EU AI Act: 8 Official Sources + Practical Tools 2026

Eight addresses, four tools. This article goes through the eight sources one by one: what a compliance officer opens them for, which provision carries the obligation behind them, and how to cite them. Three of the eight addresses are evaluated with their own entry in the sources block (EUR-Lex, AI Office, GPAI Code of Practice); for the remaining five this article names the point of contact and the provision that carries the obligation, but no individual document — scope and reason in the sources block. "Official" in the title refers to the EU and public-authority addresses among the eight; only EUR-Lex carries the obligation in an audit, the others supply interpretation, means of demonstrating compliance, working hypotheses or procedural status — the evidentiary status and character as proof of each source are set out in the table in the section "Citation practice". The factual status of the source review is 2 May 2026, updated to the legal status of 17 August 2026 for the legal acts named in the sources block. The practical part is covered by the four tools at the end of the page: quick test, fine calculator, self-assessment and AI inventory quick check.

1. EUR-Lex Regulation 2024/1689

Short answer: EUR-Lex carries the binding text of Regulation (EU) 2024/1689 with articles, annexes and recitals — every obligation under the Regulation itself can be traced back to a reference in precisely this document; national deviations are covered in the section "Germany and Austria".

The consolidated version is what counts: for high-risk systems under Annex III AI Act, Chapter III Sections 1–3 apply, since Regulation (EU) 2026/1744 (in force since 27 July 2026), from 2 December 2027, and for Annex I systems from 2 August 2028 — check against the consolidated version of Art. 113 AI Act before use.

eur-lex.europa.eu — official regulation text in 24 languages.

Six places are the ones you will need most often in daily work: the definitions in Art. 3 no. 1 AI Act (AI system) and Art. 3 no. 63 AI Act (general-purpose AI model), the prohibited practices in Art. 5(1) AI Act, the high-risk classification in Art. 6(1)–(3) AI Act together with the eight areas of Annex III AI Act, the transparency obligations in Art. 50 AI Act and the fine ranges in Art. 99(3)–(5) AI Act. The recitals explain the regulatory logic: recital 26 describes the risk-based approach, under which the type and content of the rules are tailored to the intensity and scope of the risks.

The fine ranges are set out in three tiers in Art. 99(3)–(5) AI Act: infringements of the prohibition on AI practices under Art. 5 AI Act up to EUR 35 million or — in the case of undertakings — up to 7 % of total worldwide annual turnover for the preceding financial year (paragraph 3); infringements of certain other obligations, among them those of providers under Art. 16 AI Act, of deployers under Art. 26 AI Act and of the transparency obligations under Art. 50 AI Act, up to EUR 15 million or 3 % (paragraph 4); the supply of incorrect, incomplete or misleading information to authorities up to EUR 7.5 million or 1 % (paragraph 5) — in each case whichever amount is higher.

The dates are not scattered through the text but bundled in Art. 113 AI Act. The Annex III and Annex I dates — shown in the "Caveat" column of rows 3 and 4 — have been recast by Regulation (EU) 2026/1744 (Digital Omnibus AI), in force since 27 July 2026: the high-risk obligations of Chapter III Sections 1–3 apply to Annex III systems from 2 December 2027 and to Annex I systems from 2 August 2028. The general date of application, 2 August 2026, has passed and remains authoritative, among other things for Art. 50 AI Act; Art. 27 AI Act sits in Chapter III Section 3 and therefore applies to Annex III systems from 2 December 2027 (section 8). The factual status of the dates in this table is 17 August 2026 — check every row against the consolidated version of Art. 113 AI Act before use:

That is precisely why the consolidated version matters more than any summary: the dates have been changed by Regulation (EU) 2026/1744 — Annex III from 2 December 2027, Annex I from 2 August 2028 — and the consolidation on EUR-Lex is what evidences the wording Art. 113 AI Act carries today. Practitioner recommendation: for quickly locating the text of an article, the AI Act Service Desk (sources block) is usable.

2. EU AI Office FAQs

Short answer: for general-purpose AI models, the AI Office (EU AI Office) is the competent body: it provides the template for the summary of training content (Art. 53(1)(d) AI Act), is the address for reporting serious incidents where there is systemic risk (Art. 55(1)(c) AI Act) and has published the GPAI Code of Practice; the Commission confers implementation for these models on it (Art. 88(1) AI Act).

digital-strategy.ec.europa.eu — practical FAQs.

