Direct answer

When do the AI Act obligations start to apply to high-risk AI systems?

The obligations for high-risk AI systems that carry a date all apply from 2 December 2027. 2 provisions carry no date of their own: they have no deadline of their own. Below is when each provision starts to apply.

17 obligations under the AI Act bear on this, of which 2 apply today.

Which AI is high-risk, with examples per area and what to do: Annex III of the EU AI Act.

Zahed Ashkara, AI and privacy legal adviserobligations checked against the source on 8 August 2026

Check your own case for free

First step: Set up data governance per dataset.

The conclusion and your first steps

Preliminary reading · Article 21: cooperation with competent authorities

Two things are unsettled here, and the status carries a cost that we state alongside them. First, who the competent authority is. Article 21 refers without qualification to a competent authority. Article 3(48) defines the national competent authority as a notifying authority or a market surveillance authority, and provides in the same point that, as regards AI systems put into service or used by Union institutions, agencies, offices and bodies, references to national competent authorities or market surveillance authorities are construed as references to the European Data Protection Supervisor. We read Article 21 as addressing those national competent authorities. The stronger alternative reading is that competence follows Chapter IX: the market surveillance authority designated by the Member State, the financial supervisor via Article 74(6), the data protection supervisory authority via Article 74(8), the European Data Protection Supervisor via Article 74(9), and, for a delimited group of systems, the AI Office via Article 75 as amended by Regulation (EU) 2026/1744, which already applies. On that reading you prepare for a different counterpart than the one we assume here. Second, what paragraph 2 asks when the logs are in fact held by the deployer. The text limits the duty to logs under your control but does not say whether you must arrange access contractually in order to retain that control. We read no separate duty into it; that is our reading and not the text. What is settled here is the date of application, and that corrects an earlier reading of ours. Article 1, point (40)(b), of Regulation (EU) 2026/1744 replaces Article 113, third paragraph, point (c), and expressly sets Chapter III, Sections 1, 2 and 3, with the exception of Article 6(5), at 2 December 2027 for systems classified as high-risk pursuant to Article 6(2) and Annex III and at 2 August 2028 for those classified as high-risk pursuant to Article 6(1) and Annex I. Article 21 sits in Section 3 and is therefore covered by that provision in so many words; the fact that the omnibus does not name Article 21 individually changes nothing, because the provision operates per Section. The cost of this status: a preliminary reading does not count in the per-situation derivation of obligations, so the log access of paragraph 2 and the language rule do not surface there, and Article 16(k) covers only the demonstration of conformity. The object stays reachable through its own page, the deadline index and the API.

The text of the provision is settled. How it plays out has not been decided in case law or guidance, so this is our reasoned reading and not established law. Weigh it yourself, or put it to your own lawyer.

Preliminary reading · Article 4a: legal basis for bias testing with special categories of personal data

