EU AI Act Compliance: Who It Applies To, What the Digital Omnibus Changed and What to Do This Quarter
Most companies meet the AI Act as users of AI tools: Copilot, ChatGPT, AI features inside their SaaS. For them it asks three things today: AI literacy measures (Art. 4), no prohibited practices (Art. 5) and, where triggered, transparency (Art. 50). The high-risk rules wait until 2 December 2027. This page maps each duty to its article and its post-Omnibus date, and shows where your NIS2 and GDPR work carries over.
General information based on the Official Journal texts, not legal advice. National law sets the penalties.
The EU AI Act at a glance
The AI Act is a regulation: it applies directly in every Member State, and national law adds the penalty rules (Art. 99(1)). Article 2 has no size threshold. Being small changes your fine ceiling, not whether you are covered.
What the Digital Omnibus changed, and what an SME deployer has to do this quarter
Regulation (EU) 2026/1744, the Digital Omnibus on AI, was published on 24 July 2026 and entered into force on 27 July 2026 (its Article 4). Older guidance usually gets the high-risk date or the Article 4 wording wrong. What changed for a company that uses AI:
| What changed | Before | Now | For a deployer |
|---|---|---|---|
| High-risk date, Annex III | 2 August 2026 | 2 December 2027 (Art. 113(c)(i)) | No Art. 26 duties yet |
| High-risk date, Annex I products | 2 August 2027 | 2 August 2028 (Art. 113(c)(ii)) | Matters to manufacturers |
| AI literacy, Art. 4 | “ensure, to their best extent, a sufficient level of AI literacy” | “take measures to support the development of AI literacy” | A documented programme, no guaranteed level |
| Prohibited practices, Art. 5 | Points (a) to (h) | Plus (ba) and (bb) from 2 December 2026 | Two more screening questions |
| Marking, Art. 50(2) | From 2 August 2026 | Systems on the market before 2 August 2026: by 2 December 2026 (Art. 111(4)) | Provider duty only |
| Fines, small mid-caps | No relief | Lower of two figures for Art. 99(4) and (5) fines (Art. 99(6a)) | Lower ceilings outside Art. 5 |
The takeaway: the Omnibus moved the heavy dates, not Articles 4, 5 or 50. Those three already apply.
What to do before 2 December 2026
If you only use AI tools, and none for a high-risk purpose, the list is short:
- List every AI system in use, SaaS features included, with your role for each (Art. 3(3), (4)).
- Screen each one against Article 5, including the new points (ba) and (bb). A deployer is caught by those only “where the deployer uses the system for the purpose of generating or manipulating such material or performance” (Art. 5(1a)(b)).
- Record your AI literacy measures: who uses what, what they learned, when it is refreshed (Art. 4(1)).
- Check your Article 50 triggers: deep fakes, and AI-generated text published to inform the public on matters of public interest (Art. 50(4)).
- Built your own chatbot or generative feature and offered it before 2 August 2026? You are probably its provider, and marking its output is due by 2 December 2026 (Art. 50(2), Art. 111(4)).
- Flag any Annex III use (recruitment, worker evaluation, credit scoring) and open its high-risk file for 2 December 2027.
None of this needs a certificate (see our ISO 42001 and AI Act analysis).
Who the AI Act applies to: provider, deployer, importer, distributor
Article 2(1) covers providers placing AI on the EU market wherever they are based, deployers in the Union, importers, distributors, and providers and deployers elsewhere “where the output produced by the AI system is used in the Union”. That last rule can bring a UK MSP with EU clients into scope. Duties follow the role:
| Role | Definition (Art. 3) | Typical case | Main duties |
|---|---|---|---|
| Provider (Art. 3(3)) | Develops an AI system or model, or has one developed, and places it on the market or puts it into service “under its own name or trademark, whether for payment or free of charge” | Software house shipping an AI feature; MSP running its own chatbot on a model API | Art. 50(1), (2); Arts 16–25 if high-risk |
| Deployer (Art. 3(4)) | Uses an AI system “under its authority except where the AI system is used in the course of a personal non-professional activity” | Any firm whose staff use Copilot or AI in SaaS | Arts 4, 5, 50(3), (4); Art. 26 if high-risk |
| Importer (Art. 3(6)) | EU-based; places on the market an AI system bearing the name or trademark of a non-EU person | EU firm launching a non-EU vendor’s product | Art. 23 (high-risk) |
| Distributor (Art. 3(7)) | Makes an AI system available on the EU market, other than the provider or importer | Reseller or MSP reselling an AI product | Art. 24 (high-risk) |
Roles move. Under Article 25(1) a deployer, importer or distributor becomes the provider of a high-risk system if it rebrands one, substantially modifies one, or repurposes a system so that it becomes high-risk, say a general-purpose chatbot used to filter job applications.
