. Choose your presumption strategy. Apply the clauses mapped in Annex ZA in full to claim the presumption for Article 17(1); where you depart from them, you lose it for the aspects affected. For the Section 2 requirements, EN 18286 makes you choose, requirement by requirement, between cited harmonised standards, common specifications, other standards and other technical specifications, and record in the technical documentation how each requirement not fully covered is met (clause 4.4.3).⁴⁷
4. Use the special routes where they apply. Providers already under a sectoral QMS regime may fold the Article 17 aspects into it (Article 17(3)). Financial institutions are deemed compliant through their internal governance rules, except for risk management, post-market monitoring and incident reporting, and must still take harmonised standards into account (Article 17(4)). SMEs and small mid-caps benefit from proportionality under the amended Article 17(2), without any lowering of rigour.⁷ None of these routes is covered by EN 18286's presumption, and Annex ZA states that applying the standard does not by itself ensure proportionate implementation.⁴
5. Put it into contracts. Buyers of high-risk AI systems can now ask suppliers a concrete question: does your QMS conform to EN 18286:2026, and on what evidence? The third post returns to this.

EN 18286 is available through national standards bodies, which must adopt it nationally by January 2027.⁴ The second post in this series turns to how it relates to the AI Act and to ISO/IEC 42001.

Related reading in this series

• Part 2 — The AI Act, EN 18286 and ISO/IEC 42001: How the Three Fit Together
• Part 3 — AI Act, EN 18286 and ISO/IEC 42001: A Chief Compliance Officer's Agenda

Disclaimer

This article is general information, not legal advice, and does not create a client relationship. It reflects the AI Act as amended by Regulation (EU) 2026/1744 and the standards landscape as at 29 October 2026. Standards in development change status often; check with your national standards body before relying on a status reported here. Statements about EN 18286 have been checked against the text of EN 18286:2026 and are paraphrased, not reproduced. The standard is a copyrighted document, available from national standards bodies, and should be read in full before any conformity claim is made.

Notes and sources

Footnotes

1. CEN-CENELEC, First ever European standard on AI published in the OJEU, 15 October 2026. ↩ ↩² ↩³

2. European Commission, First harmonised standard supporting the AI Act published in the Official Journal, 19 October 2026. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶

3. Commission Implementing Decision (EU) 2026/2274 of 12 October 2026 on the harmonised standards for high-risk AI systems drafted in support of Regulation (EU) 2024/1689, OJ L, 14 October 2026. ↩ ↩² ↩³ ↩⁴ ↩⁵

4. CEN, EN 18286:2026, Artificial intelligence – Quality management system for EU AI Act regulatory purposes, approved by CEN on 12 July 2026 (available from national standards bodies). ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹ ↩¹⁰ ↩¹¹ ↩¹² ↩¹³ ↩¹⁴ ↩¹⁵ ↩¹⁶ ↩¹⁷ ↩¹⁸ ↩¹⁹ ↩²⁰ ↩²¹

5. Regulation (EU) 2026/1744 of 8 July 2026 (Digital Omnibus on AI), OJ L, 2026/1744, 24 July 2026, as reflected in the EUR-Lex consolidated text of the AI Act as at 27 July 2026; see also the summary Digital Omnibus on AI: what changed in the EU AI Act, euaiact.com. ↩ ↩²

6. AI Act, Article 40, consolidated text as at 27 July 2026, AI Act Service Desk. ↩ ↩²

7. AI Act, Article 17, consolidated text as at 27 July 2026, AI Act Service Desk. ↩ ↩² ↩³ ↩⁴

8. European Commission, Standardisation of the AI Act. ↩ ↩² ↩³

9. Digital Policy Alert, Standardisation request C(2023) 3215 was repealed and replaced. ↩ ↩²

10. G. P. Krog, Edition 16 (September 2026): the essence, AI Standardisation Watch, 8 September 2026, summarising the CEN-CENELEC JTC 21 Task Group Inclusiveness newsletter. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶

11. CEN-CENELEC, EN 18286 in the Spotlight: Supporting Compliance with the AI Act, 31 July 2026. ↩ ↩²

12. Council of the EU, WK 12134/2025, Standards in support of AI Act – Presentation, 23 September 2025. ↩ ↩² ↩³

13. Joint Research Centre, Harmonised Standards for the European AI Act, Science for Policy Brief JRC139430, October 2024. ↩

14. spilma, AI Act: EN 18286 is the first harmonised standard cited, updated 16 October 2026. ↩ ↩² ↩³ ↩⁴ ↩⁵

15. AI Act, Article 43, consolidated text as at 27 July 2026, AI Act Service Desk. ↩