Two further points of contact are added: the Office conducts investigations (Art. 92 AI Act), and since 2 August 2026 the Commission's power to impose fines on GPAI providers applies in addition (Art. 101 AI Act: up to EUR 15 million or 3 % of worldwide annual turnover, whichever amount is higher). The scientific panel of independent experts acts in an advisory capacity (Art. 68 AI Act).

3. BSI: the technical side of Art. 15 AI Act and the second obligation regime, NIS2

Short answer: the BSI is the German address for the technical side of what Art. 15 AI Act requires in law for high-risk AI systems — accuracy, robustness and cybersecurity — and at the same time the supervisory authority for the second obligation regime under the NIS2 Directive (Directive (EU) 2022/2555), which can apply independently of the AI Act. As an authority, the Federal Office for Information Security (BSI) is the central body for information security at national level in Germany and a higher federal authority within the remit of the Federal Ministry of the Interior, under the German BSI Act (§ 1 BSIG 2025).

bsi.bund.de — the German view of AI security + minimum standards.

What Art. 15 AI Act requires technically. High-risk AI systems must achieve an appropriate level of accuracy, robustness and cybersecurity and perform consistently in those respects throughout their lifecycle (Art. 15(1) AI Act). Three requirements follow from this in concrete terms. First, accuracy: the levels of accuracy and the relevant accuracy metrics must be declared in the instructions for use (Art. 15(3) AI Act). Second, robustness: resilience against errors, faults and inconsistencies, achievable among other things through technical redundancy including backup or fail-safe plans; systems that continue to learn must additionally eliminate or minimise the risk of biased outputs influencing future operations — feedback loops (Art. 15(4) AI Act). Third, cybersecurity: resilience against unauthorised third parties; the technical solutions include, as appropriate, measures against manipulation of the training data set (data poisoning), against manipulation of pre-trained components (model poisoning), against inputs designed to cause the model to make a mistake (adversarial examples or model evasion), as well as against attacks on confidential data and the exploitation of model flaws (Art. 15(5) AI Act).

For general-purpose AI models with systemic risk, the cybersecurity obligation under Art. 55(1)(d) AI Act is added (section 6). This article does not name any individual BSI document — the source review of 2 May 2026 evaluated no BSI publication (sources block). Therefore record the title, version and date of access of the publication you rely on yourself.

Second obligation regime: NIS2. In Austria, responsibility under this regime lies with the Cybersecurity Authority (Cybersicherheitsbehörde, Austria) under the Austrian Network and Information Systems Security Act 2026 (§ 3a NISG 2026). Whether NIS2 applies to your organisation is determined in principle by the type of entity under Annexes 1 and 2 together with the size thresholds — in Germany under § 28 BSIG 2025, in Austria under § 24 in conjunction with § 25 NISG 2026. The figures behind this: in Germany, other entities of a type listed in Annex 1 qualify as a particularly important entity from 250 employees or from an annual turnover above EUR 50 million together with an annual balance sheet total above EUR 43 million; from 50 employees or from an annual turnover and an annual balance sheet total each above EUR 10 million, an entity of a type listed in Annex 1 or Annex 2 qualifies as an important entity, to the extent it is not already a particularly important entity (§ 28(1) and (2) BSIG 2025). Austria uses the same figures for large and medium-sized undertakings (§ 25(2) and (3) NISG 2026). Irrespective of these thresholds, operators of critical installations under § 28(1) BSIG 2025 qualify in Germany as a particularly important entity. Covered irrespective of size are, among others, qualified trust service providers, TLD name registries, DNS service providers and certain providers of public telecommunications networks and services; entities of the federal administration are covered separately in Germany by § 29 BSIG 2025. Practitioner recommendation: anyone subject to NIS2 under these thresholds should first check against the text of the legislation which requirements overlap with Art. 15 AI Act, and document identical measures in both chains of evidence with separate references — once as a requirement of the AI Act, once under § 30 BSIG 2025 or § 32 NISG 2026. The NIS2 catalogues expressly name, among other things, supply chain security, business continuity and multi-factor authentication (§ 30(2) BSIG 2025, § 32(4) NISG 2026).

4. CNIL AI recommendations

Short answer: this source leads into the data protection layer alongside the AI Act: where an AI system processes personal data, the processing needs a legal basis under Art. 6(1) GDPR — in addition to the classification under the AI Act.

cnil.fr — the French data protection authority publishes practical guidance, often relevant EU-wide.

