These are the key topics covered in this guide:

  1. What is the EU AI Act?
  2. Who does it apply to?
  3. The four risk categories explained
  4. What high-risk AI systems must comply with
  5. Obligations for general-purpose AI (GPAI) model providers
  6. The enforcement timeline: what is already in force
  7. Penalties for non-compliance
  8. Common mistakes organisations make
  9. How PrivaLex can help

ThThe EU Artificial Intelligence Act entered into force on 1 August 2024 and introduced the EU’s first comprehensive legal framework for artificial intelligence. It is not a future-only requirement. Several obligations already apply, while other requirements are being introduced in stages.

If your organisation develops, provides, deploys, imports, distributes or uses AI systems in a professional context, the first task is to understand your role, the systems in scope and the risk category that applies. The answer is not determined by company size alone. It depends on the system, its purpose, the people affected and how the system is placed on the market or used.

What is the EU AI Act?

The AI Act is a directly applicable EU regulation establishing harmonised rules for the development, placement on the market and use of AI systems. It uses a risk-based approach. Higher-risk systems face more demanding requirements, while certain practices are prohibited and some low-risk uses have limited AI Act obligations.

The regulation can apply to organisations established outside the EU. However, the scope should be assessed by reference to the activities and actors covered by the Regulation, not by assuming that every AI output used by anyone in Europe automatically brings an organisation within every requirement. The GDPR may apply at the same time where personal data is processed.

A company’s place of incorporation does not, by itself, determine whether the EU AI Act applies. For example, an organisation structured through a second-citizenship jurisdiction such as São Tomé and Príncipe may still be covered if it places an in-scope AI system or general-purpose AI model on the EU market, provides it to EU users, or operates in another situation covered by the Act’s territorial rules. The organisation should therefore document its role, activities, users and EU market connections rather than relying solely on its registered jurisdiction. 

Who does the AI Act apply to?

The AI Act uses a value-chain model. The same organisation may have different roles for different systems, and those roles can change when a company modifies or rebrands a third-party system.

Providers

Providers develop an AI system or general-purpose AI model, or have one developed, and place it on the market or put it into service under their own name or trademark. Provider obligations can include risk management, technical documentation, quality controls, instructions, monitoring and conformity assessment, depending on the system.

Deployers

Deployers use an AI system under their authority in a professional context. A company using an external recruitment, credit-scoring, customer-service or employee-monitoring tool may still have deployer obligations. Buying the tool from a third party does not transfer every responsibility to the vendor.

Importers, distributors and product manufacturers

Importers place on the EU market a system bearing the name or trademark of an organisation outside the EU. Distributors make a system available in the supply chain without being the provider or importer. Product manufacturers may become providers where an AI system is placed on the market or put into service together with their product under their own name or trademark.

General-purpose AI providers and downstream organisations

Providers of general-purpose AI models have specific obligations. Organisations that integrate, fine-tune, distribute or deploy those models may also have duties depending on what they do, whether they make significant modifications and whether they become a provider of a new system. Role classification should therefore be recorded for each use case.

Using AI in finance, healthcare, HR, education, critical infrastructure or public services does not automatically make every system high-risk. The specific intended purpose, Annex III listing, product relationship, exceptions and other facts must be assessed.

4 AI Act risk categories

1. Prohibited or unacceptable risk

Certain practices are prohibited because they create unacceptable risks to safety, fundamental rights or human autonomy. Examples include certain manipulative or subliminal techniques, social scoring and some forms of predictive policing. Real-time remote biometric identification in publicly accessible spaces is subject to strict rules and narrow law-enforcement exceptions.

The prohibited-practices rules and AI-literacy provisions began applying on 2 February 2025. Organisations should assess AI uses against the prohibited-practice rules before investing in implementation or procurement.

2. High risk

High-risk systems include certain AI uses in areas such as employment, education, access to essential services, critical infrastructure, law enforcement, migration and biometric identification. Annex III lists specific use cases, while some AI systems embedded in regulated products fall under a separate category.