A chatbot you built yourself can make you a provider without selling anything, because putting into service includes supply “for own use” (Art. 3(11)). Where configuring a vendor’s product ends and building your own begins is a question of fact; record your reading.
Outside the Act: purely personal use (Art. 2(10)), AI built solely for scientific research (Art. 2(6)), pre-market research and testing other than real-world testing (Art. 2(8)), and open-source AI unless it is high-risk or falls under Article 5 or 50 (Art. 2(12)).
AI Act dates after the Digital Omnibus
Article 113 sets the dates; Regulation (EU) 2026/1744 rewrote its third paragraph (Art. 1, point (40)). The general date is 2 August 2026, and everything else is an exception to it.
| Date | What applies | Legal basis |
|---|---|---|
| 2 February 2025 | Scope, definitions, AI literacy (Art. 4), prohibitions (Art. 5) | Art. 113(a) |
| 2 August 2025 | General-purpose AI model rules, governance, penalties chapter (not Art. 101) | Art. 113(b) |
| 27 July 2026 | Omnibus in force: new Art. 4 wording, SMC fine rule (Art. 99(6a)) | Reg. 2026/1744, Art. 4 |
| 2 August 2026 | General date: Art. 50, market surveillance and enforcement (Arts 74–94), Art. 101 | Art. 113, second paragraph |
| 2 December 2026 | Art. 5(1)(ba), (bb); Art. 50(2) marking for systems placed on the market before 2 August 2026 | Art. 113(a); Art. 111(4) |
| 2 August 2027 | Models placed on the market before 2 August 2025 must comply | Art. 111(3) |
| 2 December 2027 | High-risk rules (Chapter III, Sections 1–3, incl. Arts 26, 27) for Annex III | Art. 113(c)(i) |
| 2 August 2028 | High-risk rules for AI under Annex I | Art. 113(c)(ii) |
| 2 August 2030 | High-risk systems for public authorities, even if already on the market | Art. 111(2) |
Two readings sit behind the table. A high-risk system already on the market before its date is caught only after a significant design change, though Article 5 still applies (Art. 111(2)). And Articles 40 to 49, 72 and 73 are not postponed in the words of Article 113, but they attach to high-risk systems, so we read them as starting on the high-risk dates. The EU cyber compliance deadline tracker keeps these dates next to NIS2, the CRA and DORA.
AI literacy under Article 4: the duty every company carries
Article 4 binds every provider and deployer, whatever the risk class. Since 27 July 2026, paragraph 1 reads:
Providers and deployers of AI systems shall take measures to support the development of AI literacy of their staff and other persons dealing with the operation and use of AI systems on their behalf, taking into account their technical knowledge, experience, education and training and the context the AI systems are to be used in, and considering the persons or groups of persons on whom the AI systems are to be used. This obligation does not require providers or deployers to guarantee any specific level of AI literacy of any individual.
The old “sufficient level” standard is gone; adequate measures are not defined, and the Commission is to publish practical examples (Art. 4(2)). In practice: a role-based programme and a completion record, covering contractors who operate AI for you. Article 4 is not in the fine lists of Article 99(3) to (5), so any sanction comes from national rules under Article 99(1).
A trap: 2025 policies still promise a “sufficient level”, more than the law now asks. Our generative AI acceptable use policy guide places the literacy clause next to your NIS2 training duties.
