The EU AI Act explained: what applies now, what moved, and what European institutions must do next

16 October 2026 · 24 min read · Last reviewed 16 October 2026

Most writing on the EU AI Act in 2026 falls into one of two errors. The first treats the Digital Omnibus as a pause, as if Brussels had stepped back from AI regulation. The second treats the Act as a compliance project for the legal department, to be closed with a policy and a training module. Both are wrong, and both are expensive.

The Act is now an operating regime. Parts of it have applied for more than eighteen months. The European AI Office has begun using its formal information powers against general-purpose model providers. National authorities in Germany and Greece have been designated by law and have staff. And the high-risk obligations that did move have moved to fixed dates written into the Regulation, not to a date anyone can negotiate further.

This essay is the reference I give to boards, ministries and supervisory bodies that ask a simple question: what does the AI Act require of us, from when, and who will check? It is written from twenty-two years at the seam where US technology platforms meet European public institutions, and it is dated deliberately. Where the law is settled, I say so. Where guidance is still in draft, I say that too.

01 What the Act is, and what it is not

The AI Act is a regulation, not a directive. It applies directly in all twenty-seven Member States without national transposition. National laws such as Germany's KI-MIG and Greece's Law 5321/2026 do not restate the obligations; they designate authorities, set procedures and fill the spaces the Regulation leaves to Member States, such as penalties for public bodies and the operation of regulatory sandboxes.

It is product-safety law at its core. The Act borrows the architecture of the New Legislative Framework: providers place systems on the market, conformity is assessed against essential requirements, harmonised standards give a presumption of conformity, and market surveillance authorities police the result. That lineage explains its vocabulary (provider, deployer, placing on the market, putting into service, substantial modification) and its method. It asks for evidence, produced before the fact, kept for years.

It regulates two different things. AI systems are regulated by what they are used for: the same model can be minimal risk in one deployment and high-risk in another. General-purpose AI models are regulated by what they can do: their providers carry obligations regardless of downstream use, and the most capable models carry more.

It has global reach. Article 2 applies the Act to providers placing AI systems or general-purpose models on the EU market wherever they are established, and to providers and deployers outside the Union where the output of the system is used in the Union. A US model provider, an Indian software house and a Swiss deployer can all be in scope.

What the Act is not: it is not a data protection law (the GDPR continues to apply in full), it is not a liability regime, and it is not a sovereignty instrument. It says a great deal about how an AI system must behave and almost nothing about who controls the infrastructure underneath it. I return to that gap in section 13, because for public institutions it is the larger risk.

02 The timeline after the Digital Omnibus

The original Regulation staggered its application over three years. Regulation (EU) 2026/1744, the Digital Omnibus on AI, adopted on 8 July 2026, published in the Official Journal on 24 July 2026 and in force since 27 July 2026, moved some of those dates. This is the consolidated picture as of October 2026.

Two observations matter more than the table itself.

First, the new high-risk dates are fixed. The Commission's November 2025 proposal would have linked application to a Commission decision confirming that standards and support tools were ready, with backstop dates. The co-legislators removed the trigger and kept the backstops. The dates are now written into Article 113, and only another amending regulation could move them.

Second, the Omnibus moved the high-risk regime as a block. Chapter III Sections 1 to 3 cover classification, the requirements for high-risk systems, and the obligations of providers and deployers alike. Deployer duties under Article 26 and the fundamental rights impact assessment under Article 27 therefore also apply to Annex III systems from 2 December 2027.

03 What the Omnibus changed, and what it left alone

The Omnibus was negotiated quickly. The Commission proposed it on 19 November 2025; Parliament and Council reached a provisional agreement on 7 May 2026; Parliament approved it in plenary on 16 June and the Council on 29 June. Its recitals are unusually candid about the reason: harmonised standards and national governance structures arrived late, and the compliance burden proved heavier than expected. It is a reprieve granted because the infrastructure of compliance was not ready, not a retreat from the requirements.

