Commission draft guidance on high-risk AI classification

Article 6 of the EU AI Act gives the rules for the classification of an AI system as high-risk. The EU AI Act is Regulation (EU) 2024/1689, as amended by Regulation (EU) 2026/1744. The Commission draft guidelines on Article 6 give the interpretation of the Commission. For the Commission examples for each Annex I sector and each Annex III use case, see High-risk examples.

The Article 6 classification framework

Providers classify AI systems with two rules in Article 6 of the EU AI Act:

• Article 6(1) and Annex I. The AI system is intended for use as a safety component of a product, or the AI system is a product. The product is in the scope of the Union harmonization legislation in Annex I. Also, that product must have a third-party conformity assessment. See Article 6(1) and Annex I products.
• Article 6(2) and Annex III. The AI system has an intended use that Annex III lists. See Article 6(2) and Annex III uses.

The Commission draft guidelines also give these "two scenarios" (general principles, paragraph (7)). Article 6(3) is a derogation from Article 6(2) only. It is not a different classification rule. If its conditions apply, an AI system that Annex III refers to is not high-risk, unless the system does profiling of natural persons. See Article 6(3) conditions.

For the status of the Commission’s non-binding EU AI Act draft guidance, see Commission guidance.

AI-system test

Only an AI system in Article 3(1) of the EU AI Act can be high-risk (Commission draft, general principles, paragraph (8)). The Article 3(1) test is necessary. But it is not sufficient for the obligations of the AI Act to apply, because Article 2 sets the scope of the AI Act. See Scope of the AI Act.

Intended purpose

The provider’s intended purpose is important for an AI system’s high-risk classification in the EU AI Act (Commission draft, general principles, paragraph (10)). Article 3(12) gives this definition:

‘intended purpose’ means the use for which an AI system is intended by the provider, including the specific context and conditions of use, as specified in the information supplied by the provider in the instructions for use, promotional or sales materials and statements, as well as in the technical documentation;

Reasonably foreseeable misuse is different from the intended purpose. It is a use that is "not in accordance with its intended purpose" (Article 3(13)). The draft also keeps the two apart (paragraph (10)).

Article 9 gives an obligation to have a risk management system for high-risk AI systems. For a high-risk AI system to which Article 9 applies, the risk management system includes "the estimation and evaluation of the risks that may emerge when the high-risk AI system is used in accordance with its intended purpose, and under conditions of reasonably foreseeable misuse" (Article 9(2)(b)). This obligation is not a general obligation of all providers.

The draft gives this guidance for "systems with multiple purposes, as well as for general-purpose AI systems" (general principles, paragraphs (11)–(12)):

• The draft recommends that the provider clearly gives the intended use of the system in the information that Article 3(12) refers to. If the system is "intended for use in a variety of contexts and application", the draft recommends that the provider gives information about each of them.
• The materials of the provider can show the system as "broadly applicable across a generality of contexts and functions". If these materials do not "consistently limit its application or exclude high-risk uses", the intended purpose then also includes high-risk use cases, and the system is high-risk.
• This applies "in particular" where high-risk uses are "feasible and reasonably foreseeable given the system’s functionalities and capabilities".
• If the provider only writes that high-risk uses are not included, for example in the "terms of service", the system can be high-risk. This applies where "the provider’s overall presentation, examples, or product positioning effectively provides for or promotes such uses".
• The draft recommends that the provider gives each limit of use "clearly, concretely, and coherently" in all materials.

The draft tells providers that the "assessment" of the intended use "is responsibility of the provider that is supervised by the relevant competent market surveillance authorities" (paragraph (13)).

A third party that changes the intended purpose of an AI system, and thus makes the system high-risk, can become its provider. See Substantial modification.

Article 6(1) — Annex I products

Providers classify an AI system as high-risk in Article 6(1) of the EU AI Act where these two conditions apply:

Irrespective of whether an AI system is placed on the market or put into service independently of the products referred to in points (a) and (b), that AI system shall be considered to be high-risk where both of the following conditions are fulfilled:

(a) the AI system is intended to be used as a safety component of a product, or the AI system is itself a product, covered by the Union harmonisation legislation listed in Annex I;

(b) the product whose safety component pursuant to point (a) is the AI system, or the AI system itself as a product, is required to undergo a third-party conformity assessment, with a view to the placing on the market or the putting into service of that product pursuant to the Union harmonisation legislation listed in Annex I.

