What is the EU AI Act?
Regulation (EU) 2024/1689 is the first horizontal law on AI in a major market. It sorts AI by the harm a use can do, bans a short list of uses, puts heavy obligations on high-risk systems, adds transparency duties for chatbots and synthetic content, and has its own regime for general-purpose AI models. It applies to anyone placing AI on the EU market or whose AI output is used in the EU, wherever the company is based.
The Act is enforced by national market surveillance authorities, with the European Commission's AI Office supervising general-purpose AI models directly. The Omnibus widened the AI Office's remit to some AI systems built on a provider's own GPAI model and to AI inside very large online platforms and search engines.
The EU AI Act timeline, as amended
The Digital Omnibus on AI was approved by the Council on 29 June 2026, signed on 8 July, published in the Official Journal on 24 July as Regulation (EU) 2026/1744 and entered into force on 27 July 2026. Here is where every deadline now sits.
| Date | What applies | Status |
|---|---|---|
| 1 Aug 2024 | Regulation (EU) 2024/1689 enters into force | Done |
| 2 Feb 2025 | Prohibited practices banned; AI literacy duty (Article 4) applies | In force |
| 2 Aug 2025 | General-purpose AI model obligations; governance bodies and penalty regimes set up | In force |
| 2 Aug 2026 | Article 50 transparency: tell people they are talking to AI, label deepfakes, mark AI-generated content; Commission fines for GPAI providers become enforceable | In force |
| 2 Dec 2026 | Watermarking deadline for generative systems already on the market before 2 Aug 2026; new bans on AI that generates child sexual abuse material or non-consensual intimate imagery of real people | Next |
| 2 Dec 2027 | High-risk obligations for stand-alone Annex III systems (employment, credit, education, essential services, biometrics, law enforcement and more) | Moved from 2 Aug 2026 |
| 2 Aug 2028 | High-risk obligations for AI in products covered by EU safety law (Annex I: medical devices, machinery, toys, vehicles and others) | Moved from 2 Aug 2027 |
Two other Omnibus changes matter to engineering leaders. The AI literacy duty was softened from "ensure" a sufficient level of literacy to "take measures to support" it, which lowers the bar but does not remove it. And SME relief (simpler documentation, proportionate quality management, lower fines, priority sandbox access) was extended to small mid-cap companies.
The four risk tiers
| Tier | Examples | What it means |
|---|---|---|
| Unacceptable | Social scoring, manipulative techniques that cause harm, untargeted scraping of facial images, emotion recognition at work or school, most real-time remote biometric ID in public for police, and since the Omnibus, generating CSAM or non-consensual intimate images | Banned |
| High risk | Annex III uses such as hiring and worker management, credit scoring, insurance pricing, education access, essential public services, critical infrastructure; plus safety components of regulated products | Risk management, data governance, documentation, logging, human oversight, accuracy and robustness, conformity assessment, registration |
| Transparency risk | Chatbots, emotion recognition and biometric categorization, deepfakes, AI-generated text published on matters of public interest | Disclose AI interaction; label or mark synthetic content |
| Minimal risk | Spam filters, recommendation features, most internal productivity tools | No new obligations beyond AI literacy |
The classification step is where most mistakes happen, in both directions. An Annex III use is not high-risk if it performs a narrow procedural task, improves the result of a completed human activity, or only detects patterns without replacing human assessment, and the provider documents that reasoning. Profiling of natural persons is always high-risk. The Omnibus also narrowed what counts as a product "safety component": a feature that only assists users or optimizes performance is not automatically one if its failure creates no health or safety risk.
What engineering actually has to build
The legal obligations for high-risk providers read as documentation. In practice each one is a system that has to exist in the pipeline and produce evidence continuously. That is why the extra 16 months is less generous than it looks.
| Requirement | What it means in the stack | Usual owner |
|---|---|---|
| Risk management system (Art. 9) | A risk register per system that is updated from testing and production incidents, not written once | AI governance + product |
| Data governance (Art. 10) | Documented training, validation and test sets; bias examination; lineage from source to model | Data platform |
| Technical documentation (Art. 11, Annex IV) | Generated from the pipeline: architecture, data, metrics, versions. Hand-written docs drift within one release | ML engineering |
| Record-keeping (Art. 12) | Automatic event logs over the system lifetime; deployers keep them at least six months | Platform / SRE |
| Human oversight (Art. 14) | Interfaces that let a trained person understand, override or stop the output; measured, not assumed | Product + UX |
| Accuracy, robustness, cybersecurity (Art. 15) | Declared accuracy metrics, adversarial testing, resilience to data poisoning and prompt manipulation | ML + security |
| Post-market monitoring and incident reporting (Art. 72-73) | Production monitoring tied to a serious-incident process with regulatory reporting clocks | SRE + legal |
On top of that sit the process steps: a quality management system, conformity assessment before market placement (self-assessment for most Annex III systems, a notified body for biometrics and many Annex I products), CE marking and registration in the EU database. Deployers of high-risk systems carry a lighter but real load: use the system as instructed, assign trained human oversight, keep logs, inform affected workers, and for public bodies and some private deployers such as credit and insurance, run a fundamental rights impact assessment.
