The EU AI Act: what starts on 2 December 2026, and what waits until December 2027

Regulation (EU) 2026/1744, the Digital Omnibus on AI, was published in the Official Journal on 24 July 2026 and entered into force on 27 July 2026, the third day after publication. It amends the EU AI Act, Regulation (EU) 2024/1689.

Three of its changes to the calendar matter here. It moved the high-risk rules later. It added two prohibitions, which start on 2 December 2026. And it gave providers of generative AI systems placed on the market before 2 August 2026 until 2 December 2026 to mark their output, in a machine-readable way, as AI-generated or manipulated (the Article 50(2) marking duty).

The rules for the high-risk uses listed in Annex III, which include recruitment, managing workers, credit scoring, and risk assessment and pricing for life and health insurance, were due to apply from 2 August 2026. They now apply from 2 December 2027, sixteen months later. The rules for AI in products already covered by EU product safety law, listed in Annex I, moved from 2 August 2027 to 2 August 2028.

Both are written into Article 113 as fixed dates. They are not tied to harmonised standards being ready or to any other trigger. The European Commission's AI Act page, last updated on 3 August 2026, and the AI Office of Ireland's timeline give the same dates.

Four sets of duties already apply, each to particular operators: the original prohibitions, the AI literacy duty, the rules for general-purpose AI models, which bind only model providers, and the Article 50 transparency duties. Two more start on 2 December 2026. The high-risk regime waits until December 2027 for Annex III uses and August 2028 for AI in regulated products.

Article 86 gives a right to an explanation of certain decisions taken with an Annex III system. It sits in a chapter whose date did not move. But it attaches to high-risk systems, and Annex III systems are classed as high-risk only from 2 December 2027. Whether it reaches any decision before then is not settled.

The EU penalty provisions have applied since 2 August 2025. Breaching a prohibition carries a fine of up to 35 million euro or 7% of worldwide annual turnover. Breaching the duties of providers, deployers (those using AI systems), importers and distributors carries up to 15 million euro or 3%. Those duties include the Article 50 transparency duties now, and the Article 26 deployer duties from December 2027.

For a business, each cap is the higher of the two figures, with two exceptions. A small or medium enterprise faces the lower of the two (Article 99(6)). Since 27 July 2026, a small mid-cap enterprise also faces the lower of the two, for every tier except the prohibitions (Article 99(6a)). The caps are lower for a smaller business, but the duties are the same.

In Ireland, the Regulation of Artificial Intelligence Act 2026 sets out how those fines are imposed. For most of the designated regulators there are three steps: a notice from an authorised officer, referral to an independent adjudicator, and an adjudication that takes effect only when the High Court confirms it. The Central Bank and the Competition and Consumer Protection Commission fine under their own Acts, as the 2026 Act amends them. The Act came into operation on 31 July 2026, apart from two paragraphs in section 78 that would add Coimisiún na Meán and the Data Protection Commission to the regulators that can use the adjudication route.

Two things. The first binds anyone who supplies a particular kind of generator, or uses one for that purpose. The second is the end of the extra time Article 111(4) gave providers of generative systems placed on the market before 2 August 2026. A business that built a chatbot or generator on another company's model can be such a provider. If it built the feature only for its own use, the text gives it no extra time, as set out below.

The first is two new prohibitions in Article 5. Point (ba) bans placing on the market, putting into service or using an AI system that generates or manipulates realistic images, video, audio or similar material of an identifiable person's intimate parts, or of an identifiable person engaged in sexually explicit activity, without that person's freely given, specific, informed, unambiguous and explicit consent. Point (bb) bans the same for material or performance covered by Article 2(c) and (e) of Directive 2011/93/EU, which the Omnibus calls child sexual abuse material, except where a "without right" defence applies under national law.

The new Article 5(1a) limits who is caught. A provider breaches it only in one of two cases. One is where generating or manipulating that material is the system's intended purpose. The other is where the system's design, training, architecture, capabilities or user-facing features make that a reasonably foreseeable and reproducible outcome without significant technical modification, and the system lacks reasonable and adequate technical safety measures and other safeguards to reliably prevent it, taking account of reasonably foreseeable misuse, and to correct observed or reported misuse. A deployer, the business using the system, breaches it only where it uses the system for the purpose of generating or manipulating that material or performance. Article 5(1b) adds that editing which neither increases the exposure of intimate parts nor alters the nature of a sexually explicit activity is not manipulation for point (ba).