The two conditions

The AI system is high-risk in Article 6(1) only if the two conditions apply together. The Commission draft gives the same interpretation (Annex I part, paragraph (27)).

For condition (a), Article 3(14) gives the definition of a safety component:

‘safety component’ means a component of a product or of an AI system which fulfils a safety function for that product or AI system, or the failure or malfunctioning of which endangers the health and safety of persons or property; for the purposes of this definition, a component fulfils a safety function where its intended purpose is to prevent or mitigate risks to health and safety of persons or property;

Article 6 adds these rules to the two conditions:

• Article 6(1a). AI systems that are "solely used for non-safety related aspects of user assistance, performance optimisation, service efficiency, automation or convenience or quality control" are not safety components.
• Article 6(1b). Although Article 6(1a) applies, AI systems "the failure or malfunctioning of which would endanger health and safety" are safety components.
• Article 6(1c). Condition (b) does not apply to a product that must have a third-party conformity assessment "solely due to risks other than risks to health and safety". Article 6(1c) gives an example: "in particular risks relating to the distribution of radio spectrum or electromagnetic interference that do not affect health and safety".

Regulation (EU) 2026/1744 added the last part of Article 3(14) and the new Article 6(1a)–(1c). The amended Article 3(14) applies from 27 July 2026. On that date, Regulation (EU) 2026/1744 entered into force (its Article 4). Article 6(1a)–(1c) apply from 2 August 2028 for AI systems in Article 6(1) and Annex I (Article 113, point (c)).

The Commission draft has a related interpretation for condition (a). In the Commission draft, the safety-component concept "may be understood as excluding" "public interests protected under the Union harmonisation legislation covered in Annex I that go beyond health, safety and fundamental rights". Examples are "risks related to the use of radio spectrum or electromagnetic compatibility" (Annex I part, paragraph (25)).

For the Commission draft’s gas-appliance and smart-thermostat examples, see Annex I sectoral examples.

For condition (b), the conformity assessment procedure comes from the Union harmonization legislation in Annex I:

• Some legislation lets the manufacturer use internal control with no conditions and with no third-party conformity assessment. Then condition (b) does not apply. The draft gives the Radio Equipment Directive as an example of such legislation: "for aspect of safety the manufacturer can use Module A without mandatory use of harmonised standards or mandatory involvement of a notified body" (Annex I part, paragraphs (56)–(58) and footnote 11).
• Some legislation lets the manufacturer use internal control only if it applies harmonized standards for "all relevant health and safety requirements". The Toy Safety Regulation and the Machinery Regulation are examples (paragraph (56)). In the draft, the selection of that procedure does not change the classification as high-risk in Article 6(1) (paragraphs (57)–(59)).
• For a high-risk AI system in Annex I Section A legislation, the manufacturer can use such internal control only if it also applies harmonized standards or, where applicable, common specifications in Article 41. These must apply to all the requirements in Chapter III Section 2 (Article 43(3)).

Article 43(3) also tells manufacturers that the classification in Article 6(1) does not make a third-party procedure necessary. This applies if the Section A legislation does not make it necessary. Article 43(3) also gives a rule for an AI system in Annex I Section A and in Annex III. The provider of such a high-risk AI system follows the conformity assessment procedure of the Section A legislation.

Annex I legislation

Annex I gives Union harmonization legislation, not products (Commission draft, Annex I part, paragraph (23)). In the draft, the set of legislation in Annex I is "exhaustive" (paragraph (24)). In the draft, products can go into the scope of Annex I, or out of it, only through one of these two changes:

• A change to the scope of legislation that Annex I gives.
• New Union harmonization legislation goes into Annex I.

