EU AI Act: Risk Categories, Obligations and Deadlines Explained

The EU AI Act is the first broad law on artificial intelligence in a major market, and it reaches software companies far beyond Europe. If you build an AI feature that is used in the EU, or if a customer in the EU uses your AI output, the Act can apply to you. Its obligations depend on what the AI system does, not on the technology behind it: a simple scoring model used to screen job applicants can be high-risk, while a large language model used to draft marketing copy usually is not.

The Act has been amended once already. The Digital Omnibus on AI, Regulation (EU) 2026/1744, entered into force on 27 July 2026 and moved the dates for high-risk obligations to December 2027 and August 2028, added a new prohibition, and adjusted other provisions. This guide reflects the Act as amended at our last review. It covers who is in scope, the risk categories, the duties of providers and deployers, general-purpose AI models, the timeline, penalties, and the engineering work that makes compliance practical. Because the rules and guidance are still developing, this guide is reviewed every three months.

What is the EU AI Act?

The Artificial Intelligence Act is Regulation (EU) 2024/1689. It applies directly in all EU member states. It defines an AI system (Article 3(1)) as a machine-based system designed to operate with varying levels of autonomy, that may exhibit adaptiveness after deployment, and that infers from the input it receives how to generate outputs such as predictions, content, recommendations or decisions that can influence physical or virtual environments. The Commission has published guidelines on this definition; conventional software that only follows rules defined by people is generally outside it.

The Act takes a risk-based approach with four levels, plus separate rules for general-purpose AI models:

• Unacceptable risk: prohibited practices (Article 5).
• High risk: permitted, subject to strict requirements (Articles 6 to 49).
• Transparency risk: chatbots, AI-generated or manipulated content and some other systems must disclose that AI is involved (Article 50).
• Minimal risk: no specific obligations, beyond AI literacy and voluntary codes.
• General-purpose AI models: obligations for model providers, with more for models with systemic risk (Articles 51 to 55).

Roles

Obligations attach to roles. A provider 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. A deployer uses an AI system under its authority in a professional activity. Importers, distributors and authorised representatives have their own duties. A software company that builds an AI feature into its SaaS product is usually the provider of that AI system; its business customers are deployers.

History and why the AI Act was introduced

The European Commission proposed the AI Act in April 2021, after a white paper and public consultation. The proposal predated the wave of generative AI, so the Parliament and Council added rules for general-purpose AI models during negotiations. Political agreement was reached in December 2023, the regulation was adopted in 2024, published in the Official Journal on 12 July 2024 and entered into force on 1 August 2024.

Its stated aims are to improve the functioning of the internal market, promote human-centric and trustworthy AI, and ensure a high level of protection of health, safety and fundamental rights against the harmful effects of AI systems, while supporting innovation. It follows the EU's product safety model: requirements before a product reaches the market, conformity assessment, CE marking and market surveillance.

In November 2025 the Commission proposed a Digital Omnibus on AI to simplify parts of the Act and to delay the high-risk rules, because harmonised standards and guidance would not be ready in time. A provisional agreement was reached in May 2026, the Parliament and Council approved it in June 2026, and Regulation (EU) 2026/1744 entered into force on 27 July 2026.

Who must comply? Scope and jurisdiction

Article 2 applies the Act to:

• Providers placing AI systems on the EU market or putting them into service, or placing general-purpose AI models on the market, wherever they are established.
• Deployers established or located in the EU.
• Providers and deployers in third countries where the output produced by the AI system is used in the EU.
• Importers, distributors and product manufacturers placing AI systems on the market with their products.

So an Indian or US software company selling an AI-powered recruitment screening tool to EU employers is a provider in scope, and so is a US company whose AI output (for example, a credit score) is used by an EU bank. Exclusions include AI used exclusively for military, defence or national security, AI used for scientific research and development only, research and testing before market placement (with limits), and purely personal non-professional use. AI systems released under free and open-source licences are excluded unless they are high-risk, prohibited or subject to Article 50.

Providers outside the EU of high-risk systems and of general-purpose AI models must appoint an authorised representative in the EU.

Prohibited AI practices (Article 5)

Since 2 February 2025, these practices have been banned in the EU, subject to narrow exceptions in the text:

• Subliminal, manipulative or deceptive techniques that materially distort behaviour and cause or are likely to cause significant harm.
• Exploiting vulnerabilities due to age, disability or social or economic situation in a way that causes significant harm.
• Social scoring that leads to detrimental or unfavourable treatment in unrelated contexts or out of proportion.
• Assessing the risk that a person will commit a crime based solely on profiling or personality traits.
• Untargeted scraping of facial images from the internet or CCTV to build facial recognition databases.
• Emotion recognition in workplaces and education institutions, except for medical or safety reasons.
• Biometric categorisation to infer race, political opinions, trade union membership, religious beliefs, sex life or sexual orientation.
• Real-time remote biometric identification in publicly accessible spaces for law enforcement, except in listed situations with authorisation.