Prohibited AI practices under Article 5
Article 5 covers placing on the market, putting into service and use, so deployers are caught too. An IT firm should check point (f), emotion recognition at work, first.
| Point | Prohibited practice (summary) | Where it could come up (our examples) |
|---|---|---|
| (a), (b) | Manipulative or deceptive techniques, or exploiting vulnerabilities, that materially distort behaviour and cause significant harm | Dark patterns in AI sales or support tools |
| (ba), from 2 Dec 2026 | Realistic intimate or sexually explicit material of an identifiable person without explicit consent | Image and video generators |
| (bb), from 2 Dec 2026 | Material within Art. 2(c) and (e) of Directive 2011/93/EU (child sexual abuse material) | Image and video generators |
| (c) | Social scoring leading to unjustified or unrelated detrimental treatment | Scoring staff or customers |
| (d), (e), (h) | Crime prediction from profiling alone; untargeted facial scraping; real-time remote biometric identification for law enforcement | Rare outside law-enforcement vendors |
| (f) | Inferring emotions at the workplace or in education, except for medical or safety reasons | Sentiment tools aimed at staff; interview analysis |
| (g) | Biometric categorisation to infer race, political opinions, union membership, beliefs, sex life or sexual orientation | Analytics on CCTV or staff photos |
For the new points, Article 5(1a) narrows the ban: a provider is caught where that output is the intended purpose or a foreseeable, reproducible outcome without adequate safeguards; a deployer only where it uses the system for that purpose. Breaches carry up to EUR 35 000 000 or 7 % of total worldwide annual turnover, whichever is higher (Art. 99(3)). A dated screening record per system is the evidence to keep.
Transparency under Article 50: chatbots, deep fakes and AI-generated text
Article 50 has applied since 2 August 2026, and the grace period to 2 December 2026 covers only paragraph 2.
| Provision | Who | Duty | Applies from |
|---|---|---|---|
| Art. 50(1) | Provider | Systems that interact with people tell them they are dealing with AI, unless obvious | 2 Aug 2026 |
| Art. 50(2) | Provider | Machine-readable marking of synthetic audio, image, video or text | 2 Aug 2026; 2 Dec 2026 for systems placed on the market earlier (Art. 111(4)) |
| Art. 50(3) | Deployer | Inform people exposed to emotion recognition or biometric categorisation | 2 Aug 2026 |
| Art. 50(4) | Deployer | Disclose deep fakes; disclose AI text published to inform the public on matters of public interest, unless human-reviewed with editorial responsibility | 2 Aug 2026 |
Notices are due “at the latest at the time of the first interaction or exposure” (Art. 50(5)). For an MSP the line runs between a vendor’s chatbot you switch on (deployer) and one you build on a model API under your own name (possibly provider, with Art. 50(1) and (2)). Breaches fall under the EUR 15 000 000 or 3 % ceiling (Art. 99(4)(g)).
High-risk AI: Article 6, Annex III and the deployer’s duties
A system is high-risk if it is, or is a safety component of, an Annex I product needing third-party conformity assessment (Art. 6(1)), or if it falls in an Annex III area (Art. 6(2)):
| Annex III area | Uses listed (summary) |
|---|---|
| 1. Biometrics | Remote identification (not verification), categorisation by sensitive attributes, emotion recognition |
| 2. Critical infrastructure | Safety components in critical digital infrastructure, road traffic, water, gas, heating, electricity |
| 3. Education | Admission, assessing learning outcomes, monitoring tests |
| 4. Employment | Recruitment, filtering applications, promotion and termination, task allocation, performance monitoring |
| 5. Essential services | Public benefits, creditworthiness (not fraud detection), life and health insurance pricing, emergency triage |
| 6–8. Public functions | Law enforcement, migration and border control, justice and elections |
For most IT firms, area 4 is the realistic one. An Annex III system escapes only if it poses no significant risk and performs a narrow procedural task, improves a completed human activity, detects deviations without replacing human assessment, or does preparatory work (Art. 6(3)); never if it profiles people. The provider documents that assessment and registers the system (Art. 6(4), Art. 49(2)).
Deployer duties (Art. 26) and the FRIA (Art. 27)
- use the system according to its instructions (Art. 26(1));
- assign oversight to people with “the necessary competence, training and authority, as well as the necessary support” (Art. 26(2));
- keep input data you control relevant and sufficiently representative (Art. 26(4));
- monitor; on a risk, inform the provider and the authority and suspend use; on a serious incident, “immediately inform first the provider” (Art. 26(5));
- keep logs under your control for at least six months (Art. 26(6));
- inform workers’ representatives and affected workers before workplace use (Art. 26(7));
- tell people when an Annex III system makes or assists decisions about them (Art. 26(11)).