Annex I point	Legislation
Section A	"List of Union harmonisation legislation based on the New Legislative Framework"
2	Directive 2009/48/EC, "on the safety of toys"
3	Directive 2013/53/EU, "on recreational craft and personal watercraft"
4	Directive 2014/33/EU, "relating to lifts and safety components for lifts"
5	Directive 2014/34/EU, "relating to equipment and protective systems intended for use in potentially explosive atmospheres"
6	Directive 2014/53/EU, "relating to the making available on the market of radio equipment"
7	Directive 2014/68/EU, "relating to the making available on the market of pressure equipment"
8	Regulation (EU) 2016/424, "on cableway installations"
9	Regulation (EU) 2016/425, "on personal protective equipment"
10	Regulation (EU) 2016/426, "on appliances burning gaseous fuels"
11	Regulation (EU) 2017/745, "on medical devices"
12	Regulation (EU) 2017/746, "on in vitro diagnostic medical devices"
Section B	"List of other Union harmonisation legislation"
13	Regulation (EC) No 300/2008, "on common rules in the field of civil aviation security"
14	Regulation (EU) No 168/2013, "on the approval and market surveillance of two- or three-wheel vehicles and quadricycles"
15	Regulation (EU) No 167/2013, "on the approval and market surveillance of agricultural and forestry vehicles"
16	Directive 2014/90/EU, "on marine equipment"
17	Directive (EU) 2016/797, "on the interoperability of the rail system within the European Union"
18	Regulation (EU) 2018/858, "on the approval and market surveillance of motor vehicles and their trailers"
19	Regulation (EU) 2019/2144, "on type-approval requirements for motor vehicles and their trailers"
20	Regulation (EU) 2018/1139, "on common rules in the field of civil aviation", only "where it concerns unmanned aircraft and their engines, propellers, parts and equipment to control them remotely"
21	Regulation (EU) 2023/1230, "on machinery"

Regulation (EU) 2026/1744 removed point 1, the Machinery Directive 2006/42/EC, from Section A. It added the Machinery Regulation (EU) 2023/1230 to Section B as point 21 (its Article 1(41)). For the effect of Section B, see Annex I Section B scope.

AI system and product

Article 6(1) applies "Irrespective of whether an AI system is placed on the market or put into service independently of the products". The Commission draft gives examples: an AI system "supplied, for example, as a software update, an add-on, or a remote service" can be high-risk in Article 6(1). This applies only if all the conditions of Article 6(1) apply (Annex I part, paragraph (22)).

Exceptions

Not all AI systems in products in the scope of Annex I are high-risk. In the Commission draft, "many consumer AI systems, such as smart home appliances, will not meet one or more of the cumulative conditions in Article 6(1) AI Act" (Annex I part, paragraph (21)).

An AI system that is not high-risk in Article 6(1) can be high-risk in Article 6(2), if Annex III refers to it. The Article 6(3) derogation does not apply to an AI system that is high-risk in Article 6(1). See Scope of Article 6(3).

Annex I Section B scope

EU AI Act Article 2(2) limits the provisions for operators of Article 6(1) high-risk systems related to Annex I Section B products:

For AI systems classified as high-risk AI systems in accordance with Article 6(1) related to products covered by the Union harmonisation legislation listed in Section B of Annex I, only Article 6(1), Article 60a and Articles 102 to 112 shall apply. Articles 57, 58 and 59 shall apply only in so far as the requirements for high-risk AI systems under this Regulation have been integrated in that Union harmonisation legislation.

Thus, the other provisions of the AI Act do not apply to these AI systems. For example, Article 4 (AI literacy), Article 5 (prohibited AI practices) and Article 50 (transparency) do not apply to them.

Article 5 does not apply to all AI systems. It applies only in the scope that Article 2 sets. An AI system in Article 3(1) is not, only for that cause, in the scope of Article 5.

Article 2(2) applies only to AI systems that are high-risk in Article 6(1) and related to products in the scope of the Section B legislation. For these AI systems, the Commission draft gives only Article 6(1), Articles 102 to 109 and Article 112 as the provisions that apply (Annex I part, paragraph (60)). The amended Article 2(2) is the law.

