In July 2026, Ewa, a compliance specialist at a mid-sized services company, read the headline “AI Act postponed”. She breathed a sigh of relief and moved the AI system register project to next year. A few weeks later, HR asked for approval of a tool that scores candidates’ CVs, marketing started publishing images generated by a model, and a chatbot appeared in the customer app. Each of these decisions falls under a different provision of the AI Act, each has a different deadline — and some of them already apply today.
Regulation (EU) 2026/1744, known as the Digital Omnibus on AI, postponed only the requirements for high-risk systems. The prohibited practices, the Article 4 obligation and the rules on general-purpose AI models have applied since 2025, and transparency under Article 50 since 2 August 2026. Below we set out the timeline with article numbers, show how to determine your company’s role, what fines apply to infringements, what the Polish Act on artificial intelligence systems changes and what is worth doing before the end of 2026.
What did the Digital Omnibus change in the AI Act timeline?
The Digital Omnibus moved the deadlines for high-risk AI requirements to 2 December 2027 (Annex III) and 2 August 2028 (Annex I) and left the other AI Act application dates unchanged. Formally, it is Regulation (EU) 2026/1744 of 8 July 2026, which amends, among others, Regulation (EU) 2024/1689, that is the AI Act. It was published in the Official Journal of the EU on 24 July 2026 and entered into force on the third day after publication, that is on 27 July 2026.
Before the amendment, the timeline was simpler and stricter. The original Article 113 of the AI Act made the obligations for high-risk systems listed in Annex III applicable from the general date of 2 August 2026, and those for systems covered by Article 6(1) and Annex I from 2 August 2027. Guides and slide decks prepared before July 2026 may still repeat these dates. They should be treated as out of date.
The Omnibus also added new elements. Article 5 gained two new prohibitions (points ba and bb) applying from 2 December 2026, Article 111 gained a transition period for marking synthetic content, and Article 99 gained a more lenient rule for calculating fines for small mid-cap enterprises. The wording of Article 4 on AI literacy was also changed. New point (d) of Article 113 states that Articles 102–110 apply from 27 July 2026, but those provisions amend other EU acts and are not a deadline for companies using AI.
No further act amending the AI Act appeared after 2026/1744 — that is what the EU Publications Office relationship database shows as of 10 November 2026. The Commission has proposed further changes concerning medical devices that also touch the AI Act, but that is a proposal, not law in force. For planning inside a company, the state described in this article is what counts.
Which AI Act deadlines apply to companies between 2025 and 2030?
The AI Act applies in stages: the prohibitions and Article 4 have applied since 2 February 2025, the GPAI rules since 2 August 2025, transparency under Article 50 since 2 August 2026, the new prohibitions from 2 December 2026, and the high-risk requirements from 2 December 2027 and 2 August 2028. The table lists every date with its legal basis so that each one can be checked in the text of the regulation. As of 10 November 2026.
| Date | What starts to apply | Legal basis |
|---|---|---|
| 1 August 2024 | entry into force of the AI Act | 2026/1744, recital 1 |
| 2 February 2025 | Chapters I and II: general provisions, including Article 4 (AI literacy), and the prohibited practices in Article 5 | Article 113, third paragraph, point (a) |
| 2 August 2025 | Chapter V (general-purpose AI models), notifying authorities and notified bodies (Chapter III, Section 4), governance (Chapter VII), penalties in Chapter XII except Article 101, confidentiality (Article 78) | Article 113, third paragraph, point (b) |
| 27 July 2026 | entry into force of Regulation 2026/1744; Articles 102–110 (amendments to other EU acts) | 2026/1744, Article 4; Article 113, point (d) |
| 2 August 2026 | general date of application: among others, the transparency obligations in Article 50 and the fines for GPAI model providers in Article 101 | Article 113, second paragraph |
| 2 December 2026 | the new prohibitions in Article 5(1), points (ba) and (bb), and Article 5(1a) and (1b); Article 50(2) for generative systems placed on the market before 2 August 2026 | Article 113, point (a); Article 111(4) |
| 2 August 2027 | compliance of GPAI models placed on the market before 2 August 2025 | Article 111(3) |
| 2 December 2027 | requirements and obligations for high-risk systems under Article 6(2) and Annex III | Article 113, point (c)(i) |
| 2 August 2028 | requirements and obligations for high-risk systems under Article 6(1) and Annex I | Article 113, point (c)(ii) |
| 2 August 2030 | compliance of high-risk systems intended to be used by public authorities | Article 111(2) |
| 31 December 2030 | compliance of AI systems in the large-scale IT systems listed in Annex X, placed on the market before 2 August 2027 | Article 111(1) |
The most common mistake in reading this timeline is treating 2 August 2026 as the deadline for high-risk systems. After the Omnibus, it is the date from which Article 50 applies, among others, but not the obligations in Chapter III, Sections 1–3. The second mistake is reading 2 August 2030 as a general “transition period”. That deadline concerns only high-risk systems intended to be used by public authorities.