What changed, in substance:

• High-risk dates. Annex III systems from 2 December 2027; Annex I product systems from 2 August 2028.
• Two new prohibitions. From 2 December 2026, Article 5 bans AI systems that generate or manipulate realistic intimate imagery of an identifiable person without explicit consent, and systems that generate child sexual abuse material. The ban on placing such systems on the market is targeted at systems whose intended purpose is that output, or whose design makes it a reasonably foreseeable and reproducible outcome without adequate safeguards.
• AI literacy rewritten. Article 4 previously required providers and deployers to ensure, to their best extent, a sufficient level of AI literacy. It now requires them to take measures to support the development of AI literacy of their staff, and states that no specific level has to be guaranteed. The date, 2 February 2025, did not move.
• Bias detection data. A new Article 4a allows providers of high-risk systems to process special categories of personal data for bias detection and correction only where strictly necessary and where synthetic or anonymised data would not suffice, under listed safeguards.
• Safety components clarified. New Article 6(1a) states that AI used solely for user assistance, performance optimisation, service efficiency, automation, convenience or non-safety quality control is not a safety component. Some product-embedded AI leaves the high-risk route as a result.
• SMEs and small mid-caps. The Act now defines small mid-cap enterprises (SMCs) and extends simplified technical documentation (Article 11) and proportionate quality management (Article 17) to them. For SMCs, fines under Article 99(4) and (5) are capped at whichever of the amount or percentage is lower.
• AI Office supervision. Article 75 gives the AI Office exclusive competence over AI systems built on a general-purpose model where the model and the system come from the same provider or from providers in the same undertaking (with exceptions for Annex I products, critical infrastructure, law enforcement, border management, financial institutions and the administration of justice), and over AI systems that are or are integrated into very large online platforms and search engines. New Articles 75a to 75d give it inspection powers, binding commitments, fines and periodic penalty payments of up to 5 percent of average daily turnover.
• Value chain cooperation. The written-agreement duty in Article 25(4) now expressly covers suppliers of AI models, and an initial provider whose system becomes high-risk in someone else's hands must hand over documentation, known limitations and targeted technical access, unless it has clearly excluded that use.
• Fundamental rights and data protection. A deployer's fundamental rights impact assessment may now cross-refer to an existing data protection impact assessment rather than duplicate it, and the AI Office must provide a questionnaire template, including an automated tool.
• Cybersecurity. High-risk systems that meet the Cyber Resilience Act's essential requirements are deemed to meet the cybersecurity requirement of Article 15.
• Sandboxes. The deadline for each Member State's first AI regulatory sandbox moved to 2 August 2027, and the AI Office may run a Union-level sandbox.
• Machinery. The Machinery Regulation moved from Section A to Section B of Annex I, so AI in machinery will be governed through sector rules that the Commission must align with the AI Act's requirements by delegated act.

What did not change: the risk pyramid, the original eight prohibitions, the general-purpose model obligations, the Article 50 transparency duties (apart from the marking transition for legacy generative systems), the registration of systems a provider claims are not high-risk, the penalty ceilings, and the evidence logic of the high-risk regime. Technical documentation must still exist before a high-risk system is placed on the market; it describes decisions taken in 2026 and 2027. December 2027 is when the file is inspected. 2026 is when it is written.

04 Your role decides your obligations

Before classification, settle the role. The Act distributes obligations by role, and the same institution can hold several roles across its portfolio.

• Provider. Develops an AI system or general-purpose model, or has one developed, and places it on the market or puts it into service under its own name, whether for payment or free of charge. Putting into service includes supplying a system for your own use. An institution that builds its own citizen chatbot is normally its provider.
• Deployer. Uses an AI system under its own authority in a professional capacity. Most ministries, agencies and banks are deployers of most of their AI.
• Importer and distributor. Bring systems from outside the Union into the market or make them available along the supply chain, with verification duties.
• Authorised representative. Mandatory for non-EU providers of high-risk systems and of general-purpose models (providers of free and open-source models without systemic risk are exempt).

The trap is Article 25. A deployer becomes a provider of a high-risk system if it puts its name or trademark on it, makes a substantial modification, or changes the intended purpose of a system so that it becomes high-risk. Configuring a vendor's agent platform into a recruitment screener, or relabelling a supplier's model as the ministry's own assistant, can move an institution across that line without anyone deciding to. Record the role for each system, with the reasoning, in the inventory.

05 Four tiers of risk, plus general-purpose models

Prohibited. Article 5 bans eight practices since 2 February 2025: manipulative or deceptive techniques that materially distort behaviour and cause significant harm; exploitation of vulnerabilities due to age, disability or social or economic situation; social scoring leading to unjustified detrimental treatment; predicting the risk of a person committing a crime based solely on profiling or personality traits; untargeted scraping of facial images to build recognition databases; emotion recognition in the workplace and in education, except for medical or safety reasons; biometric categorisation to infer sensitive characteristics such as political opinions, religious beliefs or sexual orientation; and real-time remote biometric identification in publicly accessible spaces for law enforcement, subject to narrow exceptions. Two further prohibitions apply from 2 December 2026, as described above.

