Every EU AI Act deadline agencies actually face (post-Omnibus, verified August 2026)
Quick status check for anyone planning Q4 2026:
- Live since 2 Feb 2025: the Article 5 prohibitions on unacceptable-risk AI practices.
- Live since 2 Aug 2026: the Article 50 transparency duties for chatbots and synthetic content.
- Landing 2 Dec 2026: the grace period for machine-readable marking of AI-generated synthetic content ends on that date (Art. 50(2)).
- Deferred: the Annex III high-risk regime to 2 Dec 2027, with high-risk AI embedded in regulated products following at 2 Aug 2028.
- Also live: the Article 4 AI-literacy duty has applied since 2 Feb 2025, and its enforcement was activated on 2 Aug 2026.
The myth this page corrects
For months, agency forums have run two contradictory lines: “the AI Act got postponed across the board” and “the heavy deadlines already hit.” Both are wrong, and each costs money in opposite directions — one produces panic re-planning around dates that moved, the other false comfort around dates that didn’t.
The facts, verifiable against the regulation itself (Regulation (EU) 2024/1689, Art. 113 application schedule, as amended by the Digital Omnibus, Regulation (EU) 2026/1744):
- Art. 5 prohibitions apply from 2 February 2025. Never deferred, never postponed by the Omnibus.
- Art. 50 transparency duties apply from 2 August 2026 to services placed on the market or put into service from that date. Live now.
- Art. 50(2) machine-readable marking carries its own later date: providers must have synthetic-content marking in place when the transitional period for that duty ends on 2 December 2026.
- Annex III high-risk obligations were deferred to 2 December 2027, with embedded high-risk AI in regulated products following on 2 August 2028.
- Art. 4 AI literacy applies since 2 February 2025; enforcement of this duty was activated on 2 August 2026.
Deadlines by obligation — who each binds, current status
“Who it binds” matters more than the calendar. The Act distinguishes providers (who develop or place systems on the market) from deployers (who use systems under their own authority). Most agencies sit in the deployer seat for chatbots, generative tools, and published AI content — unless they white-label an AI system under their own brand, which can shift you toward provider duties for that product. Where you act “to the extent” as a provider or deployer, that partial role follows the same dates as the full one.
| Obligation | Applies / bites | Binds | Status · Aug 2026 |
|---|---|---|---|
| Art. 5 — prohibited practices (manipulation, social scoring, untargeted facial scraping…) | 2 Feb 2025 | Everyone supplying, using, or importing such systems | Live |
| Art. 4 — AI literacy: staff-level understanding of the AI tools your business uses | 2 Feb 2025 | Providers & deployers | In force · enforcement activated 2 Aug 2026 |
| Art. 50(1) — tell people when they’re talking to an AI system (chatbots, voice agents) | 2 Aug 2026 | Providers & deployers | Live |
| Art. 50(4) — disclose AI-generated or manipulated image/audio/video constituting deep fakes; parallel disclosure for public-interest text | 2 Aug 2026 | Deployers publishing or circulating such content | Live |
| Art. 50(3) — inform people exposed to emotion-recognition or biometric-categorisation systems | 2 Aug 2026 | Deployers of such systems | Live |
| Art. 50(2) — mark synthetic content in machine-readable form (automated detection tooling) | 2 Dec 2026 | Providers of generative systems | Grace period ends |
| Annex III high-risk regime (biometrics, critical infrastructure, employment screening, credit scoring…) | 2 Dec 2027 | Providers & deployers of listed systems | Deferred — not yet binding |
| High-risk AI embedded in regulated products (machinery, toys, medical devices…) | 2 Aug 2028 | Providers & deployers | Deferred — not yet binding |
What is live versus deferred — and why agencies should care about the split
The Omnibus deferral bought time only for the heavy compliance regimes: Annex III high-risk conformity work, and the embedded-products wave behind it. Nothing about chatbot disclosure, synthetic-content labeling, or staff AI literacy moved. If your agency runs client chatbots, publishes AI-assisted images or video, or briefs journalists and newsletters with generated copy, your exposure dates are already here — not in 2027.
Live today (no further transition)
- Chatbot disclosure to users — Art. 50(1).
- Visible disclosure of synthetic images, audio, and video that constitute deep fakes — Art. 50(4).
- Notice to people exposed to emotion-recognition or biometric-categorisation systems — Art. 50(3).
- Disclosure when publishing AI-generated text on matters of public interest — Art. 50(4), second subparagraph, with an exemption where human editorial review and editorial responsibility exist.
- AI literacy for staff — Art. 4 (enforcement active since 2 Aug 2026).
- Avoiding banned practices — Art. 5.
Still deferred
- Annex III high-risk regime → 2 Dec 2027.
- High-risk AI embedded in regulated products → 2 Aug 2028.
One date sits between the two groups: Art. 50(2) machine-readable marking, whose transition ends 2 Dec 2026. It binds providers of generative systems — but agencies selecting vendors should already ask those vendors how they comply, because your clients may ask you the same question.
Common misinformation, corrected
-
“The Omnibus postponed everything.” No. Regulation (EU) 2026/1744 deferred specific regimes (Annex III high-risk to Dec 2027; embedded high-risk to Aug 2028). Arts. 4, 5, and 50 kept their original dates: Feb 2025 and Aug 2026 respectively.
-
“Chatbot disclosure is a 2027 problem.” No. Art. 50(1) has been live since 2 August 2026. A client-facing chatbot without AI disclosure is a present-tense gap, not a future one.
-
“AI literacy was quietly cancelled.” No. Art. 4 still applies, and its enforcement was activated on 2 August 2026. It is one of the cheapest obligations to satisfy — documented staff training on the tools you actually use.
-
“High-risk rules already apply to HR-tech clients.” Not yet: Annex III obligations bite at 2 Dec 2027. Planning should start now for affected clients, but no Annex III duty is enforceable today.
-
“Machine-readable watermarking was dropped.” No. Art. 50(2) stands, with its own transition ending 2 December 2026. Visible labels alone do not discharge the marking expectation.
-
“SMEs are exempt.” No exemption exists from these duties. What exists is a proportionality principle and a lower penalty cap for SMEs and small mid-cap enterprises — lighter sanctions, identical obligations.
What to do with this list
- This week: audit every client-facing chatbot and contact widget against Art. 50(1) — see our chatbot disclosure checklist.
- This quarter: inventory published AI-generated media and set up visible + metadata labeling before 2 Dec 2026 — see our labeling guide for images and video.
- Ongoing: keep a short AI-literacy record per team member (tool, training date, sign-off).
- In writing: ask every AI vendor for their Art. 50(2) marking plan before renewals.
- For 2027: flag any client whose systems might fall under Annex III so classification questions reach a lawyer early.
Not sure where your agency actually stands?
Run the free OX-00 self-score: ten questions, two minutes, and you get a readiness rating mapped to the exact dates above — plus pointers on the gaps that matter before 2 Dec 2026. When you want the gaps turned into a prioritized fix list with evidence templates, that's the paid packet.
Take the free AI-readiness self-score →
Prefer a done-for-you version? Our AI Act Ready packet delivers checklists, gap lists, and templates within 72 hours.