It is also worth noting that the amended Article 111(2) refers to “the date of application of Chapter III” rather than a fixed 2 August 2026. High-risk systems placed on the market before that date fall under the regulation only when they undergo significant changes in their design. For companies that are buying HR or scoring tools today, this means that the purchase date and any later modifications will matter as evidence.
Key takeaways:
- The Article 5 prohibitions and the Article 4 obligation have applied since 2 February 2025 — the Omnibus did not postpone them.
- Article 50 transparency applies from 2 August 2026, with a transition period until 2 December 2026 for marking content from systems placed on the market earlier.
- Requirements for high-risk systems: 2 December 2027 (Annex III) and 2 August 2028 (Annex I).
- The 2 August 2030 deadline concerns only high-risk systems used by public authorities.
How do you determine whether your company is a provider or a deployer of an AI system?
A company is a provider when it develops an AI system or model (or has it developed) and places it on the market or puts it into service under its own name, and a deployer when it uses an AI system under its authority in a professional activity. Both definitions are in Article 3(3) and (4) of the AI Act. Most companies using off-the-shelf services, such as chat assistants, AI features in a CRM or document analysis tools, act as deployers.
A role is not a property of the company but a relationship with a specific system. The same organisation can be a deployer of an office assistant and, at the same time, a provider of a chatbot it built on an external model and made available to customers under its own brand. That is why an AI system register needs a “role” column filled in separately for each entry, rather than a single declaration for the whole company.
The AI Act also recognises intermediate roles. An importer is an entity established in the EU that places on the market a system bearing the name or trademark of a company established outside the Union (Article 3(6)). A distributor makes a system available on the EU market and is neither the provider nor the importer (Article 3(7)). Software integrators and resellers should check whether either of these roles describes their business model.
The most important trap lies in Article 25(1). A distributor, importer, deployer or other third party becomes the provider of a high-risk system if it puts its own name or trademark on it, makes a substantial modification to such a system, or changes the intended purpose of a system that was not high-risk so that it becomes one. An example from practice: HR takes a general-purpose language model and configures it to screen out candidates automatically. That is a change of intended purpose which can shift the provider obligations in Article 16 onto the company.
Determining the role is best closed with a short note for each system: who develops the system, under whose name it is made available, who decides on the purpose of use and whether the company has changed anything in it. Such a note serves as evidence during an inspection and as the starting point for a conversation with the supplier about how obligations are divided.
How does the AI Act classify AI systems by risk?
The AI Act divides uses of AI into four groups: prohibited practices (Article 5), high-risk systems (Article 6 with Annexes I and III), systems with transparency obligations (Article 50) and other systems with no specific requirements; a separate regime applies to general-purpose AI models (Chapter V). The category depends on how the system is used, not on the technology. The same language model can power a harmless writing assistant and a high-risk system for assessing employees.
The prohibited practices in Article 5(1) include, among others, subliminal and manipulative techniques, exploiting the vulnerabilities of persons due to age or disability, social scoring, predicting the risk of committing a crime solely on the basis of profiling, building facial recognition databases through untargeted scraping of facial images from the internet or CCTV, and emotion recognition in the workplace and in education (with medical and safety exceptions). On 4 February 2025, the European Commission published guidelines on prohibited practices, and on 6 February 2025 non-binding guidelines on the definition of an AI system. The Omnibus added points (ba) and (bb), which concern systems that generate or manipulate realistic intimate material of an identifiable person without that person’s explicit consent and child sexual abuse material — these prohibitions apply from 2 December 2026.