High-risk. Two routes. Under Article 6(1), an AI system is high-risk if it is a safety component of a product, or is itself a product, covered by the Union harmonisation legislation in Annex I and required to undergo third-party conformity assessment: medical devices, toys, lifts, radio equipment and similar. Under Article 6(2), a system is high-risk if it falls within a use case listed in Annex III: biometrics; critical infrastructure; education and vocational training; employment and worker management; access to essential private and public services and benefits, including credit scoring and life and health insurance pricing; law enforcement; migration, asylum and border control; and the administration of justice and democratic processes.

Article 6(3) offers a derogation: an Annex III system is not high-risk if it does not pose a significant risk of harm, for example because it performs a narrow procedural task or only improves the result of a completed human activity. The derogation never applies where the system profiles natural persons. A provider relying on it must document the assessment before placing the system on the market and register the system in the EU database. The Commission's draft guidelines on classification, published on 19 May 2026 and consulted on until 23 July 2026, make clear that appearing on the Annex III list is where the assessment starts, not where it ends: being listed is not sufficient on its own, and the intended purpose decides. Final guidelines are expected by the end of 2026.

Transparency. Article 50 imposes disclosure and marking duties on chatbots, generative systems, emotion recognition, biometric categorisation, deepfakes and certain published text. These apply regardless of risk tier. Section 06 covers them in detail.

Minimal risk. Everything else: spam filters, scheduling tools, most internal productivity uses. No specific obligations beyond AI literacy and the prohibitions, but keep these systems in the inventory. A minimal-risk system that acquires a new use case can quietly become something else.

General-purpose AI models sit beside this pyramid rather than inside it. Their obligations attach to the model provider, not to the use, and are covered in section 07.

06 Article 50: the transparency duties already in force

Article 50 has applied since 2 August 2026. It splits duties between providers and deployers:

• Article 50(1), provider. Systems that interact directly with people must be designed so that people know they are dealing with AI, unless that is obvious from the context.
• Article 50(2), provider. Generative systems producing synthetic audio, image, video or text must mark their output in a machine-readable format so it can be detected as artificially generated or manipulated. Systems placed on the market before 2 August 2026 have until 2 December 2026.
• Article 50(3), deployer. Deployers of emotion recognition or biometric categorisation systems must inform the people exposed to them.
• Article 50(4), deployer. Deployers must disclose deepfakes, and must disclose AI-generated or manipulated text published to inform the public on matters of public interest, unless the text has undergone human review or editorial control and someone holds editorial responsibility.

The Commission published the final Code of Practice on marking and labelling of AI-generated content on 10 June 2026. It is voluntary. Its provider section sets out multi-layered machine-readable marking and detection; its deployer section sets out labelling of deepfakes and public-interest text, with a common set of EU icons. The Code is subject to an adequacy assessment by the Commission, taking utmost account of the AI Board's opinion, and the Commission may adopt implementing rules if it finds the Code inadequate. Separate Commission guidelines on the scope of Article 50 complement it.

For public institutions the practical point is simple. Any citizen-facing assistant, any generated communication and any synthetic image used in a public campaign is already in scope. Most institutions are both provider and deployer under Article 50 at the same time: provider of the assistant they built, deployer of the generative tools their communications team uses.

07 General-purpose AI models and the AI Office

Since 2 August 2025, every provider of a general-purpose AI model placed on the EU market must keep technical documentation, provide information to downstream providers who integrate the model, maintain a policy to comply with Union copyright law, and publish a sufficiently detailed summary of the content used for training (Article 53). Models with systemic risk, presumed where cumulative training compute exceeds 10^25 floating-point operations, carry additional duties under Article 55: model evaluation including adversarial testing, assessment and mitigation of systemic risks, serious incident reporting and adequate cybersecurity. Models placed on the market before 2 August 2025 have until 2 August 2027.

The General-Purpose AI Code of Practice, published in July 2025, is the de facto benchmark the AI Office uses. Adherence is voluntary; a provider that does not sign must demonstrate compliance by other means.