Two things are open here, and the risk runs the other way round than with a right such as Article 86: a broad reading here benefits the controller and not the data subject, because this concerns data on ethnicity, health, religion, trade union membership and sexual orientation. Where in doubt the narrow reading is therefore the safe one. First, the reach of "other AI systems and models" in paragraph 2, which on its face covers any AI system and any model and for which no delimitation exists. We read it on its face, but with the threshold in paragraph 2, point (a), as the real boundary: without a consequence for health and safety, fundamental rights or prohibited discrimination there is no basis. A defensible alternative reading is that paragraph 2 is confined to systems comparable to the examples in recital 9, such as scoring tools for permits and public services. Second, the relationship with Article 9 GDPR. Recital 9 states that the extension is subject to the same limitations, conditions and safeguards and thereby ensures compliance with Article 9(2), point (g), of Regulation (EU) 2016/679, so we read Article 4a as the Union law measure that point requires, with the safeguards carried by the six conditions themselves. The counterargument stands against that and has not gone away, but it has narrowed since 27 July 2026: the anchoring sits in a recital and not in the article, and the article itself designates no ground from Article 9(2). What no longer supports that counterargument is Article 2(7). Until 27 July 2026 that paragraph left the GDPR unaffected without reservation, but it was replaced by Article 1, point 2(b), of Regulation (EU) 2026/1744 and now reads: "Union law on the protection of personal data, privacy and the confidentiality of communications applies to personal data processed in connection with the rights and obligations laid down in this Regulation. Without prejudice to Articles 4a and 59 of this Regulation, this Regulation shall not affect Regulation (EU) 2016/679 or (EU) 2018/1725, or Directive 2002/58/EC or (EU) 2016/680." The Union legislature therefore carved the reservation out for Article 4a precisely, which points towards reading Article 4a as the Union law measure itself rather than a mere cross-reference to the GDPR. Anyone citing this object while quoting the former wording of Article 2(7) is quoting a replaced provision. A defensible alternative reading remains that a national or Union measure with specific safeguards is still needed alongside it, but it now rests only on the absence of an express designation in the article itself. On the date from which the basis operates, part is settled and part is not. What is settled is what recital 9 says, namely that the basis should apply from the date of entry into application of Regulation (EU) 2024/1689; that has been read and is not a house reading. What remains open is which date this object therefore carries. We hold to 27 July 2026, the day Article 4a entered the text, because a basis that was not yet there was in fact not available. The alternative reading follows recital 9 literally and lets the basis reach back to the date of application of the base Regulation. That difference is not academic for anyone who has to justify processing from that period. The editorial statement below marks that choice as our inference.

Preliminary reading · Article 75(1a) and (1e): reporting to and assessment by the AI Office

The two duties themselves are literal in the text; the date is our derivation. Article 75 sits in Chapter IX, which has applied since 2 August 2026, but both paragraphs address only the provider of a high-risk AI system, and that status only arises when Chapter III, Sections 1 to 3, becomes applicable. That date depends on the route, which the repair after the cross-review made sharper: point (c) of the third paragraph of Article 113, as replaced by Article 1, point (40)(b), of Regulation (EU) 2026/1744, gives 2 December 2027 for the route of Article 6(2) and Annex III, and 2 August 2028 for the route of Article 6(1) and Annex I. The Annex I route does not drop out here: the four carve-outs in Article 75(1) sit inside point (a), and point (b) independently brings a system that constitutes, or is integrated into, a designated very large online platform or very large online search engine under the competence of the AI Office, including where it is high-risk through Annex I. We carry the earlier of the two in deadline_at, 2 December 2027, matching the sister object article-21-cooperation-with-authorities, which carries the same class of duty holder; for a system entering through point (b) and Annex I the date is 2 August 2028. A defensible alternative reading is that paragraph 1a, as a Chapter IX provision, already operates from 2 August 2026 and therefore applies immediately to any system that is at some point classified as high-risk; on that reading you set up the reporting route to the AI Office now. That is the safer choice and it costs little. It is further uncertain exactly when paragraph 1e bites: Article 43 sits in Chapter III, Section 5, which is not among the deferred Sections 1 to 3, but a third-party conformity assessment only arises once a system is classified as high-risk. The cost of this status is that a preliminary reading is skipped by lib/answer/derive-obligations.ts, so this reporting route does not surface in the per-situation derivation; the object stays reachable through its own page, the deadline index and the API.

Preliminary reading · Article 8: compliance with the requirements for high-risk AI systems