A system is high-risk in two cases. The first (Article 6(1)) is a system that is a product or a safety component of a product covered by the Union harmonisation legislation listed in Annex I and that requires a third-party conformity assessment. The second (Article 6(2)) covers the uses listed in Annex III: biometrics, critical infrastructure, education, employment and workers’ management, access to essential services (including creditworthiness assessment and pricing of life and health insurance), law enforcement, migration, and the administration of justice and democratic processes.
Article 6(3) allows an Annex III system to be considered not high-risk when it does not pose a significant risk of harm, for example when it performs a narrow procedural task or a preparatory task for an assessment. The exception does not apply when the system performs profiling of natural persons. A provider relying on it must document its assessment before placing the system on the market (Article 6(4)) and is subject to registration under Article 49(2). A company buying such a tool should ask the supplier for that documentation before assuming that the high-risk obligations do not apply to it.
General-purpose AI (GPAI) models are a separate category addressed to their providers. Article 53 requires them, among others, to keep technical documentation, provide information to companies integrating the model, maintain a copyright compliance policy and publish a summary of the training content. A model is considered to carry systemic risk, among others, when the cumulative computation used to train it exceeds 10^25 floating-point operations (Article 51(2)). A voluntary code of practice for GPAI models, prepared by independent experts, was published on 10 July 2025. A company using a ready-made model does not become a GPAI provider for that reason — it becomes one only when it places a model on the market itself.
What obligations does a deployer of a high-risk system have?
A deployer of a high-risk system must use it in line with the instructions for use, assign oversight to competent people, look after the quality of input data, monitor the system’s operation, keep logs and inform the people the system affects — this is the catalogue in Article 26, applicable from 2 December 2027 for Annex III systems. These obligations are narrower than a provider’s, but they require processes most companies do not have today.
Article 26(1) and (2) require technical and organisational measures ensuring that the system is used in line with its instructions, and assigning human oversight to natural persons who have the necessary competence, training and authority. In practice, this means naming a person who can challenge the system’s output and has the time to do so. Where the deployer controls the input data, it is responsible for that data being relevant and sufficiently representative (paragraph 4).
Paragraphs 5 and 6 concern operation. The deployer monitors the system’s operation on the basis of the instructions and informs the provider and the market surveillance authority when the system may present a risk. It keeps the logs automatically generated by the system, to the extent they are under its control, for a period appropriate to the system’s intended purpose and of at least six months, unless other law provides otherwise. This requirement is worth carrying over into the contract with the cloud service provider, because default log retention in a SaaS service may be shorter — check this in its documentation.
Paragraphs 7 and 11 concern informing people. Before using a high-risk system at the workplace, an employer informs workers’ representatives and the workers who will be affected. A deployer of an Annex III system that makes decisions, or assists in making decisions, about natural persons informs those persons that they are subject to such a system. Paragraph 9 links the AI Act with the GDPR: information from the provider (Article 13 of the AI Act) is used for the data protection impact assessment under Article 35 of the GDPR.
The fundamental rights impact assessment (FRIA) in Article 27 does not apply to every deployer. The obligation lies with bodies governed by public law and private entities providing public services that use high-risk systems under Article 6(2) (other than critical infrastructure in point 2 of Annex III), and with deployers of systems for assessing the creditworthiness of natural persons and for risk assessment and pricing of life and health insurance (Annex III, point 5(b) and (c)). Registration in the EU database under Article 49 falls on providers of Annex III systems and on deployers that are public authorities — a private company deploying a system usually registers nothing, but it should check whether its supplier has done so.
Key takeaways:
- Deployer obligations (Article 26) start on 2 December 2027 for Annex III systems.
- Human oversight requires a named person with the competence and authority to act.
- Logs of a high-risk system: at least six months, where they are under the company’s control.
- The Article 27 FRIA concerns the public sector, public services, credit scoring, and life and health insurance.
What transparency obligations follow from Article 50 of the AI Act?