The Digital Omnibus added a prohibition on AI systems designed to generate non-consensual sexual or intimate imagery and child sexual abuse material. The Commission has published guidelines on the prohibited practices.

High-risk AI systems

An AI system is high-risk in two ways (Article 6):

1. It is a safety component of, or is itself, a product covered by EU harmonisation legislation in Annex I (such as machinery, medical devices, toys, lifts, radio equipment) that requires third-party conformity assessment.
2. It falls in an area listed in Annex III: biometrics (remote identification, categorisation, emotion recognition where permitted); critical infrastructure; education and vocational training (admission, assessment, proctoring); employment and workers management (recruitment, screening, promotion, termination, task allocation, monitoring); access to essential private and public services (eligibility for benefits, creditworthiness and credit scoring, risk pricing in life and health insurance, emergency call triage); law enforcement; migration, asylum and border control; and administration of justice and democratic processes.

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 human activity (Article 6(3)), but a system that profiles people is always high-risk. A provider relying on this exception must document its assessment and, as reinstated in simplified form by the Digital Omnibus, register the system.

Requirements for high-risk systems (Articles 8 to 15)

Article	Requirement	What it means in practice
9	Risk management system	A continuous process to identify, evaluate and mitigate risks across the life cycle, with testing
10	Data and data governance	Training, validation and test data that are relevant, representative, as free of errors as possible, and examined for bias
11	Technical documentation	The Annex IV file: design, development, data, performance, risk management, changes
12	Record-keeping	Automatic logging of events over the system's lifetime to support traceability and monitoring
13	Transparency and information for deployers	Instructions for use: capabilities, limitations, accuracy, human oversight measures
14	Human oversight	Design that lets people understand, monitor, override and stop the system
15	Accuracy, robustness and cybersecurity	Declared accuracy metrics, resilience to errors and to attacks such as data poisoning and adversarial inputs

Provider obligations

Providers of high-risk systems must also run a quality management system (Article 17), keep documentation and logs, carry out the relevant conformity assessment (Article 43), draw up an EU declaration of conformity, affix CE marking, register in the EU database (Article 49), take corrective action and inform authorities when needed, and run post-market monitoring and serious incident reporting (Articles 72 and 73).

Deployer obligations

Deployers must use high-risk systems according to the instructions, assign human oversight to competent people, ensure input data under their control is relevant, monitor operation, keep logs under their control for at least six months, inform workers before using high-risk AI in the workplace, and inform people when high-risk AI is used to make decisions about them (Article 26). Public bodies and certain private deployers, such as those doing credit scoring or life and health insurance pricing, must carry out a fundamental rights impact assessment (Article 27). People affected by certain decisions have a right to an explanation (Article 86).

Transparency obligations (Article 50)

• Providers must design AI systems that interact with people so that people are told they are interacting with AI, unless obvious from context.
• Providers of systems generating synthetic audio, image, video or text must mark outputs in a machine-readable format and make them detectable as artificially generated or manipulated.
• Deployers of emotion recognition or biometric categorisation systems must inform the people exposed.
• Deployers of deepfakes must disclose that the content is artificially generated or manipulated, and deployers publishing AI-generated text on matters of public interest must disclose it unless it has undergone human editorial review.

These duties apply from 2 August 2026. Under the Digital Omnibus, the marking obligation for generative systems already on the market before that date applies from 2 December 2026.

General-purpose AI models

Providers of general-purpose AI (GPAI) models, such as large language models offered to others, must (Article 53) keep technical documentation, give information and documentation to downstream providers who build AI systems on the model, put in place a policy to comply with EU copyright law including text and data mining opt-outs, and publish a sufficiently detailed summary of the content used for training using the Commission's template. Models with systemic risk (presumed when training compute exceeds 10^25 floating-point operations, or designated by the Commission) must also evaluate the model, assess and mitigate systemic risks, report serious incidents and ensure cybersecurity (Article 55). A voluntary General-Purpose AI Code of Practice, published in July 2025, is a way to show compliance.

Most software companies are not GPAI model providers. If you call a model through an API to build a feature, you are usually the provider of an AI system that uses a GPAI model, and you rely on the information the model provider must give you.

Timeline: when each part applies

Date	What applies
1 August 2024	Entry into force
2 February 2025	Prohibited practices (Article 5) and AI literacy (Article 4)
2 August 2025	General-purpose AI model obligations, governance, notified bodies, penalties framework
2 August 2026	Transparency obligations (Article 50) and most remaining provisions not deferred
2 December 2026	Marking of AI-generated content for generative systems already on the market before 2 August 2026
2 August 2027	Deadline for GPAI models placed on the market before 2 August 2025
2 December 2027	High-risk obligations for stand-alone Annex III systems (as amended by the Digital Omnibus)
2 August 2028	High-risk obligations for AI in Annex I regulated products (as amended)

Article 4 requires providers and deployers to take measures to ensure a sufficient level of AI literacy among staff dealing with AI systems. The Digital Omnibus adjusted how AI literacy is framed; read the amended text for the exact obligation that applies to you.