General-purpose AI models
Since 2 August 2025, providers of general-purpose AI models must keep technical documentation, give downstream providers enough information to comply, publish a summary of training content and have a copyright policy. Models designated as having systemic risk add evaluations, adversarial testing, incident reporting and cybersecurity duties. The Commission's fining powers over GPAI providers started on 2 August 2026. If you only call a model through an API, most of this is your vendor's problem; if you fine-tune and redistribute one, check whether the modification makes you a provider.
What to do with the extra time
- Re-plan to 2 December 2027, not to never. Move the high-risk program date and keep the funding. A roadmap that quietly loses a year of budget will be late again.
- Ship Article 50 now. Chatbot disclosure and content labeling already apply, and the watermarking grace period for systems already on the market ends on 2 December 2026.
- Finish the inventory and classification. You cannot scope high-risk work until every system has a tier with written reasoning. Our AI inventory guide covers the fields.
- Build logging and oversight into the platform once. Event logging, human override and evaluation evidence are cheaper as shared platform services than as per-team projects.
- Pick a management system. ISO/IEC 42001 or the NIST AI RMF gives the program a structure auditors recognize while harmonised standards are finalized.
For how the Act compares with ISO 42001 and the NIST AI RMF, see our AI compliance guide. For tooling that maps systems to AI Act obligations, see what an AI governance platform does.
Related questions
Frequently asked questions
When does the EU AI Act apply?
In stages. Prohibitions and the AI literacy duty have applied since 2 February 2025; general-purpose AI model obligations since 2 August 2025; Article 50 transparency duties since 2 August 2026. After the Digital Omnibus on AI (Regulation (EU) 2026/1744, in force 27 July 2026), high-risk obligations apply from 2 December 2027 for stand-alone Annex III systems and from 2 August 2028 for AI embedded in products covered by EU safety law.
Was the EU AI Act delayed?
Partly. The Digital Omnibus on AI, approved by the Council on 29 June 2026, published in the Official Journal on 24 July 2026 and in force from 27 July 2026, delayed the high-risk obligations by 16 months (Annex III) and 12 months (Annex I). It did not delay the prohibitions, the general-purpose AI model rules or most Article 50 transparency duties, and it added new prohibitions.
Does the EU AI Act apply to US companies?
Yes, if you place an AI system or general-purpose AI model on the EU market, put it into service in the EU, or the output of your system is used in the EU. That last test catches US companies whose products serve EU customers even if nothing is hosted in Europe. Non-EU providers of high-risk systems and GPAI models must appoint an authorised representative in the EU.
What are the EU AI Act fines?
Up to 35 million euros or 7% of worldwide annual turnover, whichever is higher, for prohibited practices; up to 15 million euros or 3% for most other obligations; and up to 7.5 million euros or 1% for supplying incorrect information to authorities. For SMEs and start-ups the lower of the two amounts applies, and the Omnibus extended mitigated penalties to small mid-caps.
Is ISO 42001 enough for EU AI Act compliance?
No. ISO/IEC 42001 certifies an AI management system and covers much of the governance process the Act expects, but it is not a harmonised standard under the Act and does not by itself give a presumption of conformity. The European standards bodies are writing those harmonised standards separately. ISO 42001 is still a strong foundation; see our ISO 42001 guide for how the two map.
What is the difference between a provider and a deployer?
A provider develops an AI system or model and places it on the market under its own name. A deployer uses an AI system under its authority in a professional context. Most enterprises are deployers of bought AI and providers of what they build. You become a provider of a high-risk system if you put your name on it, make a substantial modification, or change its intended purpose into a high-risk use.
Sources
- Regulation (EU) 2024/1689 (Artificial Intelligence Act), Official Journal of the European Union, 12 July 2024.
- White & Case, EU AI Omnibus enters into force, amending the AI Act (July 2026); Lewis Silkin, The Digital Omnibus on AI enters into force today (27 July 2026).
- Cloud Security Alliance, EU AI Act high-risk deadline pushed to December 2027 (2026).