CyphrQ
Menu

EU AI Act Timeline: What Applies When and What to Do Now

The EU AI Act timeline from 2025 to 2028, including the 2026 AI Omnibus changes, and the practical steps an organisation should have in place today.

By CyphrQ.ai

Written with AI assistance and human editorial review. Every figure linked to its source.

12 min read

The EU AI Act timeline runs in stages, and as of 25 October 2026 the first three have already started: bans and AI literacy (February 2025), general-purpose AI model rules (August 2025) and the general date, including transparency duties (August 2026). High-risk rules now start on 2 December 2027 and 2 August 2028, after a 2026 amendment. What to have in place now: a list of your AI systems, a risk class for each, a named owner, and staff who know the rules for the tools they use.

Updated 25 October 2026. Re-check before 2 December 2026, when the next dates apply.

This is general information, not legal advice. Check your own position with a qualified lawyer.

What is the EU AI Act timeline?

The EU AI Act is Regulation (EU) 2024/1689, the European Union's law on artificial intelligence. It came into force on 1 August 2024, and its rules switch on at different dates. Article 113 sets those dates. It was amended by Regulation (EU) 2026/1744 of 8 July 2026 (the "Digital Omnibus on AI"). The dates below follow the Commission's consolidated text of Article 113, which is marked up with the 2026 amendments. I have not checked each date line by line against the amending Regulation itself, so confirm them there before you rely on them.

Date What applies Status on 25 October 2026
2 February 2025 Prohibited AI practices (Article 5) and AI literacy (Article 4) Applies
2 August 2025 Governance rules and obligations for general-purpose AI models Applies
2 August 2026 General date of application, including the transparency rules of Article 50 Applies
2 December 2026 New prohibitions on intimate-image and child abuse material generators; deadline for older systems to meet the Article 50(2) marking duty Upcoming
2 December 2027 High-risk systems listed in Annex III (for example employment, education, biometrics) Upcoming
2 August 2028 High-risk systems that are safety components of products regulated under Annex I (for example lifts, toys) Upcoming

Two caveats. First, this table gives the main dates, not every transitional rule. Article 111 has more, such as a deadline of 2 August 2027 for general-purpose AI models that were on the market before 2 August 2025. Second, the Commission's own page treats the transparency rules as taking effect in August 2026, while the consolidated Article 111(4) gives providers of generative systems already on the market before that date until 2 December 2026 to meet the marking duty in Article 50(2). Both come from the Commission's pages. I have not checked the 2 December 2026 date in the Regulation's own text, so confirm it there before you plan around it.

Did the EU delay the AI Act?

Partly, and the change is now law. In November 2025 the Commission proposed the "AI Omnibus", a package meant to simplify how the Act is implemented. The final act is Regulation (EU) 2026/1744 of 8 July 2026 (the "Digital Omnibus on AI"), which amends Regulation (EU) 2024/1689. EUR-Lex shows it published in the Official Journal on 24 July 2026 and marked "In force". According to the Commission's AI Act page, a political agreement was reached on 7 May 2026 and the Omnibus "entered into force on 27 July 2026". Those two dates are the Commission's, not something I checked in the Regulation.

So this is no longer a proposal. The practical effects:

  • High-risk systems under Annex III moved to 2 December 2027.
  • High-risk systems under Annex I moved to 2 August 2028.
  • A new set of prohibitions was added, applying from 2 December 2026.
  • The AI literacy duty in Article 4 was reworded (covered below).

If you read an older article saying the high-risk rules start in August 2026, it predates the amendment. The Commission page describes these changes as simplification and extension, not as a repeal. The obligations still arrive. They arrive later.

Which AI practices are banned under the EU AI Act?

Article 5 lists the practices banned since 2 February 2025. The Commission's consolidated Article 5 includes, in short:

  • AI that uses subliminal techniques beyond a person's consciousness to cause significant harm, or exploits vulnerabilities linked to age, disability or social or economic situation.
  • Social scoring that leads to unfavourable treatment in unrelated contexts or out of proportion.
  • Criminal risk prediction based solely on profiling or personality traits.
  • Building facial recognition databases through untargeted scraping of facial images from the internet or CCTV.
  • Inferring emotions of people in the workplace and in education institutions, except for medical or safety reasons.
  • Biometric categorisation that infers traits such as race, political opinions, religion or sexual orientation.
  • Real-time remote biometric identification in public spaces by law enforcement, with narrow exceptions.