The Commission's enforcement powers over these providers, including fines under Article 101 of up to 15 million euro or 3 percent of worldwide annual turnover, became applicable on 2 August 2026. On 29 August 2026 the AI Office used them for the first time: it sent formal requests for information under Article 91 to more than thirty providers of general-purpose AI models, reportedly among the developers of the most advanced models. The requests covered two strands: safety and security (external evaluations, post-market monitoring) and transparency and copyright (training-data summaries, copyright policies). No fine had been announced at the time of writing.

For deployers this matters indirectly but materially. If a general-purpose model sits underneath your high-risk system, you inherit none of its provider's obligations, but you depend on all of them. Require the upstream documentation as a contractual deliverable, and require notice of model changes with enough time to re-evaluate, because a silent upstream update can invalidate your own accuracy and robustness declarations.

08 High-risk systems: the work that cannot wait for 2027

The requirements for high-risk systems are in Articles 9 to 15; provider obligations in Articles 16 to 22; deployer obligations in Article 26. I have set out the sequence elsewhere in an EU AI Act compliance roadmap for high-risk deployments, and the critical infrastructure case in the EU AI Act and the operators of critical infrastructure. Here is why the extra sixteen months are shorter than they look.

Documentation is retrospective by nature. Article 11 requires technical documentation to be drawn up before the system is placed on the market or put into service. Annex IV asks for the design choices, data provenance, training methodology, evaluation results and risk management outputs. Those are decisions being made now. If they are not recorded as they are made, they will be reconstructed later, and reconstruction is visible to an inspector.

Retention starts at launch and runs for a decade. Providers keep documentation available for ten years after placing on the market (Article 18) and keep automatically generated logs for at least six months (Article 19). Logging has to be designed into the architecture, not added for the audit.

Legacy systems are not safe harbours. High-risk systems already on the market before the application date fall under the Act only if they undergo a significant change in design. That sounds like relief until you need to prove the negative. A versioned change record, kept from the start, is the only way to show a system did not change significantly.

Standards are still arriving. Harmonised standards give a presumption of conformity. The European standardisation organisations are behind schedule, which is one of the reasons the Omnibus exists. Until standards are cited in the Official Journal, providers argue conformity requirement by requirement. Plan for that.

Deployers have their own list. Under Article 26 a deployer must use the system in accordance with the instructions for use, assign human oversight to people with the necessary competence, training and authority, ensure input data is relevant and representative where it controls that data, monitor operation and report risks and serious incidents, keep logs under its control for at least six months, and inform workers' representatives before deploying high-risk AI in the workplace. Under Article 86, a person subject to a decision based on the output of an Annex III system (other than critical infrastructure), with legal or similarly significant effects, has the right to a clear and meaningful explanation of the role of the system in the decision.

09 Public authorities: duties specific to the state

Public bodies carry obligations that private deployers do not.

• Fundamental rights impact assessment. Article 27 requires deployers that are bodies governed by public law, or private entities providing public services, to assess the impact on fundamental rights before deploying an Annex III high-risk system (critical infrastructure systems under point 2 are excluded). Deployers using AI for creditworthiness assessment or for life and health insurance pricing carry the same duty. Since the Omnibus, the assessment may cross-refer to an existing DPIA.
• Registration of use. Public authority deployers must register their use of Annex III high-risk systems in the EU database, and must not use a system that is not registered. Critical infrastructure systems are registered at national level, and law enforcement, migration and border systems in a non-public section.
• The 2030 horizon. High-risk systems already on the market and intended for use by public authorities must be brought into compliance by 2 August 2030 (Article 111(2)). That is not a reason to wait: any significant design change before then brings the system into scope earlier, and procurement decisions made in 2026 will still be running in 2030.
• National registers. Member States may add their own requirements for public sector transparency. Greece has done so: Law 5321/2026 creates a central register of AI systems used by public sector bodies, and systems already in operation must be registered by 31 December 2026.

There is also a question of standing. A public institution is not only a deployer; it is the body citizens expect to set the example. An agency that cannot produce its own AI inventory will find it harder to demand one from the operators it supervises.

10 Who enforces it: the EU, Germany and Greece

At EU level. The European AI Office, part of the Commission's DG CONNECT, supervises general-purpose AI models and, since the Omnibus, certain AI systems built on them and those embedded in very large online platforms and search engines. The European Artificial Intelligence Board coordinates Member States; a scientific panel of independent experts supports the AI Office on systemic risk; an advisory forum represents stakeholders. The European Data Protection Supervisor supervises the Union institutions, bodies, offices and agencies themselves.

