EU AI Act Risk Categories: How to Classify Your AI Systems

The four EU AI Act risk categories explained, with current dates after the Digital Omnibus, obligations by category and a classification decision tree.
On this page

Last updated: 18 November 2026

Quick answer: The EU AI Act risk categories are unacceptable risk (prohibited practices), high risk (strict requirements), transparency risk (disclosure duties) and minimal risk (no specific rules). General-purpose AI models have a separate regime. Prohibitions have applied since February 2025, transparency since August 2026, and most high-risk rules from 2 December 2027.

Every AI Act compliance programme starts with the same question: which category does each of our AI systems fall into? The answer decides whether a system is banned, needs a conformity assessment and human oversight, only needs a disclosure, or can be used freely.

The timetable changed in 2026, when the Digital Omnibus on AI moved the high-risk deadlines. This guide explains each of the EU AI Act risk categories, how to classify a system step by step, what each category requires, and the dates that apply as of late 2026.

What are the EU AI Act risk categories?

The AI Act uses a risk-based approach with four tiers for AI systems: unacceptable, high, transparency and minimal risk. The heavier the potential harm to health, safety or fundamental rights, the stricter the rules.

The European Commission’s AI Act overview describes the four levels and notes that the vast majority of AI systems in the EU fall into the minimal or no risk category, such as spam filters and AI in video games. General-purpose AI models, such as large language models, sit alongside this pyramid with their own obligations.

CategoryLegal basisExamplesMain consequenceApplies from
Unacceptable riskArticle 5Social scoring, emotion recognition at work, untargeted facial image scrapingBanned2 February 2025 (new CSAM and intimate content ban: December 2026)
High riskArticle 6, Annexes I and IIICV screening, credit scoring, exam scoring, medical device AIRisk management, data governance, documentation, human oversight, conformity assessment2 December 2027 (Annex III); 2 August 2028 (Annex I)
Transparency riskArticle 50Chatbots, deep fakes, AI-generated public interest textDisclosure and labelling2 August 2026
Minimal riskNo specific rulesSpam filters, AI in video games, inventory forecastingVoluntary codes of conduct; AI literacy still appliesNot applicable
General-purpose AI modelsChapter VLarge language and multimodal foundation modelsDocumentation, copyright policy, extra duties for systemic risk models2 August 2025

Common mistake: Classifying a vendor product once and never revisiting it. The category depends on the intended purpose and how you use the system, so the same tool can be minimal risk in one department and high risk in another.

What counts as unacceptable risk?

Unacceptable-risk AI practices are banned outright under Article 5 of the AI Act, Regulation (EU) 2024/1689. The Commission now lists nine prohibited practices.

  1. Harmful AI-based manipulation and deception.
  2. Harmful AI-based exploitation of vulnerabilities.
  3. Social scoring.
  4. Individual criminal offence risk assessment or prediction based solely on profiling.
  5. Untargeted scraping of the internet or CCTV to build facial recognition databases.
  6. Emotion recognition in workplaces and education institutions.
  7. Biometric categorisation to deduce certain protected characteristics.
  8. Real-time remote biometric identification for law enforcement in publicly accessible spaces.
  9. AI systems that generate non-consensual sexually explicit or intimate content or child sexual abuse material.

The first eight have applied since 2 February 2025. The ninth was added by the Digital Omnibus on AI, Regulation (EU) 2026/1744, and the Commission states that it comes into effect in December 2026. Several of the original prohibitions have narrow exceptions, for example emotion recognition for medical or safety reasons.

Prohibited practices carry the highest fines: up to EUR 35 million or 7% of worldwide annual turnover, whichever is higher, under Article 99(3). For SMEs and start-ups, Article 99(6) applies whichever of the two amounts is lower.

How does the AI Act define high-risk AI?

An AI system is high risk in two situations under Article 6: when it is a safety component of, or itself, a product covered by EU harmonisation legislation in Annex I that needs third-party conformity assessment, or when it falls into one of the use cases listed in Annex III.

Annex I: AI in regulated products

Annex I covers product laws such as those for machinery, toys, medical devices, lifts and radio equipment. If AI is a safety component of such a product, or is the product, and the product must undergo third-party conformity assessment, the AI system is high risk. After the Omnibus, these obligations apply from 2 August 2028.

Annex III: stand-alone high-risk use cases

Annex III lists eight areas. AI systems intended for the specific uses described in each area are high risk, with requirements applying from 2 December 2027.

  • Biometrics, where permitted by law.
  • Critical infrastructure.
  • Education and vocational training.
  • Employment, workers’ management and access to self-employment.
  • Access to essential private and public services and benefits, including credit scoring and life and health insurance pricing.
  • Law enforcement.
  • Migration, asylum and border control management.
  • Administration of justice and democratic processes.

The Article 6(3) exception

An Annex III system is not high risk if it does not pose a significant risk of harm and it only performs a narrow procedural task, improves the result of a previously completed human activity, detects decision-making patterns without replacing human assessment, or performs a preparatory task. This exception never applies where the system profiles natural persons. A provider relying on it must document the assessment before placing the system on the market and is subject to registration under Article 49(2).

In practice: Record your Article 6(3) reasoning even when you are a deployer. If a regulator or customer questions why a CV-sorting tool is not treated as high risk, the provider’s documented assessment and your own notes on how you use it are your evidence.

What do the transparency and minimal risk categories require?

Transparency-risk systems must tell people they are dealing with AI or AI-generated content. Minimal-risk systems have no specific obligations under the Act, though the general duties on AI literacy and prohibited practices still apply to everyone.

