EU AI Act Article 113: Timeline and Deadlines, Implemented in Practice
EU AI Act Article 113 sets the date from which each obligation applies, in stages from February 2025 to August 2028. Our answer: every registry record carries a computed field obligations_apply_from, derived from the risk class and a versioned timeline, and that field feeds the countdown for the board.
Article 113 is the shortest article with the biggest effect on project plans. It contains no obligation for providers or deployers, it only says when the other articles start to bite. Since amending Regulation (EU) 2026/1744 it has been rewritten, and in projects I see the shift being read as an all-clear. It is not. I treat Article 113 together with Article 111, because the transition rules for existing systems cannot be read without it.
The EU AI Act has been in force since 1 August 2024. The 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. The obligations for high-risk systems under Annex III apply, after the shift by amending Regulation (EU) 2026/1744 (Digital Omnibus), from 2 December 2027, for Annex I products from 2 August 2028. Article 113 is the article that fixes exactly this timeline. The Omnibus moved dates; it removed no obligation.
What EU AI Act Article 113 requires
Three paragraphs: entry into force (para. 1), the general date of application (para. 2), exceptions in para. 3 points (a) to (d). Plus the transition rules in Article 111. The stages according to the consolidated text in the AI Act Explorer:
Date
What applies
Basis
1 August 2024
Entry into force, twenty days after publication in the Official Journal. No obligation applies yet.
Art. 113(1)
2 February 2025
Chapters I and II: scope, definitions, AI literacy (Art. 4), prohibited practices (Art. 5).
Art. 113(3)(a)
2 August 2025
Chapter III Section 4 (notified bodies), Chapter V (GPAI models), Chapter VII (governance), Chapter XII (penalties), Art. 78. Except Art. 101.
Art. 113(3)(b)
27 July 2026
Art. 102 to 110, the amendments to other Union acts (machinery, vehicles, aviation).
Art. 113(3)(d), new
2 August 2026
General date of application for everything points (a) to (d) do not list: Art. 50 (transparency), Art. 101 (fines for GPAI providers), Art. 6(5), on the wording also Chapter III Section 5 and Chapters VIII and IX (see below).
Art. 113(2)
2 December 2026
Art. 5(1) points (ba) and (bb) and Art. 5(1a) and (1b), the new prohibitions. Deadline for existing systems that generate synthetic content to comply with Art. 50(2).
Art. 113(3)(a); Art. 111(4), new
2 December 2027
Chapter III Sections 1 to 3 for high-risk systems under Art. 6(2) and Annex III: classification, requirements Art. 8 to 15, obligations Art. 16 to 27.
Art. 113(3)(c)(i), new
2 August 2028
The same for high-risk systems under Art. 6(1) and Annex I, products with third-party assessment under product law.
Art. 113(3)(c)(ii), new
Transition
GPAI models on the market before 2 August 2025: by 2 August 2027. High-risk systems used by public authorities: by 2 August 2030. Components of the large-scale IT systems in Annex X placed on the market before 2 August 2027: by 31 December 2030.
Art. 111(3), (2) and (1)
The Omnibus rewrote point (c): before, it only named Art. 6(1) with 2 August 2027, and Annex III systems fell under 2 August 2026. Now point (c) names Chapter III Sections 1 to 3 explicitly, without Art. 6(5), with two dates. Point (a) carves out the new prohibitions, point (d) is new. Art. 111(2) now refers to the date of application of Chapter III instead of a date, and para. 4 is new.
I have to be honest here. Point (c) shifts Sections 1 to 3 of Chapter III. Section 5 (Art. 40 to 49: standards, conformity assessment, declaration of conformity, CE marking, registration), Chapter VIII (Art. 71, EU database) and Chapter IX (Art. 72 onwards, post-market monitoring, incident reporting, market surveillance) appear in none of the points. On the wording, they have applied since 2 August 2026. Whether that is intended, or whether they follow the timeline of the high-risk systems they refer to, I cannot say with certainty from the Explorer text. In substance they run empty as long as Section 2 applies to no system. In my posts on Articles 43, 49, 72 and 73 I wrote that they follow the high-risk date, as my reading, not a settled one. The same goes for Art. 111(2), which speaks of “the” date of application of Chapter III, although Chapter III now has three dates.