In Germany. The KI-Marktüberwachungs- und Innovationsförderungsgesetz (KI-MIG), Article 1 of the Gesetz zur Durchführung der Verordnung über künstliche Intelligenz of 22 July 2026, was adopted by the Bundestag on 11 June 2026, approved by the Bundesrat on 10 July 2026, published in the Bundesgesetzblatt on 28 July 2026 and has applied since 29 July 2026. It makes the Bundesnetzagentur the central market surveillance authority, covering prohibited practices, Annex III high-risk systems and transparency duties. An independent KI-Marktüberwachungskammer within the Bundesnetzagentur oversees certain sensitive high-risk uses, and a central coordination and competence centre supports consistent application across authorities. The Bundesnetzagentur also serves as the central complaints point and must establish at least one AI regulatory sandbox (KI-Reallabor). The AI Act already gives SMEs and start-ups priority access to sandboxes; the KI-MIG adds priority for research institutions, universities and their spin-offs. AI systems of public bodies of the Länder and municipalities are excluded from its market surveillance. For financial institutions, the Act itself assigns market surveillance to the financial supervisor, in Germany BaFin.

In Greece. Law 5321/2026 was published in the Government Gazette on 20 July 2026 (ΦΕΚ Α' 114/20.07.2026). The Hellenic Data Protection Authority is the market surveillance authority for prohibited practices, Annex III high-risk systems and Article 50 transparency, the single point of contact, and the operator of a single complaints system. The Hellenic Telecommunications and Post Commission (EETT) is the notifying authority and hosts the AI Coordination and Expertise Centre and the national AI regulatory sandbox. Sectoral authorities keep their product remits: the National Organisation for Medicines for medical devices, for example. The law creates the public sector AI register described above, makes the removal of transparency markings from synthetic content an offence, brings AI Act infringements within whistleblower protection, and repeals the AI provisions of Law 4961/2022.

11 Penalties

Article 99 sets three ceilings, each the higher of a fixed amount or a percentage of worldwide annual turnover:

• Prohibited practices: up to 35 million euro or 7 percent.
• Most other obligations, including high-risk requirements, operator obligations and Article 50 transparency: up to 15 million euro or 3 percent.
• Supplying incorrect, incomplete or misleading information to authorities: up to 7.5 million euro or 1 percent.

For SMEs, and since the Omnibus for small mid-caps in the second and third bands, the lower of the two figures applies. Providers of general-purpose AI models face fines from the Commission of up to 15 million euro or 3 percent under Article 101. Member States decide whether and to what extent fines can be imposed on their own public authorities; Union institutions face fines from the European Data Protection Supervisor under Article 100.

Fines are not the only lever. Market surveillance authorities can order corrective action, restrict or withdraw a system, and recall it. For a public service that depends on a system, withdrawal is the sanction that matters.

12 The Act inside the wider rulebook

No institution meets the AI Act alone. The same system is often caught by four or five regimes at once.

• GDPR. Applies in full to any personal data an AI system processes. The DPIA and the fundamental rights impact assessment should be one workstream.
• DORA and NIS2. Financial entities and essential operators already run ICT risk management and incident reporting. An institution whose AI Act, DORA and NIS2 incident clocks run on different definitions will spend its first real incident reconciling its own timestamps. One incident taxonomy, one clock, several reports.
• Cyber Resilience Act. High-risk systems meeting the CRA's essential cybersecurity requirements are now deemed to meet Article 15's cybersecurity requirement.
• Data Act. Changes the data foundation beneath AI: access rights for users of connected products and services, and cloud switching obligations that matter for any exit plan.
• Sector law. Medical devices, machinery and financial supervision each bring their own conformity and supervisory channels.

The institutions that move fastest treat the AI Act as the integration spine for the others rather than as a fifth parallel programme. One inventory, one owner per system, one evidence base.

13 The question the Act does not answer: dependency

The AI Act governs how a model behaves. It does not govern who can switch it off.

On 12 June 2026, a US export control directive targeting access by foreign nationals led one US provider to suspend two of its frontier models for all customers worldwide. The controls were lifted on 30 June, and access began returning, in phases, from 1 July. Whatever one thinks of the policy, the operational lesson for European institutions is plain: a service can be fully compliant with the AI Act and still become unavailable overnight for reasons entirely outside European jurisdiction.

