EU AI Act Article 111: Grandfathering Legacy Systems, Implemented in Practice
EU AI Act Article 111 decides which AI systems and GPAI models that were already running before the application dates are grandfathered, and when they lose that status. We write the date of putting into service and a computed design-change flag into every registry record: when weights, architecture or intended purpose change, the protection falls away automatically, and every legacy system gets a migration plan with a date.
Almost every company runs AI systems that are older than the EU AI Act. A scoring model from 2019, a chatbot from 2023, an assistant on a language model from 2025. The first question in every project: do we have to retrofit all of it, and by when? Article 111 gives the answer, and it is more nuanced than most people hope. Depending on the paragraph, it addresses the operators of high-risk systems, meaning providers as well as deployers, the providers of GPAI models and the providers of generative systems. The article belongs to no phase: it has applied since entry into force (Art. 113) and governs the transitions between the phases.
The EU AI Act has been in force since 1 August 2024. Prohibitions and AI literacy have applied since February 2025, the obligations for GPAI models since August 2025, the transparency obligations under Article 50 since August 2026. After the postponement by the amending Regulation (EU) 2026/1744 (Digital Omnibus), the obligations for high-risk systems under Annex III apply from 2 December 2027, for Annex I products from 2 August 2028. The Omnibus moved deadlines, it removed no obligation, and it amended Article 111 itself, more on that below. The overview is in EU AI Act in practice.
What Article 111 requires
Art. 111(1): AI systems that are components of the large-scale IT systems under Annex X (such as the Schengen Information System, Eurodac, ETIAS) and were placed on the market or put into service before 2 August 2027 must be brought into compliance by 31 December 2030. Article 5 is unaffected. The paragraph concerns Union agencies and authorities, not companies.
Art. 111(2), first sentence: For operators of high-risk systems placed on the market or put into service before the date of application of Chapter III, the Regulation applies only if the system undergoes a significant change in its design from that date. Article 5 is excepted: the prohibitions apply to legacy systems without transition.
Art. 111(2), second sentence: Providers and deployers of high-risk systems intended to be used by public authorities must in any case comply by 2 August 2030, even without a design change.
Art. 111(3): Providers of GPAI models placed on the market before 2 August 2025 must meet the obligations of Chapter V (Articles 53 to 55) by 2 August 2027. The addressee is the model provider, not the company that uses the model.
Art. 111(4) (new, added by the Omnibus): Providers of generative systems for audio, image, video or text placed on the market before 2 August 2026 must implement the marking under Art. 50(2) by 2 December 2026. For the other obligations of Article 50, such as telling people they are talking to an AI, there is no transition.
What “significant change in design” means
Recital 177 declares the term equivalent to the “substantial modification” of Art. 3(23): a change after placing on the market that was not foreseen in the initial conformity assessment and that either affects compliance with Chapter III, Section 2 or changes the intended purpose. So a new intended purpose is a design change even if nobody touched a byte of the model. For legacy systems, “foreseen” is hard to prove, because there was no conformity assessment in 2019. Whether routine retraining counts has not been decided; I consider it the riskier reading to treat it as no change.
The Omnibus and the Annex III date
Before the Omnibus, paragraph 2 named 2 August 2026 as the cut-off. The consolidated version, as the AI Act Explorer shows it, refers instead to the date of application of Chapter III under Article 113, which is 2 December 2027 for Annex III systems and 2 August 2028 for Annex I products. My reading: an Annex III system that goes live by 1 December 2027 is grandfathered; one that starts later must comply from day one. The wording, however, speaks of one date, although Article 113 names two for Chapter III. That each system takes the date of its own category is an interpretation, not a certainty.
The technical example
An insurer asks us about three systems. Which are protected, and until when?
The scoring model from 2019: risk assessment and pricing in health insurance, built in-house, put into service in 2019, retrained every year since. That is Annex III, point 5(c), so high-risk. Grandfathered under Art. 111(2) as long as no significant design change happens from 2 December 2027. The annual retraining is exactly the question: if it counts as a design change, protection ends with the first training run after the cut-off. Article 5, the GDPR and the supervisory requirements for models apply regardless.
The chatbot from 2023: customer service, answers contract questions, decides nothing. Not high-risk, so not a case for paragraph 2. Article 50 has applied to it since August 2026 without transition; only the marking under Art. 50(2), if it falls on the provider of the chatbot, has until 2 December 2026. In practice nothing is protected.
The GPAI-based assistant from 2025: summarises claims files for handlers and runs on a hosted model from an external provider that was released before 2 August 2025. Art. 111(3) protects the model provider until 2 August 2027, not the insurer. The assistant is not a high-risk system today. If it were, its grandfathering would be worth little: hosted models are replaced by successors within months, and every new model version changes the weights and architecture of the system running on it.
The third case shows why “we never touched it” is not a strategy. The sentence is true of your own team, not of the system. When the provider retires the model version, the design changes whether the insurer wants it or not, and it stands there without grandfathering, without Article 11 documentation and without time. The question is not whether a legacy system falls into scope, but whether that happens on a date you chose yourself.