Examples can include automated recruitment tools, credit assessment, educational admissions, clinical decision support and systems used in access to public services. Classification is use-case specific, so the presence of AI in a sector is not enough on its own.

3. Transparency risk

Some systems have specific transparency duties under Article 50. Examples include chatbots and systems that interact directly with people, along with certain AI-generated or manipulated content. Depending on the system, users may need to be informed that they are interacting with AI, and synthetic content may need to be identifiable or labelled.

The main transparency obligations apply from 2 August 2026. Organisations should identify affected interfaces, content-generation workflows and user notices before that date, while checking the applicable exceptions and technical requirements.

4. Minimal or limited risk

Many systems, such as spam filters, video-game features and some recommendation tools, fall outside the Act’s specific mandatory obligations for higher-risk categories. That does not mean they are free from all legal requirements. The GDPR, consumer-protection law, employment law, copyright rules and contractual duties may still apply.

What must high-risk AI systems do?

High-risk obligations depend on the type of system and the organisation’s role. A practical compliance programme should link each requirement to an owner, a process and evidence.

Risk management

Providers need a continuous risk-management process covering the AI system’s lifecycle. Deployers also need to identify operational risks, follow instructions, monitor use and escalate problems. The assessment should consider foreseeable misuse, affected people, bias, safety, security, accuracy, human oversight and changes over time.

Data governance

Training, validation and testing data must be appropriate to the intended purpose and managed with suitable data-governance processes. Evidence may include data-source records, quality checks, provenance, representativeness assessments, access controls and decisions about limitations or bias.

Technical documentation and records

Providers must maintain technical documentation and records appropriate to the system. Automatic logging can support traceability, post-market monitoring and incident investigation. Organisations should decide which records are generated automatically, which require human review and how long they must be retained.

Transparency and human oversight

Providers must give deployers information and instructions that allow the system to be used appropriately. High-risk systems must support meaningful human oversight, including intervention, review, override or stopping the system where necessary. Human oversight should be assigned to people with sufficient competence, authority and time.

Accuracy, robustness and cybersecurity

High-risk systems must meet applicable requirements for accuracy, robustness and cybersecurity. Testing should cover expected performance, foreseeable errors, manipulation, security weaknesses and material changes. The organisation should record test results, acceptance criteria, remediation and approval.

Conformity assessment and registration

Many high-risk providers must complete a conformity-assessment process before placing the system on the market and applying CE marking. Some systems require the involvement of a notified body. Registration and post-market obligations also depend on the system and the provider’s role. An ISO 42001 programme can support governance and evidence, but it does not replace an AI Act conformity assessment.

ISO 42001 can provide a structured management-system foundation for risk management, data governance, documentation, oversight, training, internal audit and continual improvement.

Obligations for general-purpose AI models

The AI Act includes a specific framework for general-purpose AI models, such as versatile large language or image-generation models that can perform a wide range of tasks. The relevant obligations began applying on 2 August 2025, subject to transitional provisions for certain models already placed on the market.

GPAI providers may need to:

  • Maintain technical documentation about development, testing and evaluation.
  • Provide information to downstream providers about capabilities, limitations and appropriate use.
  • Publish a summary of training content using the Commission’s template.
  • Respect applicable EU copyright rules, including text-and-data-mining opt-outs.
  • Meet additional duties where the model presents systemic risk.

The term “model card” is often used for downstream information, but organisations should map the exact documentation required by the Act and its guidance rather than assume that a particular document title is sufficient.

Providers of GPAI models with systemic risk may face additional requirements, including adversarial testing, serious-incident reporting and energy-use information. The GPAI Code of Practice is a voluntary compliance tool, not a replacement for the Regulation or a guarantee of compliance.

Organisations that fine-tune, host, distribute or integrate GPAI models need to assess whether their activities create provider-level obligations. Significant modifications may change the analysis, while minor adjustments do not automatically make the organisation a provider. The Commission’s guidelines on GPAI obligations should be read alongside the facts of the specific use case.