What the provision asks of a provider is settled; when it asks it has been derived here. Article 8 sits in Chapter III, Section 2, and names no date of its own. The timeline canon dates the high-risk requirements per route and not per chapter: 2 December 2027 for the systems that are high-risk through Annex III, and 2 August 2028 for those that are high-risk through Annex I. Those two dates do not fit in a field that carries one. `deadline_at` therefore carries 2 December 2027, the earlier of the two and the same date already carried by the objects for Articles 9 to 15. A defensible alternative reading is that a chapeau provision should carry no date of its own and that the dating belongs with the seven requirements it sits above; on that reading this field reads as the earliest of the two routes and not as the date of Article 8. On either reading: for a system inside a regulated product under Annex I, 2 December 2027 is too early and 2 August 2028 is the date that counts. What also remains open is the content of the second measure in paragraph 1. The generally acknowledged state of the art is not a term the Regulation defines and not a synonym for a harmonised standard; anyone who equates it with the standard of Article 40 closes an open measure. On Regulation (EU) 2026/1744. That text was retrieved after all on 6 September 2026 from the Publications Office Cellar service and is archived in data/ai-act/review/sources/reg-eu-2026-1744-nl.txt and -en.txt. The amending regulation carries forty-three amendment points in its Article 1; which of those are reflected in the local legal texts is recorded per point in data/ai-act/review/consolidation-manifest.json. Where a point touches this Article, that is stated below with the statement concerned.

This applies now

Coming up

Article 10: data and data governancefrom 2 December 2027
Article 11: technical documentationfrom 2 December 2027
Article 12: logging and traceabilityfrom 2 December 2027
Article 13: transparency towards deployersfrom 2 December 2027
Article 14: human oversightfrom 2 December 2027
Article 15: accuracy, robustness and cybersecurityfrom 2 December 2027
Article 16: the twelve duties of a provider of a high-risk AI systemfrom 2 December 2027
Article 17: quality management systemfrom 2 December 2027
Article 18: documentation keepingfrom 2 December 2027
Article 20: corrective actions and duty of informationfrom 2 December 2027
Article 21: cooperation with competent authoritiesfrom 2 December 2027
Article 26: obligations of deployers of high-risk AI systemsfrom 2 December 2027
Article 75(1a) and (1e): reporting to and assessment by the AI Officefrom 2 December 2027
Article 8: compliance with the requirements for high-risk AI systemsfrom 2 December 2027
Article 9: risk management systemfrom 2 December 2027

