Article 17 of the AI Act requires providers of high-risk artificial intelligence systems to put in place a documented quality management system. Until a few weeks ago that obligation would have become enforceable on 2 August 2026, and the technical means of building it did not yet exist. Within a month the two swapped places: Regulation (EU) 2026/1744, the Digital Omnibus on AI in force since 27 July, moved the obligation to 2 December 2027 for Annex III systems and to 2 August 2028 for Annex I systems, and CEN and CENELEC approved EN 18286, the European standard designed for Article 17 itself. For once the tool arrives before the deadline.
What Article 17 requires
Article 17 of Regulation (EU) 2024/1689 calls for a documented system that follows the AI system across its whole life cycle, from design to data management, from post-market monitoring to the chain of internal responsibilities. The practical point on which everything turns, however, is one: the provider must put in writing which technical standards it intends to apply and, where it does not apply the harmonised standards in full, the other means by which it ensures compliance all the same. This is where the choice whether or not to adopt a standard stops being an organisational preference and becomes a duty to give written reasons.
On the change made by the Omnibus it is worth being precise, because it is narrower than it is usually reported to be. In the text of Article 17 the Digital Omnibus on AI touches only paragraph 2, where it adds that implementation shall be proportionate to the size of the organisation, in particular if the provider is an SME, including a start-up, or a small mid-cap enterprise. Proportionality, however, was already in the original text, and the saving clause, under which providers shall in any event respect the degree of rigour and the level of protection required to ensure compliance, is unchanged. Anyone expecting proportionality to work as a discount on the substance of the obligations is misreading it. The real change is the deferral: Chapter III, Sections 1, 2 and 3, which contains Article 17 as well, applies from 2 December 2027 for Annex III and from 2 August 2028 for Annex I.
Recital 40 explains why with unusual candour: the standards and the national authorities were not ready. It should be read for what it is, however, because the deferral operates by fixed dates and is not subject to any condition. No Commission act can bring it forward or push it back, and anyone reading it as conditional on standards becoming available will find the deadline arriving regardless.
The tool, and the certainty it does not yet give
On 29 July 2026 CEN and CENELEC announced, in a note titled “First Standard Approved under the AI Act”, that they had approved in June EN 18286, “Artificial Intelligence: Quality Management System for EU AI Act Regulatory Purposes”, the first European standard supporting the AI Act. The European Commission’s page on the standardisation of the AI Act describes the route: once harmonised standards are published by CEN and CENELEC, the Commission assesses whether they meet the intended objectives and legal requirements of the AI Act, and only after this final step are the standards referenced in the Official Journal of the EU. The application of standards remains voluntary and providers can choose any other framework to demonstrate compliance; harmonised standards referenced in the Official Journal, however, provide legal certainty, because those who apply them are presumed to be compliant. The same page notes that EN 18286, on 30 October 2025 still as prEN 18286, is the first harmonised standard for AI to enter public enquiry, and is designed to help providers of high-risk AI systems comply with Article 17.
That step into the Official Journal has not yet been taken: the Commission is expected to publish the reference later in 2026, and as at 30 July a search on EUR-Lex returns no results. A provider adopting the standard today is therefore building its system on the very tool commissioned for that purpose, and will have the presumption of conformity once the citation arrives; a provider choosing another route may do so, but must document how it ensures compliance. It is the same distinction already seen with the digital product passport, where six standards out of eight received the citation and two were left out.
Who verifies, and who supervises
For part of the high-risk population the quality management system will have to be verified by a notified body, and that side has moved too: on 28 July Accredia reported the new edition of document EA 2/17, which governs the accreditation of the bodies that certify for the purpose of notification, with national accreditation bodies required to align by 25 June 2027. That leaves the question of who will supervise: the newsletter of the Italian data protection authority of 29 July reports two opinions on the draft Italian decrees implementing the AI Act, which designate the Garante as market surveillance authority for high-risk systems used in justice, law enforcement, immigration, borders and democratic processes. They remain drafts, and the decree has not been published.
The dates, and what is worth doing
Few dates are worth marking. 2 August 2026 is the general date of application of the AI Act: what is added is the transparency obligations of Article 50 and market surveillance, not penalties or the designation of authorities, which have applied since 2 August 2025. 2 December 2027 and 2 August 2028 are the new Article 17 deadlines, for Annex III and Annex I respectively. In between, providers of high-risk systems would do well to use the sixteen extra months not to defer but to measure what they already do against what Article 17 requires.