Article 50 imposes four transparency obligations: informing people that they are interacting with an AI system, marking synthetic content in a machine-readable format, informing people about emotion recognition and biometric categorisation, and disclosing deepfake content; they apply from 2 August 2026. The obligations are split between providers and deployers, so the same company can have one of them as the provider of a chatbot and another as the publisher of marketing content.
Paragraph 1 concerns providers of systems intended to interact directly with people: a person must know that they are talking to an AI system, unless this is obvious from the point of view of a reasonably well-informed person. A company that offers a customer service chatbot on its website under its own brand is the provider in this respect and should make sure there is a clear notice in the interface. Paragraph 2 concerns providers of systems generating synthetic audio, image, video or text: outputs must be marked in a machine-readable format and be detectable as artificially generated or manipulated.
Paragraphs 3 and 4 are addressed to deployers. A deployer using an emotion recognition or biometric categorisation system informs the people it affects. A deployer that generates or manipulates image, audio or video content constituting a deepfake discloses that the content has been artificially generated or manipulated; a similar obligation covers text published to inform the public on matters of public interest. Marketing and corporate communications teams should have a simple rule for labelling such material.
The Omnibus added a transition period in Article 111(4). Providers of systems generating synthetic content that were placed on the market before 2 August 2026 have until 2 December 2026 to comply with Article 50(2). The transition period does not cover paragraphs 1, 3 and 4, so it does not postpone the obligation to inform about a chatbot or to disclose deepfakes.
Apart from that, the Omnibus changed only paragraph 7 of Article 50, which concerns codes of practice for detecting and labelling content. Infringing the transparency obligations is subject to a fine under Article 99(4), that is up to EUR 15 million or 3% of total worldwide annual turnover.
What does the AI literacy obligation in Article 4 mean?
After the amendment, Article 4 requires providers and deployers to take measures to support the development of their staff’s AI literacy and expressly states that this does not mean guaranteeing any specific level of AI literacy of any individual. The obligation has applied since 2 February 2025 because it belongs to Chapter I; its current wording was given by Regulation 2026/1744.
The change shifts the emphasis from outcome to action. A company does not have to prove that every employee has reached a specific level of knowledge, but it must be able to show what measures it has taken and why they are appropriate to the way it uses AI. The regulation also requires the Commission and the Member States to support organisations, especially SMEs, and the Commission to publish practical examples.
The measures can take different forms and do not have to mean courses. A sufficient foundation is often an AI use policy with clear rules (we describe its elements in our article on the AI acceptable use policy), instructions for specific tools, material on model limitations and a rule that outputs are verified before use. The scope of the measures is matched to role and risk: one way for a team publishing generated content, another for an HR department using a tool to assess candidates.
The evidence of meeting Article 4 is documentation: who received which material, which rules apply, when they were updated. During an inspection, what counts is whether the measures exist and are applied, not the number of hours. A complete absence of measures while AI tools are widely used will be hard to defend.
What fines does the AI Act provide for, and do SMEs pay less?
Using prohibited practices can lead to fines of up to EUR 35 million or 7% of total worldwide annual turnover, infringing other operator obligations up to EUR 15 million or 3%, and supplying incorrect information to authorities up to EUR 7.5 million or 1% (Article 99(3)–(5)); SMEs and small mid-cap enterprises pay according to the lower-of-the-two rule. For an undertaking, the higher of the two amounts applies as a rule, and the ceilings are maximums, not automatic amounts.
The category in Article 99(4) covers, among others, the obligations of providers (Article 16), deployers (Article 26) and transparency (Article 50). The Omnibus added point (da) to it, concerning the obligations in Article 25(2) and (4), that is cooperation along the value chain. Chapter XII on penalties has applied since 2 August 2025, and the fines for GPAI model providers in Article 101 since 2 August 2026.
For SMEs, Article 99(6) provides that each fine under paragraphs 3, 4 and 5 does not exceed the amount or the percentage, whichever is lower. The Omnibus added paragraph 6a for small mid-cap enterprises (SMCs), defined in Article 3(14b) by reference to Commission Recommendation (EU) 2025/1099. For SMCs, however, the lower-value rule covers only the fines in paragraphs 4 and 5, not the fines for prohibited practices in paragraph 3.
