3 Legal interpretation of Article 6(3)(a) and (d)

3.1.1 The overarching condition of no significant risk of harm#s. 