How to prepare, step by step

1. Inventory your AI. List every AI system and model you build, buy or use, including AI features inside third-party SaaS. Record purpose, users, data and where outputs are used.
2. Determine your role for each: provider, deployer, importer, distributor, GPAI model provider.
3. Classify the risk. Screen against Article 5, Annex I, Annex III and Article 50. Document the reasoning, especially if you rely on the Article 6(3) exception.
4. Stop or redesign prohibited uses immediately; they have been banned since February 2025.
5. Implement transparency: AI interaction notices, machine-readable marking of generated content, deepfake disclosures.
6. For high-risk systems, build the Article 9 to 15 requirements into the product: risk management, data governance and bias checks, technical documentation, logging, instructions for deployers, human oversight controls, accuracy and robustness testing, and security.
7. Set up a quality management system, which can be built on ISO/IEC 42001 and your existing ISO/IEC 27001 or ISO 9001 processes.
8. Plan conformity assessment, CE marking, EU database registration and an EU authorised representative if you are outside the EU.
9. Update contracts with customers (deployers) and with model providers, covering documentation, logs, incidents and responsibilities.
10. Train staff on AI literacy and on the oversight duties of their role.
11. Track guidance and standards from the AI Office and the harmonised standards being developed by CEN and CENELEC.

Documentation required

• AI inventory and risk classification records.
• For high-risk providers: Annex IV technical documentation, risk management file, data governance records, test reports, logs, instructions for use, quality management system documentation, EU declaration of conformity, post-market monitoring plan and incident reports.
• For deployers of high-risk systems: records of human oversight assignments, logs kept for at least six months, worker information, and fundamental rights impact assessments where required.
• For GPAI providers: model documentation, downstream information, copyright policy and the public training data summary.
• Transparency notices and content-marking implementation records.
• AI literacy training records.

Conformity assessment and certification

There is no "AI Act certified" label. High-risk systems need a conformity assessment before they are placed on the market. For most Annex III systems, providers use internal control (self-assessment against the requirements, Annex VI). Biometric systems may require a notified body if harmonised standards are not fully applied. Annex I products follow the conformity assessment of their product legislation, with AI requirements included. Systems that comply with harmonised standards published in the Official Journal benefit from a presumption of conformity (Article 40). ISO/IEC 42001 certification can support a quality management system, but does not replace conformity assessment.

Timeline and cost considerations

For most software companies, the near-term work is the AI inventory, classification, removing any prohibited uses, and transparency features for August and December 2026. High-risk providers have until December 2027 (Annex III) or August 2028 (Annex I), but building risk management, data governance, logging and oversight into a product and its documentation is a long programme. Cost factors include legal analysis, engineering for logging, oversight and labelling, evaluation and bias testing, technical documentation, a quality management system, notified body fees where required, an EU authorised representative for non-EU providers, and staff training.

Benefits and penalties

Compliance lets you keep selling AI features in the EU and answers the AI questions now appearing in procurement. The disciplines it requires (inventories, testing, logging, oversight) also reduce the risk of AI failures that damage customers and reputation.

Article 99 sets maximum fines: up to 35 million euros or 7 percent of total worldwide annual turnover, whichever is higher, for prohibited practices; up to 15 million euros or 3 percent for non-compliance with most other obligations; and up to 7.5 million euros or 1 percent for supplying incorrect, incomplete or misleading information to authorities. For SMEs and start-ups, the lower of the two amounts applies. Fines for GPAI model providers are imposed by the Commission (Article 101). Authorities can also order systems withdrawn from the market.

Common mistakes and challenges for IT companies

• Assuming the Act only covers "big AI". A small model used to rank job applicants or set credit limits can be high-risk.
• Assuming non-EU companies are out of scope. Output used in the EU brings providers and deployers in.
• Not knowing what AI is in the product. Third-party SDKs and SaaS features add AI you did not build.
• No logging. Article 12 logging and deployer log retention need infrastructure designed early.
• Human oversight in name only. A reviewer who cannot see why the system decided, or cannot override it, does not meet Article 14.
• Ignoring the GDPR. AI systems that process personal data must meet both laws.
• Planning to the old dates or to rumours. Use the amended text and check guidance regularly.