The Omnibus added two more prohibitions, shown in the Commission's consolidated text: AI that creates or alters realistic intimate imagery of identifiable people without their explicit consent, and AI that generates child sexual abuse material. According to the consolidated Article 113, these apply from 2 December 2026.

For most businesses, the one worth checking now is emotion inference. A call-centre or hiring tool that claims to read employee or candidate emotions from voice or video deserves a hard look before the next renewal.

What does the EU AI Act require for AI literacy?

Article 4 requires providers and deployers of AI systems to take measures to support the AI literacy of their staff and of other people operating AI on their behalf. A deployer is an organisation that uses an AI system under its own authority.

The Omnibus changed the wording. In the Commission's consolidated text, the duty is to "take measures to support the development of AI literacy", taking into account people's knowledge, experience, education and the context of use. The text adds that "this obligation does not require providers or deployers to guarantee any specific level of AI literacy of any individual." The Commission and Member States are to help, in particular SMEs, including with practical examples of how to comply.

In plain terms: you do not need a certificate for every employee. You do need to show you have done something proportionate. Training that matches the role is the obvious answer: what the tool is, what it gets wrong, what data must stay out of it. My earlier post on what to keep out of AI chatbots at work is a starting point for that material.

What rules apply to general-purpose AI models?

General-purpose AI (GPAI) models, the large models behind chatbots and many AI products, have had their own obligations since 2 August 2025. These fall mainly on the companies that build the models. If you only use them, your exposure is indirect: your vendor must meet them.

To help providers comply, the Commission published the General-Purpose AI Code of Practice on 10 July 2025. It has three chapters: Transparency, Copyright, and Safety and Security. The Commission and the AI Board describe it as an "adequate voluntary tool". When this post was drafted, the page listed Anthropic, Amazon, Google, IBM, Microsoft, Mistral AI and OpenAI among the signatories. xAI signed only the Safety and Security chapter. Signatory lists change, so check the page for the current one.

The Commission also issued guidelines on key concepts for GPAI models. For a buyer, the useful question is simple: has this vendor signed the code, and if not, how does it show compliance?

What do the EU AI Act transparency rules require from August 2026?

Article 50 sets four transparency duties, summarised from the consolidated text:

  1. Chatbots. Providers must design systems so people are told they are interacting with AI, unless that is obvious.
  2. Synthetic content. Providers of systems that generate audio, image, video or text must mark outputs "in a machine-readable format" so they are detectable as artificially generated.
  3. Emotion recognition and biometric categorisation. Deployers must inform the people exposed to the system.
  4. Deepfakes and AI-written public-interest text. Deployers must disclose that content has been artificially generated or manipulated. For public-interest text, there is an exception where the content has been through human review or editorial control and someone holds editorial responsibility.

If you publish AI-generated images or run a customer chatbot, items 1 and 4 are likely yours. Marketing and communications teams should read them this quarter.

How much can the EU AI Act fine?

Article 99 sets the ceilings. The figures below are quoted from the Commission's consolidated Article 99. I have not compared them with the Official Journal text, which is the authoritative version:

Breach Maximum fine
Non-compliance with the Article 5 prohibitions (Article 99(3)) up to EUR 35,000,000 or, for an undertaking, up to 7% of total worldwide annual turnover for the preceding financial year, whichever is higher
Non-compliance with listed operator obligations, including Article 50 transparency (Article 99(4)) up to EUR 15,000,000 or up to 3% of turnover, whichever is higher
Incorrect, incomplete or misleading information to authorities (Article 99(5)) up to EUR 7,500,000 or up to 1% of turnover, whichever is higher

For SMEs, including start-ups, Article 99(6) says each fine is up to the percentage or amount above, "whichever thereof is lower". These are maximums. They are not what a typical case costs, and I have no enforcement decisions to cite yet.

What should organisations have in place now?