When it starts to apply

2 December 2027

Date to follow

education
Logging in admission and assessment: what the institution keeps in its own hands
A university of applied sciences uses a purchased AI system that ranks student admissions and also raises flags during digital assessment. The logs sit in the supplier environment, which hands them over on request. The teaching organisation wonders whether that settles the matter or whether the school retains a duty of its own.
Provenance: Article 12(1) requires a high-risk AI system to technically allow for the automatic recording of events (logs) over the lifetime of the system, and Article 12(2) ties that recording to a level of traceability appropriate to the intended purpose of the system. Article 26(6) provides that deployers keep the automatically generated logs to the extent those logs are under their control, for a period appropriate to the intended purpose, of at least six months, unless applicable Union or national law provides otherwise, in particular Union law on the protection of personal data. Article 19(1) places a corresponding retention duty on providers for the logs generated by their systems that are under their control, likewise for at least six months.
On our reading, the fact that the supplier hosts the logs does not by itself place the school outside Article 26(6): what matters is whether the logs are under its control. The Regulation does not define that notion, and on our reading server location is not decisive in itself; the question is whether you can actually and contractually dispose of those logs, and it has to be answered for each procurement. So agree at procurement that the institution can request, export and itself retain the logs for the chosen period, and record which admission and assessment events that covers. Take data protection into account at the same time, because the same provision allows other Union or national law to cap the retention period.
Editorial example. The rule above is in the Regulation. The situation was written by us to show how that rule plays out in this sector, and is not taken from a worked case in official guidance.
Article 12(1) and (2), Article 19(1) and Article 26(6)
manufacturing and industry
Logging on the production line: which logs the manufacturer keeps and which the factory keeps
A manufacturer supplies an AI system that runs as a safety component inside the machinery of a production line and also drives quality control. The factory operating the line keeps only the alerts visible in the local controller; the rest of the recording flows to the supplier environment. The question is who has to keep which logs when it later has to be reconstructed why the line was halted.
Provenance: Article 12(1) requires high-risk AI systems to technically allow for the automatic recording of events (logs) over the lifetime of the system. Under Article 12(2), those logging capabilities must enable the recording of events relevant for identifying situations that may result in the system presenting a risk within the meaning of Article 79(1) or in a substantial modification, for facilitating the post-market monitoring referred to in Article 72, and for monitoring the operation of the high-risk AI systems referred to in Article 26(5). Article 19(1) obliges providers to keep those automatically generated logs to the extent they are under their control, for a period appropriate to the intended purpose of the system, of at least six months, unless applicable Union or national law provides otherwise. Article 26(6) places a corresponding retention duty on deployers for the logs under their control, with the same six-month floor.
We read Article 19 together with Article 26(6) as leaving the manufacturer, acting here as the provider, and the factory, acting as the deployer, each with a retention duty of its own for the logs under its control. The Regulation does not say when logs count as being under your control, and on our reading the place where the recording physically lands does not settle that by itself; the question has to be answered system by system. For a production line that means recording which events stay in the machinery controller and which travel to the supplier, and securing access to that second set contractually before you need it. Leave that unarranged and you risk being unable to trace a line stoppage or a quality control rejection back to the behaviour of the system.
Editorial example. The rule above is in the Regulation. The situation was written by us to show how that rule plays out in this sector, and is not taken from a worked case in official guidance.
Article 12(1) and (2), Article 19(1) and Article 26(6)
workplace and staff
Logging in task allocation at work: evidence about the system or a file on the employee
An employer deploys an AI system that handles task allocation among staff and summarises their performance for the performance review. HR wants to know what record of those outcomes has to be retained when an employee objects months later to a promotion decision.
Provenance: Article 12(1) requires high-risk AI systems to technically allow for the automatic recording of events (logs) over the lifetime of the system. Article 12(2)(c) names, as one of the purposes of those logging capabilities, monitoring the operation of the systems referred to in Article 26(5), and that paragraph obliges deployers to monitor operation on the basis of the instructions for use and, where relevant, to inform the provider in accordance with Article 72. Article 26(6) obliges deployers to keep the automatically generated logs to the extent they are under their control, for a period appropriate to the intended purpose, of at least six months, unless applicable Union or national law provides otherwise. Article 19(1) sets out a corresponding retention duty for providers, with the same six-month floor.
We read Article 12(2) as putting the logging capabilities there first of all to follow risks, substantial modifications and the operation of the system, not to sharpen judgements about individual staff. Whether that statement of purpose also limits what the retained logs may later be used for is something Article 12 does not say: on our reading that limit has to come from data protection law, to which Article 26(6) itself refers. For HR a practical line follows: keep what is needed to trace an outcome and to carry out the monitoring under Article 26(5), and settle in writing beforehand whether those same files may double as a performance record. Bear in mind that the six-month floor can be shorter than the period within which an employee challenges a promotion decision; the Regulation does not govern that evidential position.
Editorial example. The rule above is in the Regulation. The situation was written by us to show how that rule plays out in this sector, and is not taken from a worked case in official guidance.
Article 12(1) and (2)(c), Article 19(1) and Article 26(5) and (6)
retail and consumer
Webshop builds its own consumer credit check: what belongs in the file
A non-food retail chain lets customers pay later in its webshop and decides at checkout whether a consumer qualifies. The team builds that assessment system in house, on top of a pre-trained model supplied by a vendor. The question is what has to be on record before the feature goes live, and who has to put it there.
Provenance: Article 11(1) requires the technical documentation of a high-risk AI system to be drawn up before that system is placed on the market or put into service, and to be kept up to date. It must be drawn up so as to demonstrate compliance with the requirements of that Section and to give national competent authorities and notified bodies the necessary information, in a clear and comprehensive form, to assess that compliance; it contains at a minimum the elements set out in Annex IV. Annex IV lists those elements as applicable to the AI system concerned, and asks among other things for the methods and steps performed for development, including, where relevant, recourse to pre-trained systems or tools provided by third parties and how those were used, integrated or modified by the provider, for the validation and testing procedures used, and for a description of relevant changes made by the provider to the system through its lifecycle. SMEs, including start-ups, may supply those elements in a simplified manner; where they take that route, they must use the simplified form the Commission is to establish for that purpose.
Article 11 puts the documentation duty on the provider, so the first open question here is who that provider is. If the creditworthiness assessment of consumers counts as high-risk in your situation, then in our reading there is a strong case that a webshop assembling the system itself and putting it into service under its own name is no longer merely a deployer but ends up on the provider side, carrying the documentation duty that comes with it. Article 11 does not settle that allocation of roles, so test it against the definitions and against Article 25 before assuming the file is your vendor's problem. In our reading the hard part in retail is not the first version of the file but the pace afterwards, because a webshop often ships changes per release while the file stays frozen at the state it had when the feature went into service. So contract for the provenance, training and testing information that Annex IV expects from you when you buy the pre-trained model, and assign per release who updates the file.
Editorial example. The rule above is in the Regulation. The situation was written by us to show how that rule plays out in this sector, and is not taken from a worked case in official guidance.
Article 11(1) and Annex IV