So the Omnibus did not reduce the fines; it extended the more lenient method of calculating them to another group of companies and clarified that Member States may also provide for administrative fines. In Poland, the provisions on administrative fines are in Chapter 8 of the Act on artificial intelligence systems, applicable from 28 October 2026.
What does the Polish Act on artificial intelligence systems introduce?
The Polish Act of 3 July 2026 on artificial intelligence systems (Journal of Laws 2026, item 1003) creates national supervision of the AI Act: the market surveillance authority and single point of contact is the Commission for the Development and Security of Artificial Intelligence (Komisja Rozwoju i Bezpieczeństwa Sztucznej Inteligencji). The Act was published in the Journal of Laws on 27 July 2026. The name of the Commission is sometimes abbreviated to “KRiBSI”, but the Act itself uses the term “the Commission”.
The Act enters into force in stages, in line with Article 127. As a rule, it entered into force 14 days after publication, that is on 11 August 2026. Article 125(4) has applied from the day after publication, that is from 28 July 2026. Articles 8–18 and Chapters 3–5, 8 and 9 entered into force three months after publication, that is on 28 October 2026.
The titles of the chapters applicable from 28 October 2026 show what to expect: Chapter 3 is “Monitoring compliance with the provisions on artificial intelligence”, Chapter 4 “Proceedings before the market surveillance authority”, Chapter 5 “Agreement on the conditions for extraordinary mitigation of sanctions”, Chapter 8 “Provisions on administrative fines and criminal provisions”, and Chapter 9 “Amendments to provisions”. From 28 October 2026, the national provisions on inspections, proceedings before the authority and penalties therefore apply.
The Sejm appointed the Chair of the Commission by a resolution of 18 September 2026 (Monitor Polski 2026, item 945), with the consent of the Senate given on 24 September 2026 (Monitor Polski 2026, item 946). The Act also amends other acts, including the Act on the national cybersecurity system and the Act on the protection of personal data, and in the Sejm metadata it is linked to the AI Act, the Digital Services Act and the GDPR.
For companies, the conclusion is practical: the AI Act is no longer only a regulation “from Brussels”. Compliance with the obligations that already apply — the prohibited practices, Article 4 and, since August 2026, also Article 50 — is checked by a national authority with its own procedure. The broader regulatory background for the legal sector is covered in our article on AI, GDPR and ethics in law firms.
What should a company do in 2026 to prepare for the AI Act?
In 2026, it is worth building an AI system register with the company’s role for each system, checking uses against Article 5, documenting the Article 4 measures, organising content labelling under Article 50 and identifying the systems that will enter the high-risk regime in 2027. The order matters: without a register you cannot determine the role, the risk category or the deadline.
The register should also cover AI features built into SaaS applications, browser extensions and tools used by employees without formal approval. That last group is covered in a separate text on shadow AI, and the structure of roles and the approval process for new tools in our article on the AI governance framework. A broader view of the technical safeguards of an AI environment is given by the map of AI security risks and controls.
The table below is an action plan ordered by deadline. The “evidence” column answers the question that will be asked during an inspection: what can the company show?
