EU AI Act: Risk Tiers, Obligations, Deadlines, and Penalties

The EU AI Act (Regulation (EU) 2024/1689) is the first comprehensive law from a major global power governing how artificial intelligence is built, sold, and used. It sorts AI systems into risk tiers, bans the applications considered most harmful, imposes detailed compliance duties on high-risk systems, and layers lighter transparency rules on the rest. Most of its obligations take effect on August 2, 2026, and fines can reach €35 million or 7% of worldwide annual turnover, whichever is higher.¹

Who Has to Comply

The Act reaches well beyond Europe’s borders. It applies to any company that places an AI system on the EU market or puts one into service in the EU, no matter where the company is based.² A firm with no European offices is still covered if the output of its AI system, whether a score, a recommendation, a decision, or generated content, is used inside the EU.

The rules also bind deployers (organizations using AI in their operations) located in the EU, along with importers, distributors, and manufacturers who embed AI into physical products.² Non-EU providers of high-risk systems must appoint an authorized representative established in the EU to act on their behalf.³ If your AI touches the European market in any meaningful way, geography offers no shelter.

AI Practices That Are Banned Outright

Article 5 sets absolute prohibitions. These practices have been unlawful since February 2, 2025, and no amount of compliance work will make them acceptable.⁴ The banned categories are:

Manipulative or deceptive AI that distorts behavior in ways likely to cause physical or psychological harm, including systems exploiting vulnerabilities tied to age, disability, or social circumstance.⁵
Social scoring by public authorities or private companies where the resulting score leads to unfair treatment in unrelated contexts or treatment disproportionate to the behavior.⁶
Untargeted scraping of facial images from the internet or CCTV to build or expand facial recognition databases.⁵
Individual predictive policing based solely on profiling or personality traits. Systems that support human assessment already grounded in objective facts linked to criminal activity are not caught.⁶
Emotion recognition from biometric data in workplaces and schools. Medical and safety uses in those settings remain permitted.⁵
Real-time biometric identification in public spaces for law enforcement, permitted only in narrow situations: searches for specific victims of abduction or trafficking, preventing an imminent terrorist attack or threat to life, and identifying suspects for certain serious offenses.⁶

The workplace emotion recognition ban catches many products already on sale. Tools that claim to measure employee engagement or student attention through facial analysis are now illegal in the EU unless they serve a medical or safety purpose.

How the Risk Tiers Work

Everything not banned falls into a tiered framework, and the higher the risk, the heavier the compliance load.¹ High-risk systems, defined in Article 6 and Annex III, are those whose failure or bias could seriously affect safety, livelihoods, or fundamental rights.⁷ Annex III lists eight areas:

Biometrics, including remote identification and biometric categorization inferring sensitive attributes
Critical infrastructure safety components (digital infrastructure, road traffic, water, gas, electricity)
Education (admissions, evaluating learning outcomes, exam monitoring)
Employment (recruiting, filtering applicants, evaluating candidates, monitoring workers)
Essential services (creditworthiness, public benefits eligibility, emergency call triage, health and life insurance risk pricing)
Law enforcement (victim risk assessments, evidence reliability, recidivism prediction)
Migration and border control (asylum, visa and residence decisions, identification at borders)
Justice and democracy (researching and applying the law in court, tools that could influence elections or voting behavior)

Limited-risk systems face only transparency duties under Article 50, described below. Minimal-risk applications, which include the great majority of AI in use today, spam filters, recommendation engines, AI-enhanced games, carry no new obligations. Harmless AI was deliberately left alone.

What High-Risk Providers and Deployers Must Do

If your system lands in a high-risk category, Articles 8 through 15 spell out a long list of duties you must meet before you can legally sell or deploy it.⁸

Design and Development

Providers must run a risk management process across the whole lifetime of the product, not only during initial development. Training data must meet governance standards ensuring it is relevant, representative, and as free from errors as possible; this is where bias prevention sits, and the responsibility for skewed data lies with you. Systems must log events automatically so regulators can reconstruct what happened after the fact. Human oversight has to be built in so a person can understand outputs, override them, or stop the system.⁹ Technical documentation must be prepared and kept current. Only then does the system qualify for the CE marking that authorizes it for the European market.¹⁰

Post-Market Monitoring

Compliance does not end at launch. Providers must maintain a post-market monitoring system that collects and analyzes performance data in real-world conditions throughout the product’s life, documenting how they will detect drift, spot recurring problems, and report serious incidents.¹¹ Keeping a system compliant against unpredictable inputs is a permanent job, not a one-off project.

Fundamental Rights Impact Assessment

Some deployers must complete a fundamental rights impact assessment (FRIA) before switching on a high-risk system. This applies to public bodies, private organizations providing public services, and companies operating in sensitive domains such as credit scoring and health insurance.¹² The assessment must identify affected rights, describe who is likely to be impacted, evaluate specific risks of harm, and explain human oversight and complaint mechanisms. It can be combined with an existing GDPR data protection impact assessment where they overlap, but the FRIA must still cover ground beyond data protection, such as non-discrimination and access to justice.

