The EU AI Act is a comprehensive legal framework for artificial intelligence. It uses a risk-based approach to regulation, meaning the rules an AI system must follow depend on the level of risk it poses to health, safety, or fundamental human rights. This article explains the EU AI Act, how to determine if it applies to your work, and how to identify your compliance obligations.
Key takeaways
- The EU AI Act categorizes AI systems into four risk tiers: unacceptable (banned), high, limited, and minimal. The stricter rules apply to the higher-risk categories.
- The regulation has a wide reach. It applies to any organization that places an AI system on the EU market or whose AI's output is used in the EU, regardless of where the organization is based (Source).
- Your obligations depend on your role. A "provider" who builds an AI system has different duties than a "deployer" who uses one in a business context.
- High-risk systems must meet stringent requirements. These include implementing risk management, ensuring high data quality, enabling human oversight, and maintaining robust cybersecurity (Source).
- Compliance deadlines are staggered. Rules for banned AI practices began applying on 2 February 2025, with other key provisions rolling out through 2 August 2028 (Source).
What is the EU AI Act?
The EU AI Act, officially Regulation (EU) 2024/1689, is a legal framework designed to harmonize rules for artificial intelligence across the European Union (Source). Its main goal is to protect people's health, safety, and fundamental rights from the potential harms of AI. At the same time, it aims to support innovation and the development of trustworthy AI. The Act does this by establishing legally binding obligations for organizations that develop, sell, or use AI systems in the EU.
The core principle of the regulation is its risk-based approach. Instead of treating all AI equally, it concentrates the strictest rules on applications with the highest potential for harm. This means a spam filter is treated very differently from an AI system used to make hiring decisions or diagnose diseases.
How the EU AI Act Works: A 4-Step Framework
To determine your obligations under the Act, you can follow a four-step process. This framework helps you move from the general regulation to specific, actionable requirements for your project.
Step 1: Determine if the Act applies to you
First, check if your activities fall within the Act's scope. The regulation applies if you are a:
- Provider who places an AI system or general-purpose AI model on the EU market, regardless of where you are established.
- Deployer of an AI system who is located or established within the EU.
- Organisation outside the EU, if the output produced by an AI system is used within the EU (Source).
Your specific role—such as provider, deployer, importer, or distributor—determines your exact responsibilities.
Step 2: Classify your AI system's risk level
The Act creates four risk categories, and your system will fall into one of them. This classification is the most critical step, as it dictates your compliance workload. The tiers are Unacceptable Risk (banned), High-Risk, Limited Risk, and Minimal Risk. The classification depends on the AI's intended purpose and context of use, not just its underlying technology.
Step 3: Fulfill your specific obligations
Based on the risk level and your role, you must implement the required measures. For a minimal-risk system, you may have no legal obligations at all. For a limited-risk system, you might only need to ensure users know they are interacting with an AI. For a high-risk system, you will face a long list of requirements, including setting up a risk management system, ensuring data quality, creating technical documentation, and enabling human oversight.
Step 4: Track your compliance deadlines
The AI Act's rules do not all start at once. They are staggered over several years.
- 2 February 2025: Rules on prohibited AI practices apply.
- 2 August 2025: Obligations for general-purpose AI models begin.
- 2 August 2026: The broad application date for most rules, including for many high-risk systems.
- 2 December 2027 and 2 August 2028: Extended deadlines for certain high-risk systems, such as those in specific Annex III use cases (2027) or embedded in regulated products under Annex I (2028).
You must identify which deadlines apply to your specific system and role (Source).
Who Does the EU AI Act Apply To? A Breakdown of Roles
The Act defines specific roles within the AI supply chain, each with distinct responsibilities. Understanding your role is essential for compliance. A person or organization can have more than one role.
| Role | Who They Are | Key Responsibilities |
|---|---|---|
| Provider | An entity that develops an AI system or general-purpose AI model and places it on the market or puts it into service. | Bears the primary responsibility for compliance for high-risk systems. Must conduct conformity assessments, create technical documentation, and register the system in an EU database. |
| Deployer | A person or organization using an AI system under its authority for professional activities (e.g., a company using an AI recruitment tool). | Must use the system according to the provider's instructions, ensure human oversight, and monitor the system's operation. For high-risk systems, they must also report serious incidents. |
| Importer | An entity in the EU that places an AI system from a non-EU country on the EU market. | Has specific obligations for placing an AI system from a non-EU country on the market. |
| Distributor | An entity in the supply chain, other than the provider or importer, that makes an AI system available on the market. | Has specific obligations for making an AI system available on the market after it has been placed by a provider or importer. |
Which AI systems are classified as high-risk?
An AI system is classified as high-risk if it has the potential to adversely impact people's safety or fundamental rights (Source). A system can become high-risk in one of two main ways.
Route 1: It's part of a regulated product. If an AI system is a safety component of a product—or is itself a product—that is already covered by existing EU product safety laws, it is considered high-risk. This applies to products listed in Annex I of the Act, which includes machinery, toys, cars, and medical devices (Source). For example, an AI-powered diagnostic feature in a medical scanner would fall into this category.