What that means for a business: using a general image or video generator for marketing does not bring you within either ban. Recital 12 of the Omnibus says the ban on use does not cover using a system for lawful purposes, even where it lacks the safeguards its provider should have built, and does not cover accidental generation. A business that offers image, video or voice generation to the public under its own name may be the provider of that feature under Article 3(3). If it is, and generating that material is not the feature's purpose, the question from 2 December 2026 has two parts: whether its design makes that material a reasonably foreseeable and reproducible outcome without significant technical modification, and if it does, whether its safeguards are reasonable and adequate.

Both are Article 5 breaches, in the highest fine tier. Which Irish body enforces them is unclear. The Irish designation regulations, as amended on 31 July 2026, assign points (a) to (h) of Article 5(1), in whole or in part, to named regulators. They name no body for points (ba) or (bb). The AI Office of Ireland is designated as a market surveillance authority in general terms, for the purposes of Article 70(1). It is not among the regulators that section 78 of the 2026 Act lets use the adjudication route. On the text of sections 2(1) and 68 of the 2026 Act, that designation lets it appoint authorised officers under Part 5. Section 72 lets an authorised officer serve a prohibition notice on a person suspected of infringing Article 5. Neither instrument says whether the AI Office's general designation reaches points (ba) and (bb), or which Irish body enforces them.

The second is the end of the extra time for the Article 50(2) marking duty. Article 50(2) requires the provider of an AI system that generates synthetic audio, image, video or text to make sure its outputs are marked in a machine-readable format and detectable as artificially generated or manipulated. It does not apply to the extent the system performs an assistive function for standard editing, or does not substantially alter the input the deployer provides, or its meaning. It has applied since 2 August 2026. Article 111(4), added by the Omnibus, gives providers of generative systems placed on the market before 2 August 2026 until 2 December 2026 to comply. A system placed on the market on or after 2 August 2026 gets no extra time.

A business that only uses a generative tool it bought owes nothing under Article 50(2). Its own Article 50 duties, such as disclosing a deep fake it publishes, have applied since 2 August 2026 and did not move.

A business that built its own generative feature, a website chatbot included, and put it into service under its own name can be the provider of it. The Commission's Article 50 guidelines give, as an example of a provider, an organisation that developed a chatbot in-house and put it into service for its own use (paragraph 11). They let a provider rely on marking built in by the supplier of the model underneath, while it stays responsible for showing that the result complies (paragraph 74). That business may not get the extra time. Article 111(4) gives it to providers of systems placed on the market before 2 August 2026. The guidelines describe it as covering systems placed on the market or put into service before then (paragraph 153). The text does not mention putting into service, though, so do not count on the extra time for a feature built for your own use. The Article 50 guide sets out who owes which duty, and the wording that satisfies it.

Annex III lists the standalone high-risk uses. Point 4 covers AI used for recruitment and selection, including targeted job advertising, filtering applications and evaluating candidates, and AI used to make decisions on terms of work, promotion or termination, to allocate tasks based on individual behaviour or personal traits, or to monitor and evaluate performance and behaviour. Point 5(b) covers evaluating creditworthiness or setting a credit score, except to detect financial fraud, and point 5(c) covers risk assessment and pricing for life and health insurance. The other points cover biometrics, critical infrastructure, education, public services and benefits, emergency calls and triage, law enforcement, migration, and the administration of justice and democratic processes.

From 2 December 2027, the rules in Sections 1 to 3 of Chapter III apply to those uses. Until then none of them does. The GDPR still applies to an AI tool used in hiring or credit decisions wherever personal data is processed: Article 2(7) of the AI Act says the Act does not affect the GDPR, apart from Articles 4a and 59, on processing personal data to detect bias and in regulatory sandboxes. Automated decisions in hiring are covered separately.

One right tied to Annex III sits outside the deferred chapter. Article 86 covers decisions a deployer takes on the output of an Annex III high-risk system, other than one in critical infrastructure. A person affected by such a decision has the right to clear and meaningful explanations of the system's role and the main elements of the decision. It applies where the decision has legal effects, or similarly significant effects that the person considers adverse to their health, safety or fundamental rights. It sits in Chapter IX, which the Omnibus did not defer and which has applied since 2 August 2026. But Article 6(2), which makes Annex III systems high-risk, applies only from 2 December 2027. And on the reading of Article 111(2) below, Article 86 would not reach systems placed on the market or put into service before then until their design changes significantly. Article 86(3) applies it only where Union law does not already give the right.

If an Annex III use describes something you run, the important question is whether you are a deployer or a provider. Broadly, a deployer uses the system, and a provider makes it, or has it made, and puts its name on it. The obligations are very different.