Article 50 has applied since 2 August 2026. Providers must design chatbots so people know they are interacting with a machine, unless it is obvious, and must mark synthetic audio, image, video and text in a machine-readable way. Deployers must disclose deep fakes and AI-generated text published to inform the public on matters of public interest, unless the text has undergone human editorial review. The Omnibus gave generative systems placed on the market before 2 August 2026 until 2 December 2026 to meet the marking requirement.

Breaches of Article 50, like breaches of most operator obligations including high-risk duties, can lead to fines of up to EUR 15 million or 3% of worldwide annual turnover under Article 99(4).

How do you classify an AI system? A decision tree

Classify each system by asking the questions below in order and stopping at the first yes. Record the answer, the reasoning and the date for every system in your AI inventory.

  1. Is it an AI system at all? Check the Article 3(1) definition. Simple rule-based software may fall outside it.
  2. Is the use prohibited under Article 5? If yes, stop using it or change the use.
  3. Is it a safety component of, or a product under, Annex I legislation that needs third-party conformity assessment? If yes, it is high risk (Annex I).
  4. Is it intended for a use listed in Annex III? If no, go to step 6.
  5. Does the Article 6(3) exception apply, and does the system avoid profiling? If the exception applies, document it; otherwise, it is high risk (Annex III).
  6. Does it interact with people, generate synthetic content, or perform emotion recognition or biometric categorisation? If yes, Article 50 transparency duties apply.
  7. None of the above? It is minimal risk. Apply AI literacy measures and your internal AI policy.
  8. Separately: if you provide a general-purpose AI model, check Chapter V obligations regardless of the system classification.

Note that categories can stack. A high-risk recruitment chatbot must meet both the high-risk requirements and the Article 50 transparency duties.

Running this classification across dozens of systems is easier with a shared control library. Enactia’s AI-powered GRC platform maps AI Act requirements to ISO 42001, GDPR and security frameworks in its Compliance Universe, so one control can evidence several obligations. To classify your own AI inventory, start your 14-day free trial.

What obligations follow from each category?

The category determines the workload. High-risk systems carry by far the most obligations, split between providers who build them and deployers who use them.

CategoryProvider obligationsDeployer obligations
UnacceptableDo not place on the marketDo not use
High riskRisk management, data governance, technical documentation, logging, transparency to deployers, human oversight design, accuracy and cybersecurity, quality management, conformity assessment, registration, post-market monitoringUse according to instructions, human oversight, input data quality, monitoring, log retention, informing workers and affected people, fundamental rights impact assessment for certain deployers
TransparencyDesign for disclosure; machine-readable marking of synthetic contentDisclose deep fakes, public interest AI text, emotion recognition and biometric categorisation
MinimalNone specific; voluntary codesNone specific; AI literacy

High-risk obligations connect closely to existing governance. Your enterprise risk management process, vendor risk assessments for AI suppliers and AI use policy should all reference the classification. For financial entities, our article on the overlap between DORA, NIS2 and the AI Act shows how to avoid duplicate work.

Key takeaways

  • The EU AI Act risk categories are unacceptable, high, transparency and minimal risk, with a separate regime for general-purpose AI models.
  • Eight prohibitions have applied since 2 February 2025; the new ban on AI-generated CSAM and non-consensual intimate content comes into effect in December 2026.
  • High-risk rules apply from 2 December 2027 for Annex III systems and 2 August 2028 for Annex I products.
  • Article 50 transparency duties have applied since 2 August 2026.
  • Classification depends on intended purpose and use, so review it whenever use changes.

Frequently asked questions

How many risk categories are in the EU AI Act?
There are four risk categories for AI systems: unacceptable risk, which is prohibited; high risk, which carries strict requirements; transparency risk, which requires disclosure; and minimal risk, which has no specific obligations. General-purpose AI models are regulated separately under Chapter V, with additional obligations for models that pose systemic risk.

Is a chatbot high risk under the AI Act?
Usually not. Most customer service chatbots fall into the transparency category, meaning people must be told they are interacting with AI unless it is obvious. A chatbot becomes high risk if it is used for an Annex III purpose, such as assessing job candidates or deciding access to essential services, in which case both sets of rules apply.

When do the high-risk requirements apply?
Following the Digital Omnibus on AI, which entered into force on 27 July 2026, requirements for stand-alone high-risk systems listed in Annex III apply from 2 December 2027. Requirements for high-risk AI in products covered by Annex I legislation, such as medical devices and machinery, apply from 2 August 2028.

Who decides which risk category applies?
The provider classifies the system based on its intended purpose and documents that assessment, including any reliance on the Article 6(3) exception. Deployers should check the classification against how they actually use the system, because changing the intended purpose can turn a deployer into a provider of a high-risk system under Article 25.

What are the fines for each risk category?
Using or offering prohibited practices can lead to fines of up to EUR 35 million or 7% of worldwide annual turnover. Breaches of high-risk and transparency obligations can lead to fines of up to EUR 15 million or 3%. Supplying incorrect or misleading information to authorities can cost up to EUR 7.5 million or 1%. SMEs face the lower of the two amounts.

Start with an accurate AI inventory

Every obligation in the AI Act flows from classification. Build an inventory of the AI systems you provide and use, run each through the decision tree, record the reasoning and review it whenever the use changes. With high-risk rules applying from December 2027, the organisations that classify now will have time to close gaps calmly. Our guide to ISO 42001 certification shows how an AI management system can keep that inventory current.

To discuss your AI Act programme, contact us or book a demo or start your 14-day free trial.

This article is for general information and is not legal advice.

Written by
Share
Built for MSPs

Turn compliance into recurring revenue. Launch in days, not months.

Governance, risk & compliance made simple

Fewer spreadsheets. Faster compliance. See Enactia in action.

Ready to make compliance simple?

See how Enactia brings governance, risk and compliance into one AI-powered platform. Book a demo now.