The EU is beginning to legislate on this layer. On 3 June 2026 the Commission proposed the Cloud and AI Development Act as part of its Technological Sovereignty Package, introducing a single framework of sovereignty assurance levels for cloud services and sovereignty risk assessments for public authorities. It is a proposal, still before Parliament and Council. Institutions should not wait for it.

Three questions belong in every AI procurement and every AI Act inventory entry, alongside the risk tier:

• Residency. Where are the data, the prompts and the logs?
• Jurisdiction. Which laws can reach them, and which authority can compel access or suspend service?
• Control. Who operates the service, who controls access, and can we leave within an agreed time and budget?

Location alone does not establish exclusive EU jurisdiction. An exit is only real once it has been tested. And for essential functions, a fallback matters: routing sensitive workloads to models running in the institution's own environment, and keeping a local capability for critical tasks when a public service is unavailable. I use public frontier models every day. The point is not to avoid them. It is to end dependence on a single provider for anything the institution must keep doing. The distinction is set out in more detail in data residency vs data sovereignty and digital sovereignty as resilience.

14 A 90-day plan

The governance model I use with institutions has six layers grouped in three capabilities: adopt (AI inventory, responsible deployment), defend (security and access, testing and monitoring) and govern (human oversight, compliance and audit). Governance cannot be added after deployment, so the first ninety days are about visibility.

Days 1 to 30: see everything.

• Inventory every AI system in operation, procurement and pilot, including AI features inside existing software.
• Record for each: purpose, decision influenced, affected persons, data categories, operating dependency, role (provider or deployer) and the reasoning for that role.
• Identify anything that touches the Article 5 prohibitions, including the two that apply from 2 December 2026, and stop it.

Days 31 to 60: classify and close the live gaps.

• Classify each system: prohibited, high-risk (Annex I or Annex III), transparency only, minimal. Document negative classifications and any Article 6(3) derogation.
• Close Article 50 gaps now: chatbot disclosure, output marking for generative systems you provide (2 December 2026 for legacy systems), labelling for deepfakes and published text.
• Record AI literacy measures per role. The duty has applied since February 2025.
• Add residency, jurisdiction and control to each inventory entry.

Days 61 to 90: build the evidence machine.

• For every high-risk candidate, start the Annex IV file and the change log now, versioned alongside the system.
• Design logging for traceability and retention.
• Put Article 25 written agreements in place with model and component suppliers.
• Align AI, DORA, NIS2 and GDPR incident definitions on one clock.
• For public bodies: schedule the fundamental rights impact assessment with the DPIA, and in Greece register existing systems by 31 December 2026.

One test for the steering committee: for each AI system that influences a decision about a citizen, ask for a single page covering purpose, evidence, boundaries, accountability and measurement. If that page cannot be written, the system is not ready to scale.

15 Frequently asked

When does the EU AI Act apply? In stages. Prohibitions and AI literacy since 2 February 2025; general-purpose AI model obligations since 2 August 2025; general application, including Article 50 transparency, since 2 August 2026; high-risk obligations for Annex III systems from 2 December 2027 and for Annex I product systems from 2 August 2028.

Was the EU AI Act delayed? Only partly. The Digital Omnibus on AI, Regulation (EU) 2026/1744, postponed the high-risk obligations by sixteen months for Annex III systems and twelve months for Annex I systems. Everything else kept its original date, and two new prohibitions were added.

Is the Digital Omnibus law, or still a proposal? It is law. It was adopted on 8 July 2026, published in the Official Journal on 24 July 2026 and entered into force on 27 July 2026.

Does the AI Act apply to companies outside the EU? Yes. It applies to providers placing AI systems or general-purpose models on the EU market wherever they are established, and to providers and deployers outside the EU whose system output is used in the EU. Non-EU providers of high-risk systems and of general-purpose models must appoint an authorised representative in the Union, with an exemption for free and open-source models without systemic risk.

What counts as a high-risk AI system? Either an AI system that is a safety component of, or itself, a product under the legislation in Annex I and subject to third-party conformity assessment, or a system used for a purpose listed in Annex III, such as recruitment, credit scoring, access to public benefits, critical infrastructure, education, law enforcement, migration or justice. An Annex III system can escape high-risk status under Article 6(3) only if it poses no significant risk and does not profile people, and the provider must document and register that assessment.