Article 6(2) — Annex III uses

Providers classify the AI systems that Annex III refers to as high-risk in Article 6(2) of the EU AI Act, unless Article 6(3) applies.

In the Commission draft, the intended purpose is "the only relevant determinant" for Article 6(2). If the intended purpose includes an Annex III use case, the AI system is high-risk, unless Article 6(3) applies (Annex III part, paragraphs (70)–(71) and (77)). In the draft, "human involvement cannot change the purpose and area in which a system is intended to be used". Thus, it does not change this classification (paragraph (70)).

For the Annex III use cases and their conditions, see High-risk examples and the Article 6(3) exception.

Article 6(3) — Conditions for an exception

Providers can use the derogation in Article 6(3) of the EU AI Act for Annex III systems only on its conditions. It has three subparagraphs:

By derogation from paragraph 2, an AI system referred to in Annex III shall not be considered to be high-risk where it does not pose a significant risk of harm to the health, safety or fundamental rights of natural persons, including by not materially influencing the outcome of decision making.

The first subparagraph shall apply where any of the following conditions is fulfilled:

(a) the AI system is intended to perform a narrow procedural task;

(b) the AI system is intended to improve the result of a previously completed human activity;

(c) the AI system is intended to detect decision-making patterns or deviations from prior decision-making patterns and is not meant to replace or influence the previously completed human assessment, without proper human review; or

(d) the AI system is intended to perform a preparatory task to an assessment relevant for the purposes of the use cases listed in Annex III.

Notwithstanding the first subparagraph, an AI system referred to in Annex III shall always be considered to be high-risk where the AI system performs profiling of natural persons.

The Commission draft gives this interpretation of Article 6(3) (Annex III part, section 2.7):

• The four conditions are "exhaustive, but alternative". One condition is sufficient (paragraph (88)).
• In the draft, "There is no separate or independent assessment to determine whether the AI system poses a significant or any risk of harm besides those conditions" (paragraph (88)).
• In the Commission draft, Article 6(3) is an exception from rules for the protection of fundamental rights. The Commission thus gives a narrow interpretation of the conditions (paragraph (88)).
• The Commission’s interpretation also takes account of the condition "not materially influencing the outcome of decision making" (paragraph (88)).
• In the draft, the use of Article 6(3) "depends on a self-assessment by the provider" (paragraphs (88) and (113)).

The draft also gives a rule for "complex systems". An AI system cannot use Article 6(3) if it "forms part of a complex system where its combined intended purpose or joint outputs materially influence an individual decision within a high-risk use case". This applies also if one of the conditions applies to the system. The same rule applies to "complex interconnected systems, such as agentic AI systems" (paragraph (90)). In the draft, "split architectures are assessed as a whole" (paragraph (75)).

In the draft, "strictly procedural or preparatory functions of an AI system that are linked to high-risk system" can use Article 6(3) only if the two conditions that follow apply (paragraph (76)):

• The function is "genuinely separable" from the AI system.
• The function does not "structure or feed outputs that … materially influence the examination of an individual case".

Four alternative conditions

The Commission draft gives an interpretation of each condition (Annex III part, section 2.7.1):

• (a) Narrow procedural task. Recital 53 of the AI Act gives three examples: "an AI system that transforms unstructured data into structured data, an AI system that classifies incoming documents into categories or an AI system that is used to detect duplicates" (paragraph (91)). In the draft, the condition can apply to AI systems that "categorise, change the format, structure or presentation of data, or change its metadata" (paragraph (92)). A system that gives a "value judgement" of data for a decision does not do only a narrow procedural task. The draft gives two examples: a system that shows data as "‘useful’ or ‘less useful’ for the human assessment", and a system for "attributing a score or ranking to input data" (paragraph (93)).
• (b) "improve the result of a previously completed human activity". In the draft, three elements must apply together: "a human activity has been completed", "the completion of the activity led to a result", and "the result is improved by the AI system" (paragraph (94)). The changes "should not revert or lead to a full revision or replacement of the previously completed human activity" (paragraph (96)). In the Commission draft, the AI system must not be intended to give a materially different result. The draft also recommends that improvements do not change affected persons’ rights, protection, or legal or economic position. This includes persons affected by the system’s output or by decisions based on that output (paragraph (96)).
• (c) Decision-making patterns or deviations. In the draft, three limits apply. The human assessment must be complete. The AI system can only do "an ex-post comparative assessment" against previous patterns or decisions. The AI system must not be intended to replace or influence the completed human assessment without "proper human review" (paragraphs (100)–(102)).
• (d) Preparatory task. In the draft, a preparatory task occurs "prior to the actual assessment process". A narrow procedural task can also occur during the "assessment". If a narrow procedural task occurs before the "assessment" and is preparatory, (a) and (d) can apply together (paragraph (104)). Preparatory tasks, for example "indexing, searching, processing, and linking", add structure to the inputs. But they "do not provide themselves an assessment leading to an outcome" (paragraph (106)). If the output of the system goes into a human assessment, the task is preparatory only if the output is "a general input or factor". If the system gives "a specific recommendation or evaluation of the case", the task is not preparatory (paragraph (108)).