A business that uses a bought tool under its own authority, without putting its own name on it, substantially modifying it or changing what it is for, is a deployer. Article 26 requires a deployer to take measures to: use the system in line with the provider's instructions for use; assign human oversight to people with the necessary competence, training, authority and support; monitor its operation and report risks and serious incidents; and keep the logs under its control for at least six months unless other law says otherwise. An employer must inform workers' representatives and the affected workers before putting a high-risk system into service or using it at the workplace. A deployer of an Annex III system that makes or helps make decisions about people must tell those people.

Building on a tool can make you its provider instead, with the full weight of Article 16: risk management, data governance, technical documentation, conformity assessment. Article 25 makes you the provider of a high-risk system if you put your own name or trademark on one already on the market, if you make a substantial modification to one, or if you modify the intended purpose of an AI system so that it becomes high-risk. The Article 25 guide works through a CV-scoring example.

The third case, changing what a system is for, needs the closest reading. A general-purpose AI system, such as a general chat assistant, has not been classified as high-risk. If you wire it into a workflow that screens job applications, at a business of any size, then on the face of Article 25(1)(c) you have modified its intended purpose into an Annex III use. Done on or after 2 December 2027, that makes you the provider of a high-risk system, and the company that supplied the original system stops being treated as its provider. If you repurpose a system this way before that date, it is not settled whether it counts as placed on the market or put into service for the Article 111(2) transition described below. The Act does not say whether a change of purpose made before 2 December 2027 counts as placed on the market or put into service under Article 111(2), so do not plan on it.

Two of the heavier duties apply only to particular deployers. The fundamental rights impact assessment under Article 27 binds public bodies, private entities providing public services, and deployers of credit scoring and life or health insurance systems. Registration in the EU database under Article 26(8) applies to deployers that are public authorities or EU bodies. A private recruitment agency or employer is outside both unless it provides public services, which recital 96 of the AI Act illustrates with education, healthcare, social services, housing and the administration of justice. In that case the impact assessment applies to it.

Article 111(2), as replaced by the Omnibus, decides what happens to high-risk systems placed on the market or put into service before the rules start. The Regulation applies to operators of high-risk AI systems placed on the market or put into service before the date Chapter III applies to them only if, as from that date, those systems are subject to significant changes in their designs. For Annex III systems that date is 2 December 2027. For Annex I systems it is 2 August 2028.

An operator includes the provider and the business deploying the system (Article 3(8)), and putting into service includes supplying a system for the provider's own use. The transition does not affect Article 5, so a system already on the market or in service is still bound by the prohibitions from their own start dates. And providers and deployers of high-risk systems intended to be used by public authorities must comply by 2 August 2030 in any case.

Two recitals explain how the transition is read. Recital 39 of the Omnibus says it follows the type and model: if at least one unit was lawfully placed on the market or put into service before the date, other units of the same type and model are covered as long as the design stays unchanged, and the decisive date is when the first unit of that type and model was placed on the market or put into service in the EU. Recital 177 of the AI Act was written for the original Article 111(2), which was tied to 2 August 2026. It speaks of significant changes in design or intended purpose from that date, where the Article itself says only design. It says a significant change is to be understood as equivalent in substance to a substantial modification. Article 3(23) defines that as a change not foreseen or planned in the provider's initial conformity assessment that affects the system's compliance with the high-risk requirements or changes its intended purpose.

Read that way, the text covers the business deploying a system as well as its provider. It applies to operators, and its second sentence names deployers when it makes the exception for systems intended for public authorities. Recital 177 says the aim includes ensuring continuity of the use of AI systems. A business using an Annex III system whose type and model was first lawfully placed on the EU market or put into service before 2 December 2027 is then outside the Regulation, apart from Article 5, until that system's design changes significantly. That would take it outside the Article 26 duties and Article 86 too. The Article 25 guide reads it the same way. That is a reading of the text. It is not guidance from the Commission or an Irish regulator. Whether that also lifts duties that applied to the system before that date, such as Article 4 and Article 50, is not settled, so do not count on it.

Article 111(2) does not say which updates to a tool sold as a service count as significant changes in its design. That turns on what the supplier changes, so it is a question to put to the supplier: when that type and model was first placed on the EU market or put into service, and whether significant design changes are planned after 2 December 2027. Keep the answer in writing.