How PrivaLex can help with EU AI Act compliance

At PrivaLex, we help organisations understand how the AI Act applies to their specific systems and build a practical compliance programme. We begin by creating an inventory of AI systems, vendors, models and use cases, then identify each organisation’s role and the possible risk category.

We turn that assessment into a prioritised roadmap. This can include prohibited-practice screening, high-risk classification, transparency review, GPAI role analysis, supplier due diligence, AI literacy requirements, human-oversight procedures, data-governance controls, technical-documentation planning and incident workflows.

For high-risk systems, we help connect legal requirements to evidence. That may involve risk assessments, testing plans, logs, user instructions, monitoring records, approval decisions, corrective actions and management reporting. Where personal data or information security is involved, we coordinate the AI Act work with GDPR, ISO 27001, ISO 27701 and NIS2 activities to reduce duplicated work while keeping each framework’s requirements visible.

Our support can include:

  • AI system inventory and role classification
  • Risk and impact assessment
  • Gap assessment against applicable AI Act requirements
  • ISO 42001 implementation and AIMS integration
  • GPAI documentation and governance support
  • AI policies, human-oversight procedures and incident processes
  • AI literacy and role-specific training
  • Supplier and customer due-diligence support
  • Internal-audit preparation and coordination with the appropriate external assessment or certification body

We distinguish between legal interpretation, governance implementation, technical remediation and formal conformity assessment. This helps organisations build a compliance programme that is proportionate to their actual systems rather than relying on a generic checklist.

Current EU AI Act timeline

The timeline has changed for some high-risk systems, so organisations should not rely on older summaries. The current European Commission AI Act information identifies the following stages:

  • 1 August 2024: The AI Act entered into force.
  • 2 February 2025: Prohibited AI practices and AI-literacy provisions began to apply.
  • 2 August 2025: Governance rules and GPAI obligations began to apply.
  • 2 August 2026: The Act’s general application and transparency obligations apply, subject to specific exceptions and transitional provisions.
  • 2 December 2027: Rules for Annex III high-risk AI systems apply under the current extended timetable.
  • 2 August 2028: Rules for high-risk AI systems embedded in regulated products under Annex I apply.

The practical deadline therefore depends on the system. An organisation deploying a transparency-risk chatbot, using prohibited practices, providing a GPAI model or operating an Annex III high-risk system faces different dates and obligations. Maintain a controlled timeline for each system instead of applying August 2026 to every AI project.

Penalties and enforcement

The AI Act includes a tiered penalty structure, with the maximum amount depending on the infringement, the organisation and the applicable enforcement provisions. The highest fines can reach up to €35 million or 7% of worldwide annual turnover for prohibited practices and certain GPAI systemic-risk infringements. Other breaches can attract lower but still significant limits, including up to €15 million or 3% of worldwide annual turnover, while incorrect or misleading information may attract a separate lower maximum.

These are maximum statutory levels, not automatic fines for every breach. Authorities will consider the facts and applicable provisions. National competent authorities enforce the Act within Member States, while the European AI Office has an EU-level role, especially for GPAI-related oversight and coordination.

5 AI Act implementation mistakes to avoid

1. Assuming every AI use has the same obligations

An internal productivity assistant, a public chatbot and an automated employment tool require different assessments. Record the intended purpose, users, affected people and organisation’s role before choosing controls.

2. Treating a supplier’s contract as a complete compliance solution

Third-party terms do not replace your own deployer obligations. Assess instructions, data flows, monitoring, human oversight, incidents, changes, security and exit arrangements.

3. Waiting until a deadline to create evidence

Risk assessments, testing, logs, training and human-review records take time to establish. Build the evidence process while the system is being designed or procured.

4. Treating ISO 42001 as proof of AI Act compliance

ISO 42001 can support governance, but it does not replace system classification, conformity assessment, CE-marking obligations or legal analysis under the AI Act.

5. Ignoring privacy and security overlap

High-risk AI systems that process personal data may be subject to the AI Act and GDPR at the same time. Data protection impact assessments, records of processing, security controls and breach-response procedures may need to be coordinated. Integrating AI governance with your privacy and security compliance programme is usually more effective than operating separate programmes.