Best practices, tools and technologies

• Maintain an AI register alongside your asset inventory, with owners and risk classification.
• Use model cards and data sheets as the starting point for technical documentation.
• Log inputs, outputs, model versions and human interventions with retention and access control.
• Build evaluation pipelines that measure accuracy, robustness and bias on representative data before each release.
• Design review queues and override controls for decisions that affect people.
• Implement content provenance and watermarking standards for generated media where they fit.
• Use ISO/IEC 42001, ISO/IEC 23894 and the NIST AI RMF to structure governance and risk management.

We build these controls into AI features as part of our custom software development work.

How the AI Act relates to other frameworks

Framework	Relationship
ISO/IEC 42001	Certifiable AI management system that helps organise the quality management and risk processes the Act requires.
NIST AI RMF	Voluntary US framework; useful method, no legal effect in the EU.
GDPR	Applies in parallel when personal data is processed; the Digital Omnibus clarified their relationship.
Cyber Resilience Act	Security requirements for products with digital elements; high-risk AI that complies with the CRA can be presumed to meet some AI Act cybersecurity requirements.
Colorado AI Act	A US state law on automated decisions in areas such as hiring and lending, narrowed in 2026 to notices, explanations and human review, effective 2027.
India AI Governance Guidelines	Non-binding guidance in India; no equivalent statute.

Real-world implementation examples

Large general-purpose AI model providers signed the GPAI Code of Practice in 2025 and published training data summaries, giving the first public examples of AI Act documentation. National authorities and the AI Office have published guidance on prohibited practices and the AI system definition.

An illustrative case for an HR-tech SaaS vendor: its CV ranking feature falls under Annex III (employment). The company becomes a provider of a high-risk system with obligations from December 2027. It starts by logging every ranking with model version and inputs, building a reviewer interface that shows the main factors and lets recruiters override, measuring selection rates across groups on test data, and writing instructions for use that state limitations clearly. Its EU customers, as deployers, will need those logs and instructions to meet their own duties, so the work also becomes a sales advantage.

Next steps

1. Build your AI inventory and classify every system and your role.
2. Remove anything prohibited and ship transparency features for 2026.
3. For high-risk systems, plan the Articles 9 to 17 work against the 2027 or 2028 date.
4. Align with ISO/IEC 42001 or the NIST AI RMF for governance.
5. Take legal advice; the Act is complex and guidance continues to develop.

Talk to us about the engineering side.

Frequently asked questions

Does the EU AI Act apply to companies outside the EU?

Yes. It applies to providers placing AI systems or GPAI models on the EU market wherever they are based, and to providers and deployers outside the EU when the output of their AI system is used in the EU.

When do high-risk obligations apply?

After the Digital Omnibus, from 2 December 2027 for stand-alone Annex III systems and from 2 August 2028 for AI in Annex I regulated products.

Is a chatbot high-risk?

Usually not. Chatbots carry a transparency duty to tell users they are talking to AI. A chatbot becomes high-risk only if used for an Annex III purpose, such as deciding eligibility for public benefits.

We use an LLM API in our product. Are we a GPAI provider?

Usually not. You are the provider of an AI system that integrates a GPAI model. The model provider has the GPAI obligations and must give you information to help you meet yours.

What are the fines under the AI Act?

Up to 35 million euros or 7 percent of worldwide annual turnover for prohibited practices, up to 15 million or 3 percent for other obligations, and up to 7.5 million or 1 percent for misleading information, with lower caps for SMEs.

Is ISO 42001 required by the AI Act?

No. It can support a quality management system and risk management, but it is not a legal requirement and does not replace conformity assessment.

Which AI practices are banned?

Harmful manipulation, exploitation of vulnerabilities, social scoring, crime prediction based solely on profiling, untargeted facial image scraping, emotion recognition at work and school, sensitive biometric categorisation, most real-time remote biometric identification for law enforcement, and, since the Digital Omnibus, AI designed to generate non-consensual intimate imagery and child sexual abuse material.

Do open-source models fall under the AI Act?

AI systems released under free and open-source licences are largely excluded unless they are high-risk, prohibited or subject to transparency duties. Open-source GPAI models have reduced obligations unless they present systemic risk.

What is a deployer?

A natural or legal person that uses an AI system under its authority in a professional activity. Your business customers using your AI features are usually deployers.

Do we need to label AI-generated content?

Providers of generative systems must mark outputs in a machine-readable way, and deployers of deepfakes must disclose them. These duties apply from 2 August 2026, with 2 December 2026 for marking by systems already on the market.

What changed with the Digital Omnibus on AI?

It postponed high-risk obligations, gave a grace period for content marking by existing systems, added a prohibition on AI designed to generate non-consensual intimate imagery and child sexual abuse material, reinstated a simplified registration for some systems, extended some simplifications to medium-sized companies, and strengthened the AI Office's role.

Who enforces the AI Act?

National market surveillance authorities in each member state for AI systems, and the European AI Office within the Commission for general-purpose AI models and certain systems built on them.