Four references are relevant in practice: the data protection impact assessment where processing is likely to result in a high risk, in particular where new technologies are used and in the case of a systematic and extensive evaluation of personal aspects (Art. 35(1) and (3) GDPR), the general prohibition on processing special categories (Art. 9(1) GDPR), the one-month response period for data subject requests, extendable by two months (Art. 12(3) GDPR), and the processing agreement with a prescribed minimum content where external AI services are used (Art. 28(3) GDPR).

For processing in Germany and Austria, supervision is governed nationally: in Germany, the authorities competent under Land law supervise non-public bodies (§ 40(1) BDSG) under the German Federal Data Protection Act (BDSG), and the Federal Commissioner for Data Protection (BfDI) supervises federal bodies (§ 9(1) BDSG); in Austria it is the Austrian Data Protection Authority (Datenschutzbehörde) under the Austrian Data Protection Act (§ 18(1) DSG). Practitioner recommendation: record the CNIL paper with title, version and date of access in the inventory field "interpretative guidance" for the system concerned, and evidence the obligation separately via Art. 6(1) GDPR.

5. AI Act Compass: applicability test and the four assessment steps behind it

Short answer: an applicability test answers the entry question and provides a working hypothesis — you then document the classification itself, as a practitioner recommendation, with the references on which it rests, in your own AI inventory.

aiact-compass.de — free online test 'Am I affected?'.

Behind every classification there are four assessment steps that you should keep traceable independently of the tool: first, the question of a prohibited practice (Art. 5(1) AI Act). Second, allocation to one of the eight areas of Annex III AI Act — biometrics, critical infrastructure, education, employment and workforce management, essential private and public services, law enforcement, migration, asylum and border control, administration of justice and democratic processes. Third, the narrowly construed exception in Art. 6(3) AI Act, which does not apply in any event where natural persons are profiled. Fourth, the role: provider (Art. 3 no. 3 AI Act) or deployer (Art. 3 no. 4 AI Act).

How the demonstration then runs depends on that fourth question. For providers (Art. 3 no. 3 AI Act) it runs through the technical documentation under Annex IV AI Act (Art. 11 AI Act), the conformity assessment (Art. 43 AI Act) and registration in the EU database (Art. 49(1) AI Act — for all Annex III systems except no. 2). Deployers (Art. 3 no. 4 AI Act) document the relevance of the input data, the logs and the information given to workers (Art. 26(4), (6) and (7) AI Act) as well as, where applicable, the fundamental rights impact assessment (Art. 27 AI Act).

6. ENISA AI Threat Landscape

Short answer: ENISA is the European address for the security view of AI risks; the assessment grid for the technical requirements is set out in Art. 15 AI Act for high-risk AI systems and in Art. 55(1) AI Act where there is systemic risk.

enisa.europa.eu — the security view of AI risks.

Two sets of obligations are relevant here. For high-risk AI systems these are the requirements under Art. 15 AI Act (section 3). For providers of general-purpose AI models with systemic risk, Art. 55(1) AI Act requires a model evaluation in accordance with standardised protocols including adversarial testing (point (a); in practice adversarial testing or red-teaming), the assessment and mitigation of possible systemic risks at Union level (point (b) — these include bias, cybersecurity and the spread of disinformation), the recording and prompt reporting of serious incidents (point (c)) and an adequate level of cybersecurity for the model and its physical infrastructure (point (d)), including protection of the model weights. Where the cumulative amount of computation used for training is greater than 1025 floating point operations, high-impact capabilities are presumed to be present (Art. 51(2) AI Act).

Practitioner recommendation: no edition of the report has been evaluated (sources block); cite the edition you rely on with its year.

7. Code of Practice for GPAI

Short answer: the Code of Practice is a recognised means of demonstrating compliance, but not a statutory presumption of conformity. Providers of GPAI models may, under Art. 53(4) AI Act — where there is systemic risk, under Art. 55(2) AI Act — rely on codes of practice within the meaning of Art. 56 AI Act until a harmonised standard is published, in order to demonstrate compliance with their obligations.

digital-strategy.ec.europa.eu/en/library/general-purpose-ai-code-practice — final 10 July 2025.

The difference is decisive in an audit: the presumption of conformity arises expressly only from a harmonised European standard, to the extent that standard covers the obligations. Anyone who neither adheres to an approved code of practice nor complies with a harmonised standard must demonstrate adequate alternative means of compliance, which the Commission assesses. The Code is therefore strong evidence — it does not automatically protect against a divergent assessment by the Commission.