Article 4 has applied since 2 February 2025. Regulation (EU) 2026/1744 replaced it from 27 July 2026. Providers and deployers must now take measures to support the development of AI literacy of their staff and others who operate or use AI systems on their behalf, taking account of their technical knowledge, experience, education and training, the context the systems are used in and the people they are used on. The new text adds that the obligation does not require providers or deployers to guarantee any specific level of AI literacy of any individual.

Before 27 July 2026 the wording was to take measures to ensure, to their best extent, a sufficient level of AI literacy. A policy that repeats the old wording now overstates the duty. The Commission's questions and answers, last updated on 27 July 2026, say there is no need for a certificate and that organisations can keep an internal record of training and other initiatives.

The Commission also says national market surveillance authorities supervise and enforce Article 4. Article 99 sets no fine tier for it. But Article 99(1) asks Member States to lay down penalties for any infringement of the Regulation by operators, and the Commission's questions and answers say national authorities could impose penalties for breaching Article 4 under national law. The Regulation of Artificial Intelligence Act 2026 does not mention Article 4. Its contravention notices and fines cover other Articles, so the 2026 Act sets no penalty for it. That could change: the Department said on 30 July 2026 that further legislation, including amendments arising from the Omnibus, would be brought forward in the autumn.

CIPD Ireland's 2026 survey of Irish HR practice found that 33% of the organisations it surveyed offer AI training and 44% give staff clear guidance on AI use. The literacy guide sets out what counts as a measure, and what nobody has to buy.

Article 5(1)(f) bans placing on the market, putting into service for this purpose or using AI systems that infer the emotions of a person in the workplace or in education institutions, except where the system is intended for medical or safety reasons.

This has applied since 2 February 2025. It is not deferred, and it carries the highest fine tier in the Act.

It is also the one the Workplace Relations Commission has named specifically. The WRC is designated as the market surveillance authority for this prohibition outside education, and for the Annex III employment uses, and its own page identifies AI that infers the emotions of workers or job applicants as within its remit.

Recital 18 of the AI Act describes the emotions and intentions it means, such as happiness, sadness, anger, surprise, embarrassment, excitement and satisfaction, and leaves out physical states such as fatigue. It also leaves out the mere detection of readily apparent expressions, gestures or movements, such as a frown, a smile or a raised voice, unless they are used to identify or infer emotions.

The Commission's guidelines on prohibited practices read the ban as covering inferences drawn from biometric data. A tool that infers sentiment from written text is outside it, on the Commission's example. One that infers emotions from facial expressions, posture or the way someone types is inside it. If any part of your hiring process runs video interviews through software that infers how candidates feel from their face or voice, that is the provision to look at first. Not in December 2027. Now.

Ireland has spread the work across existing regulators. Each supervises AI in its own sector, and the AI Office of Ireland coordinates them. S.I. No. 366 of 2025, as amended by S.I. No. 405 of 2026 (made on 31 July 2026), designates fourteen market surveillance authorities, the AI Office among them. The AI Office's own page lists the other thirteen. Who covers what is set out separately.

Your AI Act regulator depends on what you do. The Workplace Relations Commission covers employment. The Central Bank covers AI used by regulated financial service providers in direct connection with providing financial services. The Data Protection Commission covers several prohibited practices, such as untargeted facial image scraping and biometric categorisation, and the high-risk uses in law enforcement, migration, and the administration of justice and democratic processes. Under section 11 of the Data Protection Act 2018 it is also the supervisory authority for the GDPR.

The Regulation of Artificial Intelligence Act 2026, signed into law by President Connolly on 21 July 2026, set up the AI Office of Ireland under section 8. The Office was established on 31 July 2026, the establishment day appointed by S.I. No. 404 of 2026, and its first chief executive, Paul Byrne, was announced on 30 July 2026. It coordinates the designated authorities and, under section 42, is the single point of contact under Article 70(2) of the AI Act.

The WRC has an AI Act page of its own. It is undated, but it refers to the 2026 Act having become law on 21 July 2026. It says the 2026 Act will be supplemented in the coming months by legislation providing supervision and enforcement powers for the WRC, and that public guidance will be available as Ireland's implementation framework matures. Part 5 of the 2026 Act, headed Supervision and Enforcement, has been in operation since 31 July 2026. It lets the designated authorities other than the Central Bank, the WRC among them, appoint authorised officers. Those officers can inspect premises and serve contravention and prohibition notices. Section 78 also names the WRC among the regulators that can refer a case for an adjudicated fine. The page does not say which further powers it expects.

The deferral is not a reason to do nothing, but it is a reason to do the cheap work first and leave the expensive work until it is needed.

Six moves are worth making now, and none of them requires a consultant to start.