Route 2: It's used in a sensitive area. An AI system is also considered high-risk if its intended use falls into one of the specific, sensitive areas listed in Annex III of the Act. These areas include:
- Biometrics and identification
- Management of critical infrastructure (e.g., water, gas, electricity)
- Education and vocational training (e.g., systems for student admission or evaluation)
- Employment and worker management (e.g., software for sorting résumés or evaluating employee performance) (Source)
- Access to essential services (e.g., algorithms for credit scoring or determining eligibility for public benefits)
- Law enforcement, migration, and administration of justice
However, use in one of these areas does not automatically classify a system as high-risk. The Act includes detailed conditions and exceptions that must be assessed (Source).
Worked Example: Deploying a High-Risk Recruitment Tool
Let's walk through a practical scenario to see how the rules apply.
Scenario: A European tech startup, "HireFast," decides to use a third-party AI tool to help screen and rank candidates for its engineering roles.
Step 1: Classification. HireFast is the deployer of the AI system. The system's purpose is to sort and evaluate job applicants. This use case falls under the Annex III category of "employment, worker management" (Source). Therefore, the system is presumed to be high-risk.
Step 2: Due Diligence. Before putting the tool into service, HireFast must perform due diligence. They check the EU's public database to confirm the AI system's provider has registered it. They also request the provider's technical documentation and instructions for use to ensure they understand how the system works, its limitations, and how to use it correctly.
Step 3: Implementation. HireFast cannot simply automate its hiring process. As a deployer of a high-risk system, it must implement effective human oversight. The HR team is trained to use the AI's rankings as an input, not a final decision. They commit to manually reviewing a random sample of all applications, including those ranked low by the AI, to check for potential biases.
Step 4: Monitoring. While the system is in use, HireFast must keep logs of its operation as required by the Act. They monitor the system's outputs for any signs of discriminatory patterns. If they discover a serious incident where the AI system creates a significant risk (e.g., systematically down-ranking candidates from a protected group), they are obligated to report it to the provider and the relevant national authorities.
What Are the Four Risk Tiers of the EU AI Act?
The Act's obligations are tied directly to its four-tiered risk pyramid.
| Risk Level | Description | Examples | Key Obligation |
|---|---|---|---|
| Unacceptable Risk | AI systems that are considered a clear threat to the safety, livelihoods, and rights of people. | Social scoring by governments; AI that uses manipulative subliminal techniques; real-time remote biometric identification in public spaces by law enforcement (with narrow exceptions). | Banned. These systems cannot be placed on the market or used in the EU (Source). |
| High-Risk | AI systems that can have an adverse impact on people's safety or fundamental rights. | AI in medical devices; credit scoring software; systems that sort job applications; AI used in critical infrastructure. | Strict compliance. Must meet requirements for risk management, data quality, technical documentation, record-keeping, transparency, human oversight, and cybersecurity before market entry (Source). |
| Limited Risk | AI systems that pose a specific risk of manipulation or deception. This includes much of what is called generative AI. | Chatbots; AI systems that generate deepfakes; emotion recognition systems. | Transparency. Users must be informed that they are interacting with an AI system or that content is artificially generated or manipulated. |
| Minimal Risk | All other AI systems that pose little to no risk. This is the vast majority of AI in use today. | AI-enabled spam filters; inventory management systems; AI in video games. | No mandatory obligations. Adherence to voluntary codes of conduct is encouraged. |
Common Mistakes to Avoid
As organizations prepare for the AI Act, several common misconceptions have emerged.
Mistake 1: Assuming all AI is now high-risk
The Act is designed to be proportional. The strictest rules apply only to a specific subset of AI systems that fall into the high-risk category. Most AI applications, like spam filters or recommendation engines in e-commerce, will likely be minimal-risk and face no new obligations.
Mistake 2: Believing the Act only applies to EU companies
The Act's reach is extra-territorial. If you are a provider based in the United States and you sell your AI product to customers in the EU, you must comply with the Act's rules (Source). Its scope is determined by market access and use, not company headquarters.
Mistake 3: Confusing a foundation model with a high-risk system
General-purpose AI models, including large foundation models, have their own set of rules under the Act. An application built on such a model is not automatically high-risk. The application's risk level is determined by its specific use case. A chatbot for recipes built on a powerful model is still likely minimal-risk.
Mistake 4: Focusing only on model accuracy
Compliance for high-risk systems goes far beyond technical performance. The requirements cover the entire lifecycle, including risk management processes, data governance procedures, comprehensive technical documentation, logging capabilities, and plans for human oversight (Source). Accuracy is just one piece of a much larger puzzle.
What to learn next
- What Is Generative AI? A Beginner's Guide: Understand the technology behind many of the systems facing new transparency rules.
- AI Alignment Explained: Principles and Practice: Explore the technical and ethical challenges of ensuring AI systems behave according to human values, a core goal of the AI Act.
- Foundation Models Explained: A Beginner's Guide: Learn about the large, general-purpose models that have their own distinct obligations under the new regulation.