For a task that is not in an Annex III use case, Article 6(3) is not necessary. In the draft, the "anonymisation or pseudonymisation of judicial decisions" is an example. The Commission draft gives this interpretation of Recital 61: AI systems for "ancillary administrative activities that do not affect the actual administration of justice in individual cases" are not in Annex III point 8(a) (paragraph (419)).

Profiling override

In the EU AI Act, an AI system that Annex III refers to is always high-risk if it does profiling of natural persons (Article 6(3), last subparagraph). This applies also if one of the conditions (a) to (d) applies to the system. The rule applies only to AI systems that Annex III refers to. Profiling does not make an AI system high-risk if the system is not in an Annex III use case.

Article 3(52) of the AI Act refers to Article 4(4) of Regulation (EU) 2016/679 for the definition of profiling. The Commission draft also refers to Article 3(4) of Directive (EU) 2016/680. It also refers to Article 3(5) of Regulation (EU) 2018/1725 (Annex III part, paragraph (89)).

In the draft, profiling has three elements, and all of them must apply (paragraphs (110)–(111)):

1. "there must be an automated form of processing"
2. "which is carried out on personal data"
3. "the objective of which must be to evaluate personal aspects relating to a natural person"

In the draft, "a simple classification of individuals based on personal characteristics such as their age, sex, and height does not necessarily lead to profiling" (paragraph (112)). The draft also gives an example of profiling. A system finds "deviations from previous recruitment decision-making patterns", and it also examines "the personal characteristics of the recruiters". This system does profiling. Thus, it cannot use Article 6(3) (paragraph (112)).

Scope of Article 6(3)

In the EU AI Act, Article 6(3) applies only to AI systems that Annex III refers to. It does not apply to an AI system that is high-risk in Article 6(1) (Commission draft, Annex III part, paragraph (87)). Thus, Article 6(3) does not apply to a safety component of a product in Annex I that is high-risk in Article 6(1).

Article 6(4) documentation and Article 49(2) registration

The EU AI Act gives these obligations to "A provider who considers that an AI system referred to in Annex III is not high-risk":

• The provider must document its "assessment" before the system is placed on the market or put into service (Article 6(4)).
• Before the system is placed on the market or put into service, "that provider or, where applicable, the authorised representative shall register themselves and that system in the EU database referred to in Article 71" (Article 49(2)).
• The provider must give the documentation of the "assessment" to national competent authorities "Upon request" (Article 6(4)).

Annex VIII Section B gives the information for this registration. Regulation (EU) 2026/1744 removed points 7 and 9 of Annex VIII Section B (its Article 1(42)). The registration obligation in Article 49(2) did not change.

The Commission draft recommends that the "assessment" includes these four items (Annex III part, paragraph (115)):

• the intended purpose of the AI system,
• why the system is high-risk in Article 6(2),
• which conditions in Article 6(3) apply and why,
• why the system does not "perform profiling".

The draft also recommends a record that the provider can give "at any time upon the request of a market surveillance authority" (paragraph (116)).