ISO/IEC 5259 series: data quality for analytics and machine learning
The ISO/IEC 5259 series (Artificial intelligence: Data quality for analytics and machine learning) comprises five parts: part 1 (overview, terminology and examples), part 2 (data quality measures), part 3 (data quality management requirements and guidelines) and part 4 (data quality process framework), all published in 2024, plus part 5 (data quality governance framework), published in February 2025. CEN-CENELEC has adopted parts as European standards, including EN ISO/IEC 5259-4:2025 and EN ISO/IEC 5259-3:2025. No part is cited in the Official Journal, so no presumption of conformity under Article 40 arises. The deliverable intended to do so for Article 10 is prEN 18284.
prEN 18284: quality and governance of datasets in AI
prEN 18284 (Artificial intelligence: Quality and governance of datasets in AI) is the JTC 21 deliverable under M/613 for Article 10 of the AI Act, which sets requirements for the training, validation and testing datasets of high-risk AI systems. As at June 2026 the deliverable was at the drafting stage. It has not yet been published as an EN and is not cited in the Official Journal.
EN 18286:2026: quality management system for EU AI Act regulatory purposes
EN 18286:2026 (Artificial intelligence: Quality management system for EU AI Act regulatory purposes) was drafted by CEN/CLC/JTC 21 and approved by CEN-CENELEC on 12 July 2026. By Implementing Decision (EU) 2026/2274 of 12 October 2026 the Commission published the reference of the standard in the Official Journal of 14 October 2026; the decision entered into force that day. EN 18286:2026 is thereby the first harmonised standard under Regulation (EU) 2024/1689, and since 14 October 2026 compliance with it confers under Article 40(1) a presumption of conformity with the requirements the standard covers. The decision does not list those requirements and attaches no restriction to the reference. According to a published coverage statement accompanying the standard, not yet confirmed by a second independent source, it addresses Article 17(1) points (a) to (m) and Article 11(1) first sentence, and expressly not Article 17(2) to (4) or Article 72.
EN ISO/IEC 42001: artificial intelligence management system
ISO/IEC 42001:2023 is the first certifiable international standard for an AI management system, published on 18 December 2023 and structured on the plan-do-check-act cycle. The text was adopted unchanged as EN ISO/IEC 42001:2026, approved by CEN on 13 March 2026, with national implementation by the member standards bodies. This adoption is not a deliverable under standardisation request M/613: the standard is not cited in the Official Journal and therefore confers no presumption of conformity under Article 40. For Article 17, the designated deliverable under M/613 is EN 18286:2026; its reference has been in the Official Journal since 14 October 2026 (Implementing Decision (EU) 2026/2274), so applying that standard does confer a presumption of conformity for the requirements it covers.

Read the official sourceChecked on 8 August 2026

General interpretation, not legal advice. Checked against Regulation (EU) 2024/1689 and the Digital Omnibus (EU) 2026/1744; the official source remains authoritative.