How we implement it
Every registry record in our governance layer gets three fields for Article 111. The date of placing on the market or putting into service, always with evidence: go-live protocol, contract or deployment ticket, hashed into the record. The legal basis of the transition: paragraph 2, paragraph 3 at the model provider, paragraph 4, or none. And the end of the grace period, computed: a fixed date for public-authority systems and for the model provider, “until design change” for everything else. From the provider dossier under Article 53 we add the end of support for the model version.
The gateway computes the design-change flag from the version diff every time a record is updated. It compares the hash of the weights or the provider’s model version, the architecture descriptor (model family, parameter count, modalities, context window), the text of the intended purpose, and the scope with its allowed actions. If any of them changes, the flag flips with a timestamp, the status moves from “legacy” to “in scope”, and the system loses its route until it has passed the gates for high-risk systems: the business unit’s evaluation set, regression tests, bias check, Article 11 documentation, Article 49 registration, three-way sign-off. No registration, no route also applies to systems that were protected until yesterday.
Two decisions are not dictated by the law. For us, prompt changes do not count as design changes as long as the intended purpose stays the same; they go through the regression tests. Retraining counts as a design change unless the registry record fixed the training cycle in advance, with data source, metric and acceptance criterion, by analogy with the pre-determined changes of Art. 43(4). That analogy is our choice, not the text of the law, and that is how the registry records it.
Every legacy system also gets a migration plan: owner, gap list against Articles 9 to 15, 17 and 26, target date, and one of three decisions: migrate, freeze or retire. The target date sits before the earliest of three events: end of the grace period, end of provider support, next planned training run. A frozen system is no longer trained, but keeps writing to the evidence log. Per legacy system the auditor sees the evidenced date, the version history with diffs, the history of the flag and the signed plan.
Typical mistakes
Grandfathering assumed across the board. “It has been running since 2019” without evidence and without a training log. Then the date is a claim, and whether the design changed is for the auditor to decide.
Paragraph 3 misread. The company believes its assistant is protected until August 2027. The model provider is protected.
Chatbots placed under grandfathering. Article 50 has no transition except the marking under paragraph 2, and that one ends on 2 December 2026.
No plan for end of support. The provider retires the model version, and the company has weeks to catch up on what had years.
EU AI Act Article 111 checklist
Inventory of all AI systems with the date of placing on the market or putting into service and evidence in the registry.
Every legacy system classified: Annex X, high-risk under Annex I or III, public-authority use, generative, GPAI model.
For every high-risk legacy system, the legal basis of grandfathering and its end date or end condition in the registry; public-authority systems scheduled against 2 August 2030.
Design-change flag defined, computed from the version diff and stored with history.
Articles 5 and 50 applied to all legacy systems, regardless of grandfathering.
Provider dossier with model version and end of support for every system on a GPAI model.
Migration plan per legacy system with owner, gaps, date and decision, signed off three ways.
Open questions
Paragraph 2 speaks of one date of application, Article 113 names two for Chapter III. Which date applies to which system is my interpretation; Commission guidance is missing.
Whether routine retraining is a significant change has been settled by neither standard nor case law. The research question behind it: which metric, weight distance or behavioural diff on the evaluation set, makes “significant” measurable?
Whether public-law bodies such as statutory health insurers are “public authorities” within paragraph 2 is open. It decides whether the fixed August 2030 date applies to them.
Frequently asked questions
Does Article 111 also apply to us as a deployer?
Yes. Paragraph 2 speaks of operators, which includes deployers. Your legacy system stays outside the Regulation until it undergoes a significant design change, Article 5 excepted. But you must be able to prove the date, and the deployer obligations under Article 26 bite on the day the protection ends. How that belongs in the registry is described under AI governance.
Do we lose grandfathering when the provider updates the model?
On my reading, yes, as soon as the update changes weights or architecture. That you did not initiate it changes nothing. That is why the provider’s end of support belongs in the migration plan.
What happens when grandfathering ends?
Then the Regulation applies in full and with no further transition: provider obligations under Article 16, conformity assessment under Article 43, registration under Article 49, deployer obligations under Article 26. Whoever chose the date themselves already has the documentation by then. The other articles in the series: all articles at a glance.
EU AI Act Article 111: Grandfathering Legacy Systems, Implemented in Practice
EU AI Act Article 111 decides which AI systems and GPAI models that were already running before the application dates are grandfathered, and when they lose that status. We write the date of putting into service and a computed design-change flag into every registry record: when weights, architecture or intended purpose change, the protection falls away automatically, and every legacy system gets a migration plan with a date.
Almost every company runs AI systems that are older than the EU AI Act. A scoring model from 2019, a chatbot from 2023, an assistant on a language model from 2025. The first question in every project: do we have to retrofit all of it, and by when? Article 111 gives the answer, and it is more nuanced than most people hope. Depending on the paragraph, it addresses the operators of high-risk systems, meaning providers as well as deployers, the providers of GPAI models and the providers of generative systems. The article belongs to no phase: it has applied since entry into force (Art. 113) and governs the transitions between the phases.