Practitioner recommendation: check whether the provider of a GPAI model you use has signed the Code at the time of classification, and record this with the date in the AI inventory. The substantive framework is formed by the basic obligations under Art. 53(1) AI Act: technical documentation under Annex XI AI Act, information for downstream providers under Annex XII AI Act, a copyright policy and a published summary of the training content.

8. Stiftung Neue Verantwortung — AI policy tracker

Short answer: a tracker answers what is currently going through a procedure — and that is exactly where its value and its limit lie: a proposal in the legislative process is not applicable law and carries no statement about obligations.

stiftung-nv.de — ongoing tracking of EU and DACH AI legislation.

The object lesson here is the Digital Omnibus AI: the European Commission's proposal of 19 November 2025 has become Regulation (EU) 2026/1744, in force since 27 July 2026. It postpones the Annex III dates from 2 August 2026 to 2 December 2027 and the Annex I dates from 2 August 2027 to 2 August 2028; for the conformity assessment (Art. 43 AI Act), the EU database entry (Art. 49 AI Act) and post-market monitoring (Art. 72 AI Act), read the authoritative wording in the consolidated AI Act. The procedure is complete: signature 8 July 2026, publication in the Official Journal 24 July 2026, entry into force 27 July 2026. Not postponed are the GPAI obligations and the general date of application, 2 August 2026, with Art. 50 AI Act; Art. 4 AI Act (AI literacy) has applied in a new version since 27 July 2026, Art. 5 AI Act has been supplemented by a prohibition applying from 2 December 2026, and Art. 27 AI Act, as part of Chapter III Section 3, applies to Annex III systems from 2 December 2027.

Practitioner recommendation: read the tracker entry, follow up with the primary source, cite the entry in line with the citation-practice table — and mark proposals visibly as proposals in internal documents.

Citation practice: evidencing statements traceably

Short answer: a robust reference names the legal act, article and paragraph, plus, where applicable, the point, number or annex — together with the language version and date of access. Anything that does not yield this form is context, not proof: the obligation is carried by the legal act, not by the source through which you found it. Which of the eight sources carries a statement in an audit is set out in the column "Evidentiary status / character as proof".

Two patterns hold things up in an audit interview: an obligation evidenced only through a summary, and a citation without a version. Practitioner recommendation: quarterly source review using this checklist — to be ticked off for each source:

• version or consolidation date recorded?
• date of access recorded?
• references last used listed?
• pending procedures with a bearing on deadlines checked against Art. 113 AI Act in the consolidated version?
• result updated in the AI inventory?
• ownership and review cycle entered?

Practitioner recommendation: what belongs in each entry of the AI inventory — assembled from the elements this article has already named:

• system and intended purpose
• role: provider (Art. 3 no. 3 AI Act) or deployer (Art. 3 no. 4 AI Act)
• classification with reasons and the references assessed for it (Art. 5 AI Act, Art. 6 AI Act, Annex III AI Act)
• for each statement, the reference in the form legal act, article, paragraph, and where applicable point, number or annex
• version, date of access, status of pending procedures, ownership and review cycle — as in the quarterly checklist above

Germany and Austria: what differs nationally

Short answer: nothing in the text of the Regulation itself. As an EU regulation, the AI Act applies directly and in identical wording in Germany and Austria; what is regulated nationally are the competent authorities and the penalty regime within the framework of Art. 99(1) AI Act.

Each Member State establishes or designates at least one notifying authority and at least one market surveillance authority; one market surveillance authority is designated as single point of contact, and the contact details had to be made publicly available by 2 August 2025 (Art. 70(1) and (2) AI Act). The Regulation does not name names: designation takes place nationally, so the contact details published by the Member State under Art. 70(2) AI Act are the official reference — and the proof that carries competence in an audit. In Germany, the market surveillance authority has been the Federal Network Agency (Bundesnetzagentur) since 29 July 2026 (§ 2(1) of the German AI Market Surveillance and Innovation Promotion Act (KI-MIG)), to the extent sectoral authorities are not competent; as at the date of this article (17 August 2026), Austria has not designated a market surveillance authority under Art. 70 AI Act — the RTR's AI service point is an information and advisory body without supervisory powers. Practitioner recommendation: take the authority, reference and date of access into the AI inventory separately for each country, and include them in the quarterly source review (checklist in the section "Citation practice").

For general-purpose AI models, by contrast, the question of competence is already settled in the text of the Regulation (Art. 88(1) AI Act, see section 2); for data protection supervision in Germany and Austria, section 4 applies, and for NIS2 supervision section 3 — the national bodies are named there.