# How to classify a high-risk AI system under the European AI Act

> To know if your AI is high-risk according to the AI Act, check if it is integrated into a regulated product or if its use falls under one of the 8 domains of Annex III. An exception clause exists, but profiling excludes any derogation. Obligations from August 2026.

- Tags: classification système ia haut risque, ai act haut risque, annexe iii ai act, conformité ai act, obligations réglementaires ia, article 6 ai act, exception article 6 ai act, sanctions ai act
- Main keyword: classification système IA haut risque

## Key points

- The AI Act imposes an AI system classification by risk level, with two main paths, 8 critical domains, and an exception clause to master before August 2026.
- The status of your AI determines your obligations, from transparency to full technical documentation.

## Introduction

The European AI Act imposes a classification of artificial intelligence systems according to their risk level. To determine if your AI is high-risk, you must understand two classification mechanisms, examine the 8 critical sectors of Annex III, and master the exception clause before August 2026.

This distinction is not trivial: a CV screening tool may be subject to heavy regulatory obligations, while a simple customer support chatbot generally is not. **Article 6 of Regulation (EU) 2024/1689** structures this classification, which is essential for anticipating your compliance steps.

## Content

## The four risk levels defined by the AI Act

The AI Act framework is based on a risk scale: the more likely an AI system is to impact health, safety, or fundamental rights, the stricter the requirements. Four categories structure the regulation:

### Unacceptable risk: prohibited practices from 2025

Certain AI uses are completely banned as of **February 2, 2025** (see Article 5). This includes subliminal manipulation, social scoring by authorities, predictive policing based solely on profiling, or emotion analysis in the workplace and educational institutions.

### High risk: maximum requirements

AI systems classified as high-risk are permitted but subject to enhanced obligations: technical documentation compliant with [Annex IV](https://compaia.eu/obligations), risk management, data governance, human oversight, and CE marking. This is the most demanding category of the regulation.

### Limited risk: mandatory transparency

AI systems that interact with the public (chatbots, content generators, deepfakes) must comply with the **transparency rules** set out in Article 50: informing the user of the interaction with an AI, clearly labeling generated content, and ensuring technical detection.

### Minimal risk: no specific requirements

Most AI systems, such as spam filters, recommendation tools, or video games, are considered minimal risk and have no specific obligations.

## The two mechanisms for "high-risk" classification

Article 6 of the regulation distinguishes between **two main modalities** for an AI system to be considered high-risk. This distinction directly influences the implementation timeline for obligations.

### 1. Safety component of a regulated product (Article 6§1)

An AI system is classified as high-risk if it meets two cumulative criteria:

- It is integrated as a safety component in a product covered by European harmonization legislation listed in Annex I, or is the product itself.
- This product is subject to a third-party conformity assessment before being placed on the market.

Annex I groups more than 20 European sectoral texts, covering, among others, medical devices, toys, elevators, radio equipment, agricultural vehicles, civil aviation, and maritime safety. For example, an AI software embedded in a Class IIa medical device will automatically fall into this category.

The obligations associated with this classification will apply from **August 2, 2027**, one year after those for path 2.

### 2. Systems falling under Annex III (Article 6§2)

Regardless of any physical product, an AI system is classified as high-risk if its use corresponds to one of the **8 domains of Annex III**. This mechanism concerns the majority of companies and comes into force on **August 2, 2026**.

## The 8 critical sectors of Annex III

Annex III lists eight types of use for which an AI system is **presumed to be high-risk**, due to their impact on rights, health, or safety:

### 1. Biometrics

Includes **remote biometric identification** (real-time or post-event), categorization based on sensitive attributes (ethnic origin, sexual orientation), and **emotion recognition**. Simple identity verification is excluded from the scope.

### 2. Critical infrastructure

Covers systems used as **safety components** in the management of road traffic, water, gas, heating, electricity supply, or in the operation of essential digital infrastructure.

### 3. Education and vocational training

Applies to systems that determine access or assignment to institutions, evaluate learning outcomes, education levels, or monitor behavior during exams.

### 4. Employment and HR management

Concerns **recruitment** tools (job posting, CV screening, candidate evaluation), management of working conditions, promotions, terminations, or employee performance monitoring.

### 5. Essential public and private services

Includes systems that evaluate eligibility for social or health benefits, establish a **credit score** (excluding fraud detection), life and health insurance pricing, emergency call management, or patient triage.

### 6. Law enforcement

Targets systems evaluating risks of criminal offenses, serving as **polygraphs**, analyzing the reliability of evidence, profiling individuals during investigations, or processing criminal offenses.

### 7. Migration, asylum, and border control

Groups systems evaluating security or irregular migration risks, assisting in the examination of asylum or visa applications, or detecting persons in the context of **border surveillance**.

### 8. Justice and democratic processes

Integrates systems assisting judicial authorities in researching and interpreting facts and law, or used to **influence the outcome of elections** or referendums (excluding campaign organization tools without direct interaction with voters).

## Exception to the high-risk presumption: Article 6§3

Listing in Annex III **does not automatically lead to high-risk classification**. Article 6§3 provides for a derogation: a system may be excluded if it does not present a significant risk to health, safety, or fundamental rights.