The EU AI Act has been in force since 1 August 2024. Prohibitions and AI literacy have applied since February 2025, the obligations for GPAI models since August 2025, the transparency obligations under Article 50 since August 2026. After the postponement by the amending Regulation (EU) 2026/1744 (Digital Omnibus), the obligations for high-risk systems under Annex III apply from 2 December 2027, for Annex I products from 2 August 2028. The Omnibus moved deadlines, it removed no obligation, and it amended Article 111 itself, more on that below. The overview is in EU AI Act in practice.
What Article 111 requires
What “significant change in design” means
Recital 177 declares the term equivalent to the “substantial modification” of Art. 3(23): a change after placing on the market that was not foreseen in the initial conformity assessment and that either affects compliance with Chapter III, Section 2 or changes the intended purpose. So a new intended purpose is a design change even if nobody touched a byte of the model. For legacy systems, “foreseen” is hard to prove, because there was no conformity assessment in 2019. Whether routine retraining counts has not been decided; I consider it the riskier reading to treat it as no change.
The Omnibus and the Annex III date
Before the Omnibus, paragraph 2 named 2 August 2026 as the cut-off. The consolidated version, as the AI Act Explorer shows it, refers instead to the date of application of Chapter III under Article 113, which is 2 December 2027 for Annex III systems and 2 August 2028 for Annex I products. My reading: an Annex III system that goes live by 1 December 2027 is grandfathered; one that starts later must comply from day one. The wording, however, speaks of one date, although Article 113 names two for Chapter III. That each system takes the date of its own category is an interpretation, not a certainty.
The technical example
An insurer asks us about three systems. Which are protected, and until when?
The third case shows why “we never touched it” is not a strategy. The sentence is true of your own team, not of the system. When the provider retires the model version, the design changes whether the insurer wants it or not, and it stands there without grandfathering, without Article 11 documentation and without time. The question is not whether a legacy system falls into scope, but whether that happens on a date you chose yourself.
How we implement it
Every registry record in our governance layer gets three fields for Article 111. The date of placing on the market or putting into service, always with evidence: go-live protocol, contract or deployment ticket, hashed into the record. The legal basis of the transition: paragraph 2, paragraph 3 at the model provider, paragraph 4, or none. And the end of the grace period, computed: a fixed date for public-authority systems and for the model provider, “until design change” for everything else. From the provider dossier under Article 53 we add the end of support for the model version.
The gateway computes the design-change flag from the version diff every time a record is updated. It compares the hash of the weights or the provider’s model version, the architecture descriptor (model family, parameter count, modalities, context window), the text of the intended purpose, and the scope with its allowed actions. If any of them changes, the flag flips with a timestamp, the status moves from “legacy” to “in scope”, and the system loses its route until it has passed the gates for high-risk systems: the business unit’s evaluation set, regression tests, bias check, Article 11 documentation, Article 49 registration, three-way sign-off. No registration, no route also applies to systems that were protected until yesterday.
Two decisions are not dictated by the law. For us, prompt changes do not count as design changes as long as the intended purpose stays the same; they go through the regression tests. Retraining counts as a design change unless the registry record fixed the training cycle in advance, with data source, metric and acceptance criterion, by analogy with the pre-determined changes of Art. 43(4). That analogy is our choice, not the text of the law, and that is how the registry records it.
Every legacy system also gets a migration plan: owner, gap list against Articles 9 to 15, 17 and 26, target date, and one of three decisions: migrate, freeze or retire. The target date sits before the earliest of three events: end of the grace period, end of provider support, next planned training run. A frozen system is no longer trained, but keeps writing to the evidence log. Per legacy system the auditor sees the evidenced date, the version history with diffs, the history of the flag and the signed plan.
Typical mistakes
EU AI Act Article 111 checklist
Open questions
Frequently asked questions
Does Article 111 also apply to us as a deployer?
Yes. Paragraph 2 speaks of operators, which includes deployers. Your legacy system stays outside the Regulation until it undergoes a significant design change, Article 5 excepted. But you must be able to prove the date, and the deployer obligations under Article 26 bite on the day the protection ends. How that belongs in the registry is described under AI governance.
Do we lose grandfathering when the provider updates the model?
On my reading, yes, as soon as the update changes weights or architecture. That you did not initiate it changes nothing. That is why the provider’s end of support belongs in the migration plan.
What happens when grandfathering ends?
Then the Regulation applies in full and with no further transition: provider obligations under Article 16, conformity assessment under Article 43, registration under Article 49, deployer obligations under Article 26. Whoever chose the date themselves already has the documentation by then. The other articles in the series: all articles at a glance.
Related: EU AI Act · All articles · AI Governance · AI Consultant
Recent Posts
Recent Comments