Full map for your situation

Your route

From your situation, via your role, to what applies and when it starts.

2 now · 15 later

Your situation

Applies now

From 2 December 2027

To record: Data governance file · Technical file (Annex IV) · Logs and retention regime

Source and locator

Every statement above rests on these texts. The locator points to the place in the text, the version and the date say which edition was checked.

EU Artificial Intelligence Act 2024/1689
European Parliament and Council, version original-oj-2024-07-12, checked on 8 August 2026, ELI http://data.europa.eu/eli/reg/2024/1689/oj
Locators in this source
- Article 10(1)-(6)
- Article 11(1)-(3) and Annex IV
- Article 12, Article 19 and Article 26(6)
- Article 13(1)-(3)
- Article 14(1)-(5)
- Article 15(1)-(5)
- Article 16(a)-(l)
- Article 17(1)-(3)
- Article 18(1)-(3)
- Article 20(1)-(2)
- Article 20(2), Article 73(1)-(2) and Article 79(1)
- Article 21(1)-(3)
- Article 12(1); Article 19(1); Article 21(1)-(3); Article 22(3); Article 26(6); Article 74(1); Article 78; Article 99(5)
- Article 26(1)-(12)
- Article 63(1) and (2)
- Article 17(1); Article 62(2); Article 113, second paragraph
- Article 8(1)-(2)
- Article 6(1); Article 9(1) and (2); Article 16, point (a); Article 40(1); Article 43(3) and (4); Annex I, Section A
- Article 9(1)-(10)
Digital Omnibus on AI 2026/1744
European Parliament and Council, version official-journal-2026-07-24, checked on 8 August 2026, ELI http://data.europa.eu/eli/reg/2026/1744/oj
Locators in this source
- Amended Article 113 application dates
- Article 1, point (40)(b), of Regulation (EU) 2026/1744, replacing Article 113, third paragraph, point (c), of Regulation (EU) 2024/1689
- Article 1, point (34), amending Article 77; Article 1, point (38)(b), inserting point (da) into Article 99(4)
- Article 1, point (40)(b), replacing Article 113, third paragraph, point (c)
- Article 1, point (39)(a), replacing Article 111(2)
- Article 1, point (2)(a), replacing Article 2(2); Article 1, point (41), amending Annex I
- Article 1, point 6, inserting Article 4a: Article 4a(1), points (a) to (f), and Article 4a(2), points (a) and (b)
- Article 1, point 6 (insertion) and point 9 (Article 10 amended, paragraph 5 deleted)
- Article 1, point 2(b), replacing Article 2(7)
- Recital 9, Article 4 (entry into force) and Article 1, point 40(a), replacing Article 113, third paragraph, point (a)
- Regulation (EU) 2026/1744, Article 1, point (31), Article 75(1a) and (1e)

Dataset praxikon:sys:registry:dataset:ai-act-implementation-graph 2.5.0, schema 1.5.0.

Have it done

Embed this obligation in policy, process and ownership

A legal action only works when it is clear who performs it, where the evidence sits and when it is reviewed. As your external DPO with AI oversight I assign your AI obligations, keep the register up to date and report on where you stand every quarter. The source interpretation above is separate from this offer: the explanation is the same for everyone, whether you buy anything or not.

External DPO and AI oversight · Fixed monthly fee · matched to your organisation, agreed up front · 12 months, then cancellable per quarter

See external DPO and AI oversight

Carried out by Zahed Ashkara, the legal adviser behind this explanation.

Does this answer your question?

Behind this page

Zahed Ashkara

Freelance AI and privacy legal adviser · AI governance & safety

View profile LinkedIn

When do the AI Act obligations start to apply to high-risk AI systems?

This applies now

Coming up

When it starts to apply

How this looks in practice4 practice examples

What the Commission says1 guidance statement

Standards that apply here4 standards

Follow-up questions

Applies now

From 2 December 2027

Source and locator

Embed this obligation in policy, process and ownership

Zahed Ashkara

Related questions