This exception is applicable if one of the following situations is verified:

1. **Narrow procedural task**: the system performs a well-defined and low-criticality mission (format conversion, document sorting, duplicate detection).
2. **Support for an existing human activity**: the AI complements a human action without decision-making autonomy (text reformulation, linguistic improvement).
3. **Pattern detection without impact on the decision**: the system identifies deviations from previous decisions without influencing human evaluation.
4. **Preparatory action**: the AI performs a task prior to a human evaluation in one of the Annex III domains.

> "An AI system referred to in Annex III shall not be considered to be high-risk where it does not present a significant risk of harm to the health, safety or fundamental rights of natural persons." — Article 6§3, Regulation (EU) 2024/1689

### Strict limit: profiling

A major restriction applies: any AI system performing **profiling of natural persons** (automated data analysis to evaluate performance, health, preferences, behavior) remains **obligatorily high-risk**, with no possibility of exception.

### Mandatory documentation

The provider invoking the exception must **document its analysis before placing the system on the market** (Article 6§4) and make it available to competent authorities. It must also register in accordance with Article 49. [Article 80](https://compaia.eu/diagnostic) provides a procedure allowing authorities to challenge the classification chosen by the provider.

## Deciding on the risk level: a 5-step decision tree

To classify your AI system, follow this path:

1. **Does your solution fall under the definition of AI in Article 3?** If it does not rely on machine learning or inference, it might be out of scope. February 2025 guidelines exclude simple mathematical optimization and deterministic processing.
2. **Does it fall under a prohibited practice (Article 5)?** If yes, it is prohibited as of February 2, 2025.
3. **Is it integrated as a safety component in an Annex I product?** If yes, and if a third-party assessment is required, it is high-risk (path 1, deadline August 2027).
4. **Does it fall into one of the 8 domains of Annex III?** If yes, it is presumed to be high-risk (path 2, deadline August 2026), unless an exception is documented according to Article 6§3.
5. **Does it interact with people or generate synthetic content?** If yes, transparency obligations of Article 50 apply, even without high-risk classification.

Tools like the [free compaia diagnostic](https://compaia.eu/diagnostic) facilitate this journey and quickly identify the risk level and associated obligations.

## Classification examples: typical cases

### High-risk systems

- **Bank credit scoring**: creditworthiness evaluation → Annex III, domain 5, high-risk without exception (profiling).
- **CV screening software**: application analysis → Annex III, domain 4, high-risk, no exclusion possible (profiling).
- **Facial recognition at an airport**: biometric identification → Annex III, domain 1, high-risk.
- **Embedded medical AI**: safety component of a medical device → Annex I, path 1, high-risk.

### Systems potentially excluded thanks to the Article 6§3 exception

- **HR grammar checker**: improves writing without influencing selection → human improvement task, exclusion possible.
- **Duplicate detection in files**: procedural task, no profiling → exclusion possible if documented.

### Systems not concerned by high risk

- **Customer support chatbot**: outside Annex III, but subject to Article 50 transparency.
- **Spam filter**: minimal risk, no specific obligation.
- **E-commerce recommendation**: minimal risk, unless profiling manipulates behavior (in which case, Article 5 may apply).

## Regulatory framework: timeline, delays, and uncertainties

Article 6§5 provided that the European Commission publish, **no later than February 2, 2026**, guidelines to clarify the classification, with concrete examples. This deadline was not met.

After the June 2025 public consultation, the Commission announced it is preparing a draft for comments in early 2026, with final adoption expected in spring 2026.

### Digital Omnibus: towards a delay?

In November 2025, the Commission proposed the **Digital Omnibus**, planning to postpone the application of high-risk obligations for Annex III systems to **6 months after the publication of harmonized standards**, with a deadline of December 2, 2027. This shifts the initial August 2026 deadline by 16 months ([see timeline](https://compaia.eu/echeancier-ai-act)).

This project remains debated in Parliament and the Council. It is not guaranteed to be adopted before August 2026. Companies must therefore prepare based on the initial calendar while monitoring legislative developments.

### Delay in technical standards

The European standardization organizations (CEN/CENELEC) responsible for drafting harmonized standards for AI have also fallen behind. The final texts are not expected to be available before the end of 2026, which reinforces the argument for a delay. In the meantime, one must rely on the regulation and the future common specifications of the Commission.

## Preparing for compliance: recommended steps

Despite the uncertainties regarding the calendar, the substantive requirements of the regulation remain. Here are the actions to anticipate:

1. **Map your AI systems**: inventory all AI systems developed, deployed, or used, including third-party tools and integrated models.
2. **Classify each system**: apply the decision tree (Annex I, Annex III, Article 6§3 exception, Article 50 transparency, minimal risk).
3. **Document exclusions**: if you invoke the Article 6§3 exception, write a detailed assessment justifying the absence of significant risk, to be kept before any market placement.
4. **Identify your role**: provider, deployer, importer, or distributor? [Obligations vary according to your function](https://compaia.eu/obligations) in the value chain.
5. **Launch technical documentation**: for high-risk, prepare a file compliant with Annex IV (description, risk management, data governance, testing, post-market monitoring).
6. **Anticipate conformity assessment**: depending on the sector, an internal or notified body assessment will be required before CE marking.

> Doubting the risk level of your AI? The [free compaia diagnostic](https://compaia.eu/diagnostic) analyzes your use case and tells you precisely your obligations according to the AI Act regulation.

## Official source
- [Source](https://www.aiacto.eu/fr/blog/classifier-systeme-ia-haut-risque-ai-act)