A fundamental rights impact assessment is needed only by public bodies, private entities providing public services and deployers of credit-scoring or life and health insurance systems (Annex III, 5(b), (c)); area 2 is excluded (Art. 27(1)). It is made before first use and its results go to the market surveillance authority (Art. 27(2), (3)); it may cross-refer to your GDPR DPIA (Art. 27(4)). Article 26 is fined under Article 99(4)(e) at up to EUR 15 000 000 or 3 %; Article 27 is not listed.
All of this starts on 2 December 2027. For the NIS2 side, read our guide to dual NIS2 and AI Act obligations of high-risk deployers and the Annex III and NIS2 sector matrix. Providers carry Articles 9 to 15 and a quality management system (Art. 17); its harmonised standard, EN 18286:2026, was cited in the Official Journal by Implementing Decision (EU) 2026/2274 on 14 October 2026.
General-purpose AI models, in short
These rules have applied since 2 August 2025 (Art. 113(b)) and bind model makers, who must keep technical documentation for the AI Office, give integrators information on the model, have a copyright policy and publish a training-content summary (Art. 53(1)). Models trained with more than 1025 floating-point operations are presumed to carry systemic risk, with extra duties (Art. 51(2), Art. 55(1)). Older models have until 2 August 2027 (Art. 111(3)).
Build on a model API and you are a downstream provider of your own system (Art. 3(68)). The Article 53(1)(b) information the model provider owes you belongs in your NIS2 Article 21(2)(d) supplier file.
Penalties under Article 99
Member States set the rules, which “may also include administrative fines, warnings and non-monetary measures”, and must weigh the economic viability of SMEs and small mid-caps (Art. 99(1)). For everyone else the higher of the two figures is the ceiling:
| Infringement | Maximum fine | SMEs | Small mid-caps |
|---|---|---|---|
| Prohibited practices (Art. 5) | EUR 35 000 000 or 7 % (Art. 99(3)) | Lower (Art. 99(6)) | Higher: Art. 99(6a) skips para. 3 |
| Provider, importer, distributor, deployer duties; Art. 25(2), (4); Art. 50 | EUR 15 000 000 or 3 % (Art. 99(4)) | Lower | Lower (Art. 99(6a)) |
| Misleading information to bodies or authorities | EUR 7 500 000 or 1 % (Art. 99(5)) | Lower | Lower (Art. 99(6a)) |
| AI literacy (Art. 4), FRIA (Art. 27) | No EU ceiling (Art. 99(1)) | National rules | National rules |
An SME is defined by Recommendation 2003/361/EC (Art. 3(14a)). A small mid-cap is not an SME, employs fewer than 750 persons and has turnover up to EUR 150 million or a balance sheet up to EUR 129 million (Art. 3(14b), Recommendation (EU) 2025/1099). The Commission fines model providers itself, up to 3 % or EUR 15 000 000, whichever is higher (Art. 101(1)). Compare NIS2’s fine ceilings.
How the AI Act overlaps with NIS2, GDPR and the CRA
NIS2 Article 21 is technology-neutral: its measures cover the network and information systems an entity uses, AI tools included. The AI Act adds role-specific duties on top, and GDPR keeps applying alongside it (AI Act Art. 2(7)).
| Topic | NIS2 | GDPR | CRA | AI Act |
|---|---|---|---|---|
| Risk | Risk analysis, all-hazards (Art. 21(2)(a)) | DPIA (Art. 35) | Essential requirements (Annex I) | Art. 9 (providers); FRIA (Art. 27) |
| Incidents | Early warning within 24 hours (Art. 23(4)(a)) | Breach notice within 72 hours where feasible (Art. 33(1)) | Art. 14, since 11 Sep 2026 | Provider: 15 days, 2 for critical infrastructure, 10 for a death (Art. 73); deployer tells provider first (Art. 26(5)) |
| Suppliers | Supply chain security (Art. 21(2)(d)) | Processors (Art. 28) | — | Supplier agreements (Art. 25(4)); model information (Art. 53(1)(b)) |
| Cybersecurity | Art. 21 measures | Art. 32 | CRA Art. 12(1) met = AI Act Art. 15 met | Art. 15; Art. 42(3) |
| Training | Arts 20(2), 21(2)(g) | — | — | Art. 4 |
The takeaway: a NIS2 programme already holds the risk register, supplier file and training records the AI Act leans on. What is usually missing: role classification, the Article 5 screening, the Article 4 record and Article 50 notices. See NIS2 for MSPs and MSSPs, supply chain security, NIS2 vs GDPR and DPIA vs NIS2 risk assessment.
