A municipality has used a welfare risk model since 2021, a bank its own credit scoring model for years, both within Annex III, point 5. Must they fully comply with the high-risk rules on 2 December 2027?
The answer is in Article 111, read here paragraph by paragraph in the text that applies since Regulation (EU) 2026/1744, the Digital Omnibus, entered into force on 27 July 2026. The examples are fictional.
In short
A high-risk AI system placed on the market or put into service before the date of application of Chapter III is caught by the high-risk rules only once its design is subject to significant changes; systems intended to be used by public authorities must comply by 2 August 2030 in any case (paragraph 2). Further: large-scale EU IT systems by 31 December 2030 (paragraph 1), older general-purpose AI models by 2 August 2027 (paragraph 3), marking by 2 December 2026 (paragraph 4). Article 5 applies in full, and Articles 4 and 50 also apply to existing systems.
Separate posts cover the Omnibus postponement, the scope of the AI Act, the obligations of model providers, when you become a provider and the fundamental rights impact assessment.
Paragraph 1: the large-scale EU IT systems
Annex X lists the Union legislation on large-scale IT systems in the area of freedom, security and justice, such as the Schengen Information System, the Visa Information System, Eurodac and the Entry/Exit System. AI systems which are components of those systems and have been placed on the market or put into service before 2 August 2027 shall be brought into compliance with the Regulation by 31 December 2030. The requirements are also taken into account in each evaluation of such a system provided for in its own legislation, and where that legislation is replaced or amended.
Article 5 applies here too. A component placed on the market or put into service after 2 August 2027 is not covered by paragraph 1. In our reading it falls under paragraph 2 if that happens before the date of application of Chapter III, with the 2 August 2030 deadline because it is intended for public authorities, and otherwise under the ordinary regime.
Paragraph 2: existing high-risk AI systems
Since Article 1, point 39, of Regulation (EU) 2026/1744, the Regulation applies to operators of high-risk AI systems placed on the market or put into service before the date of application of Chapter III "only if, as from that date, those systems are subject to significant changes in their designs". The original text referred to 2 August 2026. Operators include providers and deployers (Article 3, point 8).
Which date? Article 113, third paragraph, point (c), sets two dates for Sections 1, 2 and 3 of Chapter III: 2 December 2027 for AI systems classified as high-risk under Article 6(2) and Annex III, and 2 August 2028 for those under Article 6(1) and Annex I. Section 4 applies since 2 August 2025 and Section 5 (including conformity assessment and registration) since 2 August 2026. In our reading point (c) still counts: it covers the requirements and obligations the transitional rule is meant for, and an earlier date would undo the Omnibus postponement for existing systems. Recital 39 of Regulation (EU) 2026/1744 supports this with "the relevant provisions".
Type and model. According to the same recital, what counts is the type and model: if one individual unit was lawfully placed on the market or put into service before that date, other units of the same type and model also benefit from the grace period, as long as the design remains unchanged.
What is a significant change? Recital 177 says the concept "should be understood as equivalent in substance to the notion of substantial modification", which Article 3, point 23, defines as a change not foreseen or planned in the provider's initial conformity assessment that affects compliance with the requirements of Chapter III, Section 2, or modifies the intended purpose. Paragraph 2 only mentions the design, the recital design or intended purpose.
An existing system never had a conformity assessment under the AI Act. In our reading what then counts is whether the change affects the requirements of Section 2, such as data, accuracy or human oversight, or the intended purpose. Article 96(1)(c) requires the Commission to develop guidelines on this; the Commission says the AI Office is still working on them.
After a significant change the provider must fully comply, including the conformity assessment requirements (recital 39 of Regulation (EU) 2026/1744). Each further substantial modification then requires a new conformity assessment (Article 43(4)). A deployer, distributor, importer or other third party that makes a substantial modification such that the system remains high-risk becomes a provider (Article 25(1)(b)).
Public authorities. Providers and deployers of high-risk AI systems "intended to be used by public authorities" shall take the necessary steps to comply with the requirements and obligations by 2 August 2030, even if nothing changes. In our reading this includes the fundamental rights impact assessment (Article 27) and registration (Article 49(3) and Article 26(8)). A private provider of public services is covered by Article 27, but in our reading not by this deadline if the system is not intended for use by a public authority.
Paragraph 3: general-purpose AI models
Chapter V applies since 2 August 2025 (Article 113, third paragraph, point (b)). Providers of models placed on the market before that date shall take the necessary steps to comply by 2 August 2027 with, among others, Article 53, Article 54 and, for models with systemic risk, Article 55.
