Skip to main content

Omnibus grandfathering: when a significant design change ends it

Which design changes count as significant under Art 111(2) of the AI Act: the texts, the open points and a documented-decision method for each release.

By Scanara

Key Takeaways

  1. 1.Who and what: Earlier high-risk systems (our label) are high-risk AI systems, other than those referred to in Art 111(1). They were placed on the market or put into service before the date of application of Chapter III referred to in Art 113. The Act does not say which of Chapter III's dates that is. Art 111(2) applies the Act to operators of earlier systems only if, as from that date, their designs are subject to significant changes. Art 3(8) defines an operator as “a provider, product manufacturer, deployer, authorised representative, importer or distributor”.
  2. 2.When: Chapter III, Sections 1, 2 and 3, sets the high-risk obligations. They start on 2 December 2027 for one group of high-risk systems and 2 August 2028 for another (Art 113). Which of Chapter III's dates Art 111(2) means is not stated in the Act. Providers and deployers of high-risk AI systems intended to be used by public authorities must take the necessary steps to comply by 2 August 2030. This applies in any case, regardless of design changes (Art 111(2)).
  3. 3.Why it matters: A significant change to the design after the Art 111(2) date should trigger the obligation of the provider to comply fully with the high-risk rules. Source: recital (39) of Regulation (EU) 2026/1744.
  4. 4.Next step: For every release, record what changed in the design and who decided whether the change is significant (our suggestion).

Regulation (EU) 2026/1744 (the Digital Omnibus on AI, called the Omnibus here) amended Art 111(2) of the AI Act, Regulation (EU) 2024/1689. Article numbers refer to the AI Act as amended by the Omnibus. Whether an engineering change is significant is open. Only the Court of Justice of the European Union can give an authoritative interpretation (Commission draft guidelines, para (6)).

Art 111(2) applies the AI Act to earlier systems only if their designs undergo significant changes

In Art 111(2), the Act applies to earlier systems “only if” those systems “are subject to significant changes in their designs”. Earlier means placed on the market or put into service before a date that Art 111(2) ties to Art 113. Placing on the market means first making an AI system available on the Union market (Art 3(9)). Putting into service means supplying an AI system for first use (Art 3(11)). Art 111(2) excludes AI systems that are components of large-scale IT systems under Art 111(1) and Annex X. Art 5, the prohibited practices, applies regardless (Art 111(2), first clause).

The Omnibus replaced the fixed date in Art 111(2) with a reference to Art 113

The Omnibus replaced the fixed date “2 August 2026” in Art 111(2) with a reference to Art 113. Amended Art 111(2) refers to “the date of application of Chapter III referred to in Article 113”. Art 113 sets different dates for different parts of Chapter III. Section 4 of Chapter III applies from 2 August 2025 (Art 113, third paragraph, point (b)). Art 113, third paragraph, point (c), sets two dates for Chapter III, Sections 1, 2 and 3 (except Art 6(5)). The first is 2 December 2027 for “AI systems classified as high-risk pursuant to Article 6(2) and Annex III”. Art 6(2) adds the AI systems that Annex III lists, which “shall be considered to be high-risk”. Annex III lists them by area; its first area is biometrics. Art 6(3) derogates from Art 6(2) where an Annex III system “does not pose a significant risk of harm to the health, safety or fundamental rights of natural persons”. The derogation applies where “any of the following conditions is fulfilled”, and the Act lists four. That derogation does not apply to an Annex III system that “performs profiling of natural persons”, which “shall always be considered to be high-risk” (Art 6(3)). The second is 2 August 2028 for “AI systems classified as high-risk pursuant to Article 6(1) and Annex I”. Art 6(1) covers AI systems that are, or are safety components of, products under the Union harmonisation legislation listed in Annex I. The product must also require third-party conformity assessment (Art 6(1)(b)). The Omnibus added Art 6(1a) to (1c). AI systems used solely for non-safety related aspects, such as user assistance or convenience, do not qualify as safety components (Art 6(1a)). AI systems whose failure or malfunctioning would endanger health and safety qualify as safety components (Art 6(1b)). This applies even where Art 6(1a) would otherwise exclude them. A product that needs third-party conformity assessment solely because of risks other than health and safety does not meet Art 6(1), point (b) (Art 6(1c)). Other parts of Chapter III, which neither point (b) nor point (c) names, follow the general date in Art 113, second paragraph: “It shall apply from 2 August 2026”. Recital (40) of the Omnibus gives the same two dates for Sections 1, 2 and 3. The Act does not state in terms which of Chapter III's dates Art 111(2) means. This article uses the dates that Art 113, third paragraph, point (c), sets for Sections 1, 2 and 3.

Recital (39) of the Omnibus ties the Art 111(2) rule to the first unit and to an unchanged design

Recital (39) makes the first unit's date decisive and keeps an unchanged design under the Art 111(2) rule. The first unit is the first one of its type and model placed on the Union market or put into service. At least one unit must be lawfully placed on the market or put into service before the date specified in Art 111(2). Other units of the same type and model are then subject to the same rule. That holds “as long as the design of that high-risk AI system remains unchanged”. Neither recital (39) nor Art 111(2) says how a unit maps to a software release. The recital calls the Art 111(2) rule “the grace period provided in Article 111(2)”.

Recital (177) of the AI Act treats a significant change as equivalent in substance to a substantial modification, which Art 3(23) defines

Recital (177) says the concept of significant change “should be understood as equivalent in substance to the notion of substantial modification”. Art 3(23) defines a substantial modification as a change after placing on the market or putting into service, subject to two further conditions. The change must be “not foreseen or planned in the initial conformity assessment carried out by the provider”. It must also affect compliance with “the requirements set out in Chapter III, Section 2” or modify the intended purpose. The Omnibus amends Art 3 only in points (14), (14a) and (14b), so Art 3(23) keeps its original wording. The recital's first sentence speaks of “significant changes in their design or intended purpose”. Art 111(2) says “significant changes in their designs”. Neither text says whether this difference matters. The AI Act Service Desk page for recital (177) has the same wording as the original.