The technical example
An insurer keeps twelve AI systems in its registry. Two are high-risk under Annex III: a risk grading for life insurance applications (point 5(c)) and a pre-screening assistant in recruiting (point 4). Four fall under Article 50: customer chatbot, voice bot in the hotline, generator for customer letters, image generator in marketing. Six have no specific AI Act obligation beyond Articles 4 and 5, among them claims pre-classification and an internal RAG; internally they still run as tier 2 because they prepare decisions about people.
The board asks: when does what apply to which system, and will we be done by then? One date per company does not answer it. The chatbot has had obligations since August 2026, the letter generator was live before that and has until 2 December 2026 for the marking under Art. 50(2), the risk grading has until December 2027, Article 5 has applied to all twelve since February 2025. That is why every registry record carries a computed field obligations_apply_from, derived from the signed-off classification and a versioned timeline table that encodes Articles 113 and 111. It feeds a countdown dashboard: per system the next deadline, days until then, owner, state of the gates. Whoever changes the class changes the deadline. And when the Omnibus arrived, we versioned one table instead of editing twelve records.
How we implement it
At SIFAMO we keep the timeline in the registry of the Governance Layer as its own versioned table. A registry record references it; simplified:
{
"model_id": "DE-V-2026-0031",
"owner": "Head of Underwriting, Life",
"classification": {"ai_act_class": "high_risk_annex_iii", "basis": "Art. 6(2), Annex III point 5(c)",
"placed_on_market": "2027-03-01 (planned)", "signed_off": ["business", "risk", "data protection"], "version": "class-v2"},
"obligations_apply_from": [
{"scope": "Art. 4, Art. 5", "date": "2025-02-02", "status": "applies"},
{"scope": "Chapter III Sections 1 to 3 (Art. 8 to 27)", "date": "2027-12-02", "days_left": 743},
{"scope": "Art. 43, 49, 72, 73", "date": "2027-12-02", "note": "SIFAMO reading, wording of Art. 113 does not list them"}
],
"timeline_version": "aiact-timeline-2026-1744",
"gates": {"registry": "ok", "gateway_only": "ok", "evidence_log": "ok", "oversight_tier": "full", "eval_set": "in progress", "annex_iv": "open"}
}
Computation, not maintenance
The field is a list, not a date, because several stages apply to every system. The rule is small: Art. 4 and 5 for all; Art. 50 for systems that interact with people or generate content, with the deadline from Art. 111(4) for existing systems; Chapter III from December 2027 or August 2028 depending on the annex. For the articles whose date the wording leaves open, we set the high-risk date and write the reading as a note into the record. The classification is versioned and signed off three ways: business, risk, data protection. Without sign-off the field does not compute, without the field there is no route at the gateway.
What we do in 2026 and 2027 regardless
The shift changes nothing about four things we build first in every project, because supervisors already expect them today. BaFin has demanded an inventory, owners and validation for models for years, DORA has required logging and reporting of major ICT incidents since January 2025. That matches what I describe under AI Governance:
Registry. Without an inventory you cannot demonstrate Article 5, and that has applied since February 2025. Every system, one class, one owner, one expiry date.
Gateway. One entry point for all model calls. That is where we enforce prohibitions, marking under Article 50 and pseudonymisation, today, not in 2027.
Evidence log. Append-only per decision. Article 12 bites later, DORA and MaRisk demand traceability now. Whoever starts logging in 2027 has no history in 2028.
Human oversight. Tiered, with blind re-checks. We use the two years to settle reviewer rates and escalation paths before Article 14 demands them.
For the two high-risk systems, the time until December 2027 is build time: an evaluation set owned by the business, the Annex IV dossier, the conformity pipeline, registration under Article 49. The auditor sees the timeline table in all its versions, per system the computed deadlines with basis and note, the signed-off classification and monthly snapshots of the dashboard in the evidence log. So not only when a deadline applies, but when we recognised it.