The Commission's guidelines of 19 November 2025, C(2025) 7719 final, apply the period to the entire lifecycle of such models (point 106), which begins at the large pre-training run (point 22); subsequent development by or on behalf of the provider forms part of the same model. In our reading a model from a new pre-training run therefore no longer falls under paragraph 3. Under point 108 these providers need not retrain or unlearn models where this is impossible for past actions, training data information is unavailable or retrieval would be disproportionate, provided they clearly disclose and justify this in the copyright policy and the training content summary. Point 51 gives examples of placing on the market, such as via an API or a public repository.
Paragraph 4: marking of generated content
Article 50(2) requires providers of AI systems generating synthetic audio, image, video or text content to mark the outputs in a machine-readable format and make them detectable as artificially generated or manipulated. Article 50 applies since 2 August 2026. For systems placed on the market before that date, the new paragraph 4 allows until 2 December 2026 for Article 50(2); recital 38 of Regulation (EU) 2026/1744 speaks of four months. The other paragraphs of Article 50 get no extra time. Paragraph 4 only names systems placed on the market; the Article 50 guidelines (point 153), more broadly, also name systems put into service.
What Article 111 does not defer
The prohibitions of Article 5 apply since 2 February 2025 to every system, old or new (Article 113, third paragraph, point (a)). The prohibitions added by the Omnibus, Article 5(1), first subparagraph, points (ba) and (bb), with their delimitation in Article 5(1a) and (1b), apply from 2 December 2026. Article 4 on AI literacy also applies since 2 February 2025, Article 50 since 2 August 2026.
Paragraph 2 does not affect Article 50, even though the text says that "this Regulation" applies only after a significant change. In its guidelines on Article 50 (point 153, footnote 50) the Commission states that the grandfathering rule in paragraph 2 is limited to the high-risk requirements and obligations and that Article 50 applies cumulatively. In our reading the same holds for Article 4: Recital 177 concerns the high-risk requirements.
If an existing system processes personal data, the obligations of the GDPR apply now. Where processing is likely to result in a high risk, a data protection impact assessment had to be carried out before processing started (Article 35 GDPR(1)); this is in particular the case for profiling on which decisions with legal effects are based (paragraph 3(a)). A missing one is a shortcoming now. Where necessary a review follows, at least when the risk changes (paragraph 11).
Example: a municipality with a risk model
In 2021 a municipality bought a supplier's model that gives social assistance applications a risk score. If it is an AI system, it is in our reading high-risk under Annex III, point 5(a). A derogation under Article 6(3) does not help here: a risk score per applicant is profiling, and an Annex III system that performs profiling is always high-risk (Article 6(3), last subparagraph).
Now. The prohibition of social scoring (Article 5(1)(c)) already applies, as do Article 4 and the GDPR.
On 2 December 2027. Without a design change, the high-risk rules do not yet apply. For this municipality 2 December 2027 is therefore not automatically the date for the fundamental rights impact assessment: Article 111(2) moves it to a significant change in the design, and at the latest to 2 August 2030.
With a new version. In 2028 the supplier delivers a version with new variables and a retrained model. In our reading that is a significant change, because the new data affect Section 2 requirements such as Article 10. The supplier must then fully comply as provider. In our reading the municipality, as deployer, carries out the Article 27 assessment before first use of the new version, registers the use (Article 49(3)) and does not use an unregistered system (Article 26(8)). If another municipality buys the same, unchanged model in 2028, that unit benefits from the grace period under recital 39 of Regulation (EU) 2026/1744.
By 2 August 2030. Even without a new version, both must then comply: systems under Annex III, point 5(a), are by definition intended to be used by or on behalf of public authorities. Recital 178 encourages providers to start to comply voluntarily during the transitional period.
Example: a bank with a credit scoring model
A bank has used a self-built model since 2023 to evaluate the creditworthiness of natural persons; that too is profiling, so high-risk under Annex III, point 5(b). Having built it itself, the bank is both provider and deployer (Article 3, points 3 and 4). No public-authority deadline applies. As long as the design does not change significantly, the high-risk rules and the Article 27 assessment do not apply.
Periodic retraining is harder. In our reading retraining on new data in the same way for the same purpose is not in itself a significant change, but new variables, a different model type or a new purpose, such as pricing, probably would be. The line is not yet settled; record why each change is or is not significant. An insurer with a model for life or health insurance (point 5(c)) is in the same position.
Frequently asked questions about Article 111 AI Act
Sources
Newsletter
Every Tuesday, the AI Act week ahead in 5 minutes
A practical briefing on deadlines, new guidance and enforcement, so you know what matters this week. No spam and you can unsubscribe in one click.
Practical and short · No spam · One-click unsubscribe