Conclusion

The EU AI Act requires organisations to understand their role, inventory their AI systems and apply obligations according to the system’s risk and use. The most reliable approach is to classify each system, assign ownership, document the applicable duties and build evidence as part of normal operations.

The current timeline includes obligations that already apply, transparency requirements from August 2026, and later dates for different high-risk categories. Do not assume that ISO 42001 certification alone proves legal compliance. Use it as a management-system foundation alongside specific AI Act analysis, GDPR review and technical implementation.

Book a call with PrivaLex to assess how the AI Act applies to your organisation and define a realistic compliance roadmap.

Frequently Asked Questions (FAQs)

Yes. The AI Act entered into force on 1 August 2024. Several obligations are already active: prohibited AI practices have been banned since 2 February 2025, GPAI obligations have applied since 2 August 2025, and the penalty regime is operational. The main obligations for high-risk AI systems apply from 2 August 2026.

Yes, if your AI systems are placed on the EU market or their outputs are used within the EU. The Act has extraterritorial effect comparable to the GDPR. Providers outside the EU that offer AI systems to EU users or organisations must designate an authorised representative established in the EU.

High-risk AI systems are listed in Annex III of the AI Act. They cover eight domains: biometric identification, critical infrastructure, education, employment, essential services (credit, benefits), law enforcement, migration and administration of justice. If your AI system falls within one of these domains and performs one of the listed functions, it is high-risk. The Commission is also developing guidance to clarify borderline cases. A structured classification exercise, which PrivaLex can support, is the right starting point.

A general-purpose AI (GPAI) model is an AI model trained on large amounts of data that can perform a wide range of tasks. Large language models such as GPT, Claude or Llama are typical examples. If your organisation provides (develops, fine-tunes significantly, or distributes) a GPAI model, you have obligations under the Act that have been in force since August 2025. If you use a GPAI model as a deployer without significant modification, you are treated as a deployer rather than a provider.

Yes, significantly. High-risk AI systems that process personal data must comply with both the AI Act and the GDPR simultaneously. Requirements around data quality, data minimisation, documentation and impact assessments overlap and reinforce each other. A Data Protection Impact Assessment (DPIA) is often required under the GDPR for high-risk AI processing, independently of the AI Act’s own conformity assessment requirements. PrivaLex can help you integrate both into a single, coherent compliance programme.

ISO 42001 is the international standard for AI management systems. It provides a structured framework for governing AI risk, documentation, oversight and continuous improvement that maps closely onto the AI Act’s requirements for high-risk systems. Implementing ISO 42001 is not legally required under the AI Act, but it provides a strong, auditable foundation for compliance and can significantly reduce the effort required to meet the Act’s documentation and risk management obligations.

The most important immediate steps are: map all AI systems in use across your organisation, classify each by risk tier, assess whether any GPAI model obligations apply to systems your organisation provides or modifies, and start building your governance documentation for any high-risk systems ahead of the August 2026 deadline. If you are unsure where to start, contact PrivaLex for an initial assessment.

The AI Act can apply to providers, deployers, importers and distributors connected to the EU market, including organizations outside the EU when their AI systems or outputs are placed on the Union market or affect people in the EU. The exact duties depend on the organization’s role, the system’s intended purpose and its risk classification. Companies should map those roles across vendors, subsidiaries and business use cases rather than relying only on where the software is hosted.

Next Step

The AI Act is not a future problem. Prohibited AI has been banned since February 2025. GPAI obligations are active. High-risk deadlines arrive in August 2026. The organisations that will navigate this most smoothly are those that start their classification and documentation work now, not in the months before enforcement begins. Book a call with PrivaLex to understand exactly where your organisation stands and what a proportionate, practical AI Act compliance programme looks like for your business.


Free self-assessment
The EU AI Act deadlines are coming. Is your organisation ready?
Download our self-assessment and find out in under 10 minutes whether your organisation meets the requirements before the obligations kick in.
Download Free Self-Assessment