Article 80 of the AI Act gives the procedure for a market surveillance authority. If it has "sufficient reason to consider" that such a system is high-risk, it "shall carry out an evaluation of the AI system concerned in respect of its classification as a high-risk AI system based on the conditions set out in Article 6(3) and the Commission guidelines" (Article 80(1)). Article 80(7) makes the provider subject to Article 99 fines where the market surveillance authority establishes these facts during its Article 80(1) evaluation:

• The provider misclassified the AI system as non-high-risk.
• The provider misclassified the system "in order to circumvent the application of requirements in Chapter III, Section 2".

Substantial modification

Article 25(1) of the EU AI Act makes a distributor, importer, deployer or other third party the provider of a high-risk AI system "in any of the following circumstances". It then has the obligations of the provider in Article 16. In each of these points, the AI system is "already placed on the market or put into service":

• (a) It puts its name or trademark on a high-risk AI system. This is "without prejudice to contractual arrangements stipulating that the obligations are otherwise allocated".
• (b) It makes a substantial modification to a high-risk AI system, and the system stays high-risk in Article 6.
• (c) It changes the intended purpose of an AI system that is not classified as high-risk, and the system becomes high-risk in Article 6. This includes a general-purpose AI system.

The Commission draft gives the same three points (general principles, paragraph (14)).

Article 25(1) applies only where the result is a high-risk AI system. It is not a general rule for all AI systems that a third party changes. For a general-purpose AI model, a different test applies to a third party that changes the model. See GPAI models.

Article 3(23) of the EU AI Act gives the definition of a substantial modification:

‘substantial modification’ means a change to an AI system after its placing on the market or putting into service which is not foreseen or planned in the initial conformity assessment carried out by the provider and as a result of which the compliance of the AI system with the requirements set out in Chapter III, Section 2 is affected or results in a modification to the intended purpose for which the AI system has been assessed;

Article 43(4) gives two rules for a high-risk AI system that had a conformity assessment:

• After a substantial modification, the system must have a new conformity assessment procedure. This applies "regardless of whether the modified system is intended to be further distributed or continues to be used by the current deployer".
• Some high-risk AI systems "continue to learn after being placed on the market or put into service". For these systems, some changes are not a substantial modification. These are the changes "that have been pre-determined by the provider at the moment of the initial conformity assessment", and that are in the technical documentation in Annex IV point 2(f).

The draft gives this information: the Commission "is preparing separate Guidelines on the practical application of rules for responsibilities along the AI value chain under Article 25 of the AI Act" (general principles, footnote 6). These guidelines will give more information about the three points of Article 25(1). The definition in Article 3(23) is in the AI Act, and it applies without these guidelines.

Application dates

For EU AI Act application dates and transitional rules for operators, see Application dates, which covers Articles 113 and 111.

Use Modulos

In Modulos, the EU AI Act Frameworks are OFF-1 at Organization level and MFF-1 at AI System level. For the Requirements and the steps, see Operationalizing the EU AI Act in Modulos.

Sources

• Regulation (EU) 2024/1689 (AI Act), as amended, text of 27 July 2026. Articles 2(2), 3(12)–(14), 3(23), 3(52), 6, 9(2), 25(1), 43(3)–(4), 49(2), 80 and 113. Annexes I, III and VIII.
• Regulation (EU) 2026/1744 of 8 July 2026 (Digital Omnibus on AI), Official Journal of 24 July 2026: Article 1(4), (8), (40), (41) and (42), and Article 4.
• The draft guidelines of the European Commission, "Draft Commission guidelines on the classification of high-risk AI systems under Article 6 of Regulation (EU) 2024/1689 (AI Act) for stakeholder consultation", published on 19 May 2026, in three parts:
  • General principles: paragraphs (6)–(14) and footnote 6.
  • Annex I of AI Act: paragraphs (21)–(27), (43), (49), (56)–(60) and footnote 11.
  • Annex III of AI Act: paragraphs (70)–(71), (75)–(77), (87)–(116) and (419).
• The European Commission publication page, "Draft Commission guidelines on the classification of high-risk AI systems", 19 May 2026.

Disclaimer

This page gives general information. It is not legal advice.