Manufacturers have a second overlap: the CRA applies from 11 December 2027, and its Article 12 lets a high-risk AI product meet the AI Act’s cybersecurity requirement through the CRA (CRA compliance guide, NIS2 vs CRA for manufacturers, NIS2 for software vendors). On standards: ISO/IEC 42001 is not a harmonised standard cited in the Official Journal, so it gives no presumption of conformity (Art. 40(1)); compare it with the NIST AI RMF in NIST AI RMF vs ISO 42001 vs the EU AI Act.
What the EU AI Act Compliance Pack contains
Written from the Official Journal texts of the AI Act and the Omnibus: 31 editable files (6 guides and 25 documents; 27 Word, 4 Excel). A 10-minute route places you in one of five profiles; most companies that use AI tools need six core documents, seven if they run a chatbot or publish AI-generated content.
The Excel workbooks were tested by recalculating their formulas against written test vectors in LibreOffice; no macros, sheets not protected. 34 interpretation points give the pack’s default reading where the law is unclear. Not legal advice, no guarantee of compliance.
One-time payment · single legal entity licence · 30-day update-or-add pledge: if a template does not fit your situation, email us within 30 days and we will update it or add what is missing.
Frequently asked questions
Does the EU AI Act apply to a company that only uses ChatGPT or Copilot?
Yes, as a deployer (Art. 3(4)): AI literacy measures (Art. 4), no prohibited practices (Art. 5) and Art. 50(3) or (4) disclosures where triggered. High-risk duties apply only to Annex III uses, from 2 December 2027.
When do the AI Act high-risk rules apply after the Digital Omnibus?
From 2 December 2027 for Annex III systems and 2 August 2028 for AI under Annex I (Art. 113(c), as amended by Regulation (EU) 2026/1744).
Is there a deadline for AI literacy?
Article 4 has applied since 2 February 2025, in new wording since 27 July 2026: providers and deployers “take measures to support the development of AI literacy”, without guaranteeing any specific level.
What changes on 2 December 2026?
The prohibitions in Art. 5(1)(ba) and (bb) apply, and providers of generative systems placed on the market before 2 August 2026 must mark their output (Art. 50(2), Art. 111(4)).
What are the fines under the EU AI Act?
Up to EUR 35 000 000 or 7 % for Article 5 and EUR 15 000 000 or 3 % for most other duties, Articles 26 and 50 included. SMEs face the lower figure (Art. 99(6)); small mid-caps too, except for Article 5 (Art. 99(6a)).
Does a UK company have to comply?
Yes, if it places AI systems on the EU market or the output of its AI system is used in the Union (Art. 2(1)(a), (c)).
Sources
- European Parliament and Council. Regulation (EU) 2024/1689 (Artificial Intelligence Act). OJ L, 2024/1689, 12.7.2024. EUR-Lex
- European Parliament and Council. Regulation (EU) 2026/1744 (Digital Omnibus on AI). OJ L, 2026/1744, 24.7.2026. EUR-Lex
- European Commission. Implementing Decision (EU) 2026/2274 on harmonised standards for high-risk AI systems (EN 18286:2026). OJ L, 2026/2274, 14.10.2026. EUR-Lex
- European Parliament and Council. Directive (EU) 2022/2555 (NIS2 Directive). EUR-Lex
- European Parliament and Council. Regulation (EU) 2016/679 (GDPR). EUR-Lex
- European Parliament and Council. Regulation (EU) 2024/2847 (Cyber Resilience Act). EUR-Lex
- European Commission. Recommendation (EU) 2025/1099 on the definition of small mid-cap enterprises. EUR-Lex
The AI Act is quoted as amended by Regulation (EU) 2026/1744 and checked against the Official Journal texts; no English corrigendum to either regulation had been published when this page was written. Commission guidelines and codes of practice are not cited as obligations. General information, not legal advice; not affiliated with any EU institution.