Do chatbots have to say they are AI? Yes, since 2 August 2026, under Article 50(1), unless it is obvious from the context. The duty sits with the provider, which is often the institution that built and runs the chatbot.

What are the fines under the EU AI Act? Up to 35 million euro or 7 percent of worldwide annual turnover for prohibited practices; up to 15 million euro or 3 percent for most other obligations, including transparency; up to 7.5 million euro or 1 percent for misleading information to authorities. General-purpose AI model providers face Commission fines of up to 15 million euro or 3 percent.

Who enforces the AI Act in Germany? Mainly the Bundesnetzagentur, under the KI-MIG in force since 29 July 2026, with an independent AI market surveillance chamber for certain high-risk uses. BaFin supervises AI used by financial institutions, and sector authorities keep their product remits.

Who enforces the AI Act in Greece? Mainly the Hellenic Data Protection Authority, under Law 5321/2026, with EETT as notifying authority and host of the AI Coordination and Expertise Centre and the regulatory sandbox.

Is AI literacy still mandatory? Yes. Article 4 has applied since 2 February 2025. Since the Omnibus it requires providers and deployers to take measures to support the AI literacy of their staff and others operating AI on their behalf, without guaranteeing a specific level. Keep a record of the measures taken.

What should a public institution do first? Build the inventory. Every other obligation, from classification to registration to the fundamental rights impact assessment, depends on knowing which AI systems you have, what they decide and who depends on them.

Sources

Primary legal texts and official publications:

• Regulation (EU) 2024/1689 (Artificial Intelligence Act), consolidated text as of 27 July 2026, EUR-Lex: https://eur-lex.europa.eu/legal-content/EN/TXT/HTML/?uri=CELEX%3A02024R1689-20260727
• Regulation (EU) 2026/1744 (Digital Omnibus on AI), Official Journal, 24 July 2026: https://eur-lex.europa.eu/eli/reg/2026/1744/oj/eng
• European Commission, Regulatory framework on AI: https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai
• European Commission, Draft guidelines on the classification of high-risk AI systems, 19 May 2026: https://digital-strategy.ec.europa.eu/en/library/draft-commission-guidelines-classification-high-risk-ai-systems
• AI Act Service Desk, Article 50: https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-50
• Deutscher Bundestag, KI-MIG (Drucksache 21/4594), first reading of 20 March 2026: https://www.bundestag.de/dokumente/textarchiv/2026/kw12-de-kuenstliche-intelligenz-1151800
• IHK, KI-MIG veröffentlicht (BGBl. 2026 I Nr. 223, in force 29 July 2026): https://www.ihk.de/ihklw/produkte/beratung-und-service/ecommerce-und-internet/ki-mig-veroeffentlicht-7142702
• Hellenic Republic, Law 5321/2026, ΦΕΚ Α' 114/20.07.2026: https://search.et.gr/el/fek/?fekId=802524

Further reading on this site:

• An EU AI Act compliance roadmap for high-risk deployments: https://uberconsul.com/insights/eu-ai-act-compliance-roadmap
• The EU AI Act and the operators of critical infrastructure: https://uberconsul.com/insights/eu-ai-act-critical-infrastructure
• Building an AI governance framework a European institution can defend: https://uberconsul.com/insights/ai-governance-framework-european-institutions
• Data residency vs data sovereignty: https://uberconsul.com/insights/data-residency-vs-data-sovereignty-europe
• The UberConsul Resilience Audit: https://uberconsul.com/methodology

About the author

Vasilios Kotitsas is the founding principal of UberConsul, a European institutional AI advisory in Frankfurt am Main. He has spent twenty-two years at the seam where US technology platforms meet European public institutions, including seven years in a hyperscaler's Worldwide Public Sector organisation, where 64 European institutions were onboarded under a single framework agreement, and a matrix programme coordinating 93 EU and EEA national competent authorities, including four central banks, for a cross-border supervisory data platform. He works in English, German and Greek.

This essay is general information on the law as it stood on 16 October 2026. It is not legal advice.

How to cite: Kotitsas, V. (2026). The EU AI Act explained: what applies now, what moved, and what European institutions must do next. UberConsul. https://uberconsul.com/insights/eu-ai-act-explained

First published 16 October 2026 · Frankfurt am Main.

← All insights