Typical mistakes
Reading the Omnibus as an all-clear. Projects stopped although Articles 4, 5 and 50 have long applied and BaFin and DORA do not wait for 2027.
One date per company. “The AI Act applies to us from 2027” overlooks the chatbot, the letter generator and every question under Article 5.
Deadlines in the project plan instead of the registry. When the timeline changes, nobody knows which systems are affected.
Classification without sign-off. The deadline hangs on a class a developer ticked and nobody signed.
Article 113 checklist
Every AI system is in the registry with a signed-off classification and placing-on-the-market date.
The timeline under Articles 113 and 111 exists as a versioned table, at the state of Regulation (EU) 2026/1744.
Per system, obligations_apply_from is computed, as a list with basis, not as one date.
Systems with interaction or content generation are handled under Article 50, existing systems by December 2026.
Registry, gateway, evidence log and oversight run for all systems, regardless of class.
For high-risk systems there is a build plan to December 2027 with evaluation set, dossier and conformity pipeline.
Open questions
Section 5, Chapters VIII and IX. Do conformity assessment, registration, post-market monitoring and incident reporting apply on the wording since August 2026, or with the high-risk dates? A clarification from the Commission is missing.
Art. 111(2). Which of the three dates of Chapter III is “the” date of application, and what is a significant change in design for a model that is retrained regularly?
Standards and guidelines. Whether harmonised standards and the guidelines under Art. 6(5) arrive in time before December 2027 is open.
Frequently asked questions
Does this apply to us as deployers too?
Yes. Article 113 does not distinguish by role. Your obligations as a deployer under Article 26 follow the high-risk dates, Articles 4 and 5 have applied to you since February 2025, Article 50(3) and (4) since August 2026. Whoever significantly modifies a purchased system becomes a provider under Article 25.
Did the Omnibus remove any obligations?
No. It moved the dates for Chapter III, added two prohibitions to Article 5 that apply from December 2026, and introduced a transition period for existing systems under Article 50 with Art. 111(4). What you have to build is the same as before.
What if we put a high-risk system into service before December 2027?
Under Art. 111(2), the Regulation covers existing systems only in case of a significant change in design, for public authorities by 2030 at the latest. I would not build on that. Models get retrained, purposes get extended, the term is open, and supervisors expect the controls anyway. We build as if Chapter III already applied.
EU AI Act Article 113: Timeline and Deadlines, Implemented in Practice
EU AI Act Article 113 sets the date from which each obligation applies, in stages from February 2025 to August 2028. Our answer: every registry record carries a computed field obligations_apply_from, derived from the risk class and a versioned timeline, and that field feeds the countdown for the board.
Article 113 is the shortest article with the biggest effect on project plans. It contains no obligation for providers or deployers, it only says when the other articles start to bite. Since amending Regulation (EU) 2026/1744 it has been rewritten, and in projects I see the shift being read as an all-clear. It is not. I treat Article 113 together with Article 111, because the transition rules for existing systems cannot be read without it.
The EU AI Act has been in force since 1 August 2024. The 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. The obligations for high-risk systems under Annex III apply, after the shift by amending Regulation (EU) 2026/1744 (Digital Omnibus), from 2 December 2027, for Annex I products from 2 August 2028. Article 113 is the article that fixes exactly this timeline. The Omnibus moved dates; it removed no obligation.
What EU AI Act Article 113 requires
Three paragraphs: entry into force (para. 1), the general date of application (para. 2), exceptions in para. 3 points (a) to (d). Plus the transition rules in Article 111. The stages according to the consolidated text in the AI Act Explorer:
The Omnibus rewrote point (c): before, it only named Art. 6(1) with 2 August 2027, and Annex III systems fell under 2 August 2026. Now point (c) names Chapter III Sections 1 to 3 explicitly, without Art. 6(5), with two dates. Point (a) carves out the new prohibitions, point (d) is new. Art. 111(2) now refers to the date of application of Chapter III instead of a date, and para. 4 is new.