| Action | Who it concerns | Deadline (as of 10 November 2026) | Evidence of application |
|---|---|---|---|
| AI system register with the company’s role (provider, deployer, importer, distributor) | all companies using AI | now — a precondition for the other steps | register with the owner, purpose and supplier of each system |
| Review of uses against Article 5 | all companies | applies since 2 February 2025; new points (ba) and (bb) from 2 December 2026 | assessment note for systems in risk areas |
| Measures supporting AI literacy under Article 4 | providers and deployers | applies since 2 February 2025 | AI policy, tool instructions, list of recipients and update dates |
| Notice of interaction with AI (Article 50(1)) | providers of chatbots and assistants offered to customers | from 2 August 2026 | interface screenshot and description of the notice |
| Disclosure of deepfakes (Article 50(4)) | companies publishing content generated or altered by AI | from 2 August 2026 | labelling rule in the publication procedure |
| Marking of synthetic content (Article 50(2)) | providers of generative systems | from 2 August 2026; systems placed on the market earlier by 2 December 2026 | description of the machine-readable marking mechanism |
| List of systems in Annex III areas (e.g. recruitment, scoring) and verification of the Article 6(3) exception | deployers and providers | preparation in 2026; obligations from 2 December 2027 | classification sheet and supplier documentation |
| Contracts with suppliers: instructions, logs, division of obligations (Article 26(1) and (6)) | deployers of high-risk systems | before 2 December 2027 | contract clauses, log retention configured for at least six months |
| Procedure for informing workers and the persons affected by decisions (Article 26(7) and (11)) | employers and deployers of Annex III systems | before 2 December 2027 | notice template, confirmation of consultation with workers’ representatives |
| FRIA (Article 27) | public bodies, public services, credit scoring, life and health insurance | before the first use of the system, from 2 December 2027 | FRIA report linked to the DPIA under Article 35 of the GDPR |
The plan can be delivered without a large project if the company builds it on its existing information security management system. The AI system register can sit next to the asset register, and the assessment of uses can follow the same cycle as the risk analysis. Organisations that want a management system framework for AI can use the standard described in our article on ISO/IEC 42001; the AI Act does not require it.
The biggest risk for 2027 concerns HR tools. Systems for recruitment, candidate selection and employee assessment fall under point 4 of Annex III, and they are sometimes bought without the compliance team’s involvement. If such a tool is already in use, it is worth establishing now whether the supplier is preparing the documentation, logs and instructions required from December 2027.
How can nFlo help you prepare for the AI Act?
nFlo provides an AI Act readiness and AI governance service that covers the inventory of AI systems, their classification, determining the organisation’s role, a gap analysis against the deadlines after the Digital Omnibus and an AI governance model. We work within the governance, risk and compliance pillar, connecting the AI Act with the GDPR, NIS2 and ISO/IEC 27001, which companies already know.
The service does not replace legal advice in an individual case. Its outcome is an organised picture: which AI systems operate in the company, which of them fall under which provisions and what needs to be done by when. The broader context of AI security and compliance is described on the AI security page, and articles in this area are collected in the AI security knowledge base.
If you want to establish which deadlines on this timeline apply to your organisation, contact us.
What else is worth knowing about the AI Act timeline?
Below are answers to the questions that come up most often when planning. The definition of the act itself and its basic concepts are in the AI Act glossary entry.
Do the requirements for high-risk AI systems apply from 2 August 2026?
No. After Regulation (EU) 2026/1744 amended Article 113, the requirements in Chapter III, Sections 1–3 apply from 2 December 2027 to systems listed in Annex III and from 2 August 2028 to systems covered by Annex I. 2 August 2026 remains the general date of application, from which Article 50, among others, applies (as of 10 November 2026).
Is a company that only uses ChatGPT or Copilot subject to the AI Act?
Yes, as a deployer, although to a narrower extent than a provider. Article 4 applies to it (measures supporting the AI literacy of staff), Article 50 applies in specific situations (for example disclosing deepfake content), and the Article 26 obligations arise only when the tool serves a purpose listed in Annex III, such as screening job candidates.
Who supervises the application of the AI Act in Poland?
The Act of 3 July 2026 on artificial intelligence systems (Journal of Laws 2026, item 1003) designates the Commission for the Development and Security of Artificial Intelligence as the market surveillance authority and the single point of contact. The Sejm appointed the Chair of the Commission by a resolution of 18 September 2026 (Monitor Polski 2026, item 945) with the consent of the Senate.
Did the Digital Omnibus reduce the AI Act fines?
No. The maximum amounts in Article 99(3)–(5) are unchanged: EUR 35 million or 7% of turnover for prohibited practices, EUR 15 million or 3% for other obligations and EUR 7.5 million or 1% for incorrect information. New Article 99(6a) applies the lower-of-the-two rule to small mid-cap enterprises, but only to the fines in paragraphs 4 and 5.
Does the AI Act require ISO/IEC 42001 certification?
No. The AI Act does not impose an obligation to certify against ISO/IEC 42001. The standard describes an artificial intelligence management system and can be a convenient framework for the roles, policies, risk assessment and reviews on which a company bases ongoing compliance, but adopting it is the organisation’s decision, not a requirement of the regulation.