Recital (128) of the AI Act gives “change of operating system or software architecture” as an example

Recital (128) gives “change of operating system or software architecture” as an example of a change that may affect compliance. For such a change, the recital says the AI system “should be considered to be a new AI system which should undergo a new conformity assessment”. The recital concerns substantial modification, not Art 111(2) directly. Use the example as a reference point for comparing your own changes.

Art 43(4) excludes pre-determined changes in learning systems; whether it reaches Art 111(2) is open

Art 43(4) excludes pre-determined changes to learning systems from substantial modification. It covers high-risk AI systems that “continue to learn after being placed on the market or put into service”. The provider must have pre-determined the changes at “the moment of the initial conformity assessment”. The changes must also be part of the technical documentation referred to in point 2(f) of Annex IV. Neither Art 111(2) nor recital (39) mentions Art 43 or learning systems, so whether Art 43(4) carries over is open. The Omnibus replaced only paragraph 3 of Art 43 (point (19)), so Art 43(4) keeps its original wording.

Neither Regulation defines a “significant change” as such, and the two draft guideline documents we read give no criterion

We searched the English text of the Omnibus and the original 2024 text of the AI Act, including recitals, for “significant change” and “design change”. “Significant change” appears only in Art 111(2), recital (39) and recital (177). On 19 May 2026 the Commission published draft guidelines on the classification of high-risk AI systems for stakeholder consultation. The draft has three documents: a general-principles document, an Annex I chapter and an Annex III chapter. We did not read the draft's Annex III chapter. The consultation page, updated on 16 June 2026, gave 23 July 2026 as the consultation deadline. It said the final guidelines were to be adopted by the end of 2026. Para (6) of the draft says the guidelines are not binding. In the draft's general-principles document, the only passage on Art 111(2) is para (449). Para (449) restates the wording of Art 111(2) from before the Omnibus, including the date 2 August 2026. It gives no criterion for “significant changes”. The draft's Annex I chapter, read in full (13 pages), has no text on substantial modification, Art 111 or “significant change”. The AI Office page, published on 4 December 2025 and updated on 31 July 2026, lists planned guidelines. One entry reads “Guidelines on the practical application of the provisions related to substantial modification”, with no date. Art 96(1)(c) obliges the Commission to develop guidelines on “the practical implementation of the provisions related to substantial modification”. Neither Regulation defines “significant change” itself. Recital (177) equates it in substance with “substantial modification” (Art 3(23)). Neither Regulation, nor the two draft documents we read, says which engineering changes meet the Art 3(23) conditions for Art 111(2). Recital (128) gives one example for substantial modification. Each operator has to make its own decision.

Keep one record per release; this is our suggestion, not a rule of the Act

FieldWhat to record
System and versionName and release identifier
Chapter III date used (Art 113)2 December 2027 (Art 6(2) and Annex III), 2 August 2028 (Art 6(1) and Annex I) or another Chapter III date
Design changeWhat changed: code, model, architecture or training data pipeline (our categories)
Reference releaseThe release you treat as the first unit placed on the market, and why (recital (39) speaks of units, not releases)
Intended purposeChanged or unchanged, with the evidence
Requirement areaWhich requirements in Chapter III, Section 2 the change touches
Art 43(4) checkFor a learning system: whether the change was pre-determined and documented (open for Art 111(2))
DecisionSignificant, not significant or undecided
Decided by, dateName, role and date
ReasoningTwo sentences
Review triggerThe event that reopens the decision, such as the next release or a new guideline

An illustrative record shows how the fields are filled, with the decision left open

The record below is not an example from the Act, and its decision field stays open.

FieldIllustrative entry
System and versionExample system A, release 2.2
Chapter III date used (Art 113)2 December 2027, assuming classification under Art 6(2) and Annex III
Design changeInference service: model file replaced by a retrained model
Reference releaseRelease 1.0, the first placed on the market
Intended purposeUnchanged, per release notes
Review triggerNext release

The team fills the other fields: requirement area, Art 43(4) check, decision, decided by and date, and reasoning.

Scanara gives the record dated scan evidence, and you make the decision

Scanara structures the evidence for the record; the team decides, with counsel where stakes warrant. Scanara scans repositories on demand with rule-based checks and detects standard ML framework patterns. Scanara can miss non-standard or obfuscated code, and confidence varies by language. A re-run after a release gives a dated result that you can attach to the record. Scanara generates drafts of Annex IV technical documentation, which you review and complete. A passing scan does not mean an organisation is compliant.

Related Reading

Sources & References

Frequently Asked Questions

Does a change of intended purpose matter under Art 111(2)?

Open. Art 111(2) names “designs” only, while recital (177) names “design or intended purpose”. Art 3(23) lists a modification to the intended purpose as one way a change becomes a substantial modification. The sources we used leave open which wording governs; the record notes any intended-purpose change.

Is “grandfathering” a term of the Act?

No. A search of the English text of the Omnibus and the original 2024 text of the AI Act, including recitals, for “grandfather” found no hit. We use the word as a label for the Art 111(2) rule.

Scanara prepares evidence from your repositories

Scanara reads your connected GitHub repositories and runs rule-based scans for standard AI and machine-learning patterns. It can miss non-standard code. Separately, it generates draft documentation from templates and your assessment answers as evidence for your compliance work. You make the classification decisions.