I have to be honest here. Point (c) shifts Sections 1 to 3 of Chapter III. Section 5 (Art. 40 to 49: standards, conformity assessment, declaration of conformity, CE marking, registration), Chapter VIII (Art. 71, EU database) and Chapter IX (Art. 72 onwards, post-market monitoring, incident reporting, market surveillance) appear in none of the points. On the wording, they have applied since 2 August 2026. Whether that is intended, or whether they follow the timeline of the high-risk systems they refer to, I cannot say with certainty from the Explorer text. In substance they run empty as long as Section 2 applies to no system. In my posts on Articles 43, 49, 72 and 73 I wrote that they follow the high-risk date, as my reading, not a settled one. The same goes for Art. 111(2), which speaks of “the” date of application of Chapter III, although Chapter III now has three dates.
The technical example
An insurer keeps twelve AI systems in its registry. Two are high-risk under Annex III: a risk grading for life insurance applications (point 5(c)) and a pre-screening assistant in recruiting (point 4). Four fall under Article 50: customer chatbot, voice bot in the hotline, generator for customer letters, image generator in marketing. Six have no specific AI Act obligation beyond Articles 4 and 5, among them claims pre-classification and an internal RAG; internally they still run as tier 2 because they prepare decisions about people.
The board asks: when does what apply to which system, and will we be done by then? One date per company does not answer it. The chatbot has had obligations since August 2026, the letter generator was live before that and has until 2 December 2026 for the marking under Art. 50(2), the risk grading has until December 2027, Article 5 has applied to all twelve since February 2025. That is why every registry record carries a computed field obligations_apply_from, derived from the signed-off classification and a versioned timeline table that encodes Articles 113 and 111. It feeds a countdown dashboard: per system the next deadline, days until then, owner, state of the gates. Whoever changes the class changes the deadline. And when the Omnibus arrived, we versioned one table instead of editing twelve records.
How we implement it
At SIFAMO we keep the timeline in the registry of the Governance Layer as its own versioned table. A registry record references it; simplified:
Computation, not maintenance
The field is a list, not a date, because several stages apply to every system. The rule is small: Art. 4 and 5 for all; Art. 50 for systems that interact with people or generate content, with the deadline from Art. 111(4) for existing systems; Chapter III from December 2027 or August 2028 depending on the annex. For the articles whose date the wording leaves open, we set the high-risk date and write the reading as a note into the record. The classification is versioned and signed off three ways: business, risk, data protection. Without sign-off the field does not compute, without the field there is no route at the gateway.
What we do in 2026 and 2027 regardless
The shift changes nothing about four things we build first in every project, because supervisors already expect them today. BaFin has demanded an inventory, owners and validation for models for years, DORA has required logging and reporting of major ICT incidents since January 2025. That matches what I describe under AI Governance:
For the two high-risk systems, the time until December 2027 is build time: an evaluation set owned by the business, the Annex IV dossier, the conformity pipeline, registration under Article 49. The auditor sees the timeline table in all its versions, per system the computed deadlines with basis and note, the signed-off classification and monthly snapshots of the dashboard in the evidence log. So not only when a deadline applies, but when we recognised it.
Typical mistakes
Article 113 checklist
Open questions
Frequently asked questions
Does this apply to us as deployers too?
Yes. Article 113 does not distinguish by role. Your obligations as a deployer under Article 26 follow the high-risk dates, Articles 4 and 5 have applied to you since February 2025, Article 50(3) and (4) since August 2026. Whoever significantly modifies a purchased system becomes a provider under Article 25.
Did the Omnibus remove any obligations?
No. It moved the dates for Chapter III, added two prohibitions to Article 5 that apply from December 2026, and introduced a transition period for existing systems under Article 50 with Art. 111(4). What you have to build is the same as before.
What if we put a high-risk system into service before December 2027?
Under Art. 111(2), the Regulation covers existing systems only in case of a significant change in design, for public authorities by 2030 at the latest. I would not build on that. Models get retrained, purposes get extended, the term is open, and supervisors expect the controls anyway. We build as if Chapter III already applied.
This post is part of my series EU AI Act in practice. Here you find all articles at a glance.
Related: EU AI Act · All articles · AI Governance · AI Consultant
Recent Posts
Recent Comments