The dates are staggered, but the preparation is the same work. Do it in this order:

  1. Name an owner. One person accountable for AI governance, with the authority to say no.
  2. Build an AI inventory. Every AI system and AI feature in use, including ones staff adopted without asking. My shadow AI post covers how to find them.
  3. Sort by role. For each system, are you the provider (you build or brand it) or the deployer (you use it)? Duties differ.
  4. Check the bans. Compare each system with Article 5. Emotion inference and biometric categorisation are where ordinary business tools most often get close.
  5. Mark likely high-risk systems. Tools that screen candidates, assess students, or affect access to essential services are the classic Annex III areas. The Commission's page lists biometrics, critical infrastructure, education, employment, migration and justice.
  6. Plan the transparency labels. Where do your chatbots, generated images and AI-written public text need a disclosure?
  7. Run proportionate AI literacy training. Role-based, short, repeated.
  8. Ask vendors in writing. Which AI Act role do they take, which risk class do they assign, and have they signed the GPAI code?
  9. Set a review date. The rules have already moved once. Re-check the consolidated text every quarter.

Going deeper: for security and IT teams

The quotations in this section come from the Commission's consolidated text. The Official Journal text is the authoritative version.

Start with an inventory that can be audited. A spreadsheet of tool names is not enough. For each system record: owner, vendor, model and version, purpose, data categories it touches, who is affected by its outputs, whether you are provider or deployer, where it runs, and the date of the last review. Pull from several sources: SSO and OAuth consent grants, procurement and expense records, browser extension inventories, API gateway and egress logs, and SaaS admin consoles where AI features can be switched on by a vendor update. Treat a vendor adding an AI feature to a tool you already approved as a change that needs re-review.

How do you classify an AI system under the AI Act?

  1. Is it an AI system as the Regulation defines the term? Do not assume every analytics or rules engine qualifies, and do not assume none does.
  2. Does the use match an Article 5 prohibition? If yes, stop the use.
  3. Does it fall under Article 6(2) and an Annex III area, or is it a safety component of a product under Article 6(1) and Annex I? If yes, it is high-risk and the dates are 2 December 2027 and 2 August 2028 respectively.
  4. Does it interact with people or generate content? If yes, Article 50 applies from August 2026.

Write down the reasoning for each decision. An auditor will ask why you concluded a system was not high-risk.

Article 15: accuracy, robustness, cybersecurity. For high-risk systems, the consolidated Article 15 requires an "appropriate level of accuracy" and consistent performance "throughout their lifecycle". Paragraph 4 requires resilience against "errors, faults or inconsistencies", in particular from interaction with people or other systems, and asks developers to reduce, as far as possible, the risk of "possibly biased outputs". Paragraph 5 requires resilience "against attempts by unauthorised third parties to alter their use, outputs or performance by exploiting system vulnerabilities." In engineering terms: define accuracy metrics per use case and test against them on a schedule; test for drift after model or data changes; red-team for adversarial input; and treat model, prompt and data pipelines as assets under your normal vulnerability management. The OWASP risks in my posts on prompt injection and AI agents are a ready checklist for the attacks paragraph 5 describes.

What logs does the AI Act require?

Article 12 requires high-risk systems to "technically allow for the automatic recording of events (logs) over the lifetime of the system", supporting traceability, risk identification and post-market monitoring. On the deployer side, Article 26 requires keeping the logs under your control "for a period appropriate to the intended purpose of the high-risk AI system, of at least six months". Article 26(2) requires human oversight assigned to people with "the necessary competence, training and authority". Practical steps: confirm each vendor exposes logs (inputs, outputs, model version, overrides) by export or API; route them to your SIEM; set retention to at least six months and check it against your data protection retention rules; and record who exercised oversight and when they overrode the system.

Evidence to keep. Inventory with change history, classification memos, training records, vendor responses, test results for Article 15 metrics, and an incident log. If a question comes from a regulator, you want to answer from records, not memory.

Ask your own team. Can we list every AI system and feature in use today? Which ones do emotion or biometric inference? Which are used in hiring or student assessment? Where do we publish AI-generated content without a label? How long are vendor logs kept, and can we export them?

Key takeaways

  • The ban and AI literacy rules have applied since 2 February 2025. GPAI model rules since 2 August 2025. The general date, including transparency, was 2 August 2026.
  • The AI Omnibus, Regulation (EU) 2026/1744, is in force (27 July 2026, per the Commission). High-risk dates are now 2 December 2027 (Annex III) and 2 August 2028 (Annex I).
  • Article 4 now asks you to support AI literacy, not to guarantee a level for each person.
  • Maximum fines run to EUR 35 million or 7% of turnover for prohibited practices (Article 99(3)).
  • Start with an owner, an inventory, a classification and vendor questions. That work serves every deadline.

Sources