Transparency Rules for Chatbots, Deepfakes, and Generated Content

AI systems that don’t reach the high-risk threshold but still interact with people or generate content face disclosure duties under Article 50. Providers of chatbots and similar systems must make sure users know they are dealing with a machine, unless that fact would be obvious to any reasonable person. Providers of systems that generate synthetic audio, images, video, or text must mark the outputs in a machine-readable format so they can be detected as AI-generated.¹³

Deployers of deepfake technology must disclose that content is artificially created or manipulated, with a narrower disclosure duty for artistic, creative, satirical, or fictional works. Deployers of permitted emotion recognition or biometric categorization systems (outside the banned workplace and school contexts) must inform the people exposed to them.¹³

Rules for General-Purpose AI Models

Chapter V creates a separate regime for general-purpose AI models, the large-scale systems, including large language models, capable of a wide range of tasks rather than a single specialized function.¹⁴ Because these models are integrated into countless downstream applications, the rules apply at the model level.

Every provider of a general-purpose AI model must prepare technical documentation, adopt a policy to comply with EU copyright law, and publish a summary of the content used to train the model.¹⁴ That training-data summary gives rights holders a route to see whether their copyrighted material was used, an actively disputed issue in the industry.

Models whose training used more than 10^25 floating-point operations are presumed to carry systemic risk.¹⁵ Providers of those models must run model evaluations and adversarial testing, report serious incidents to the European AI Office, and ensure adequate cybersecurity.¹⁴ The Commission can update the threshold as computing capabilities evolve.

When the Rules Take Effect

The Act entered into force on August 1, 2024, and obligations arrive in phases.⁴

February 2, 2025: Prohibitions on banned practices took effect, along with general AI literacy provisions.
August 2, 2025: Rules for general-purpose AI models became enforceable, along with governance and penalty provisions.
August 2, 2026: The bulk of the Act applies. Annex III high-risk rules and Article 50 transparency obligations kick in, member states must have at least one AI regulatory sandbox operational, and national enforcement begins.
August 2, 2027: Rules for high-risk AI embedded in regulated products (such as medical devices and machinery) take effect. Providers of general-purpose AI models already on the market before August 2025 must reach full compliance by this date.

The high-risk timeline is not settled. Legislative proposals in the EU’s Digital Omnibus package could push the Annex III deadline to December 2027 if harmonized technical standards are not finalized in time. If those amendments are not enacted before August 2026, the original date stands. Plan for August 2026 and watch the legislative process.

Penalties

Enforcement is split between the European AI Office at the EU level and national competent authorities in each member state. These bodies can investigate companies, demand technical documentation, and impose fines calculated on worldwide revenue. The tiers are:

Prohibited practices: up to €35 million or 7% of global annual turnover, whichever is higher.
Other compliance failures (data governance, transparency, deployer obligations, and similar): up to €15 million or 3% of global annual turnover, whichever is higher.
Providing misleading information to regulators or notified bodies: up to €7.5 million or 1% of global annual turnover, whichever is higher.

For the largest AI developers, 7% of worldwide turnover dwarfs any flat figure. Non-compliance with the fundamental rights impact assessment carries a separate operational risk: regulators can order the suspension of the system, which for some organizations would hurt more than any fine.

Relief for Smaller Companies

The Act is not purely restrictive. Member states must establish at least one AI regulatory sandbox by August 2026, giving companies a supervised environment to test systems without facing the full compliance load during testing.¹⁶ Small and medium-sized enterprises and startups get priority access when capacity is limited.

Penalties are also softened for smaller players. For SMEs and startups, each fine is capped at the lower of the flat euro amount or the percentage-of-turnover figure, rather than the higher of the two used against large companies. A startup with modest revenue pays based on the percentage, not the multimillion-euro flat number.

1
Shaping Europe’s digital future. AI Act
2
AI Act Service Desk. Article 2 – Scope
3
EU Artificial Intelligence Act. Article 22 – Authorised Representatives of Providers of High-Risk AI Systems
4
AI Act Service Desk. Timeline for the Implementation of the EU AI Act
5
AI Act Service Desk. Article 5 – Prohibited AI Practices
6
EU Artificial Intelligence Act. Article 5 – Prohibited AI Practices
7
AI Act Service Desk. Guidelines on the Classification of High-Risk AI Systems
8
EU Artificial Intelligence Act. Section 2 – Requirements for High-Risk AI Systems
9
EU Artificial Intelligence Act. Section 2 – Requirements for High-Risk AI Systems – Section: Article 14
10
EU Artificial Intelligence Act. Article 48 – CE Marking
11
EU Artificial Intelligence Act. Article 72 – Post-Market Monitoring by Providers and Post-Market Monitoring Plan for High-Risk AI Systems
12
AI Act Service Desk. Article 27 – Fundamental Rights Impact Assessment for High-Risk AI Systems
13
EU Artificial Intelligence Act. Article 50 – Transparency Obligations for Providers and Deployers
14
Shaping Europe’s digital future. General-Purpose AI Models in the AI Act – Questions and Answers
15
EU Artificial Intelligence Act. Article 51 – Classification of General-Purpose AI Models as General-Purpose AI Models with Systemic Risk
16
EU Artificial Intelligence Act. Article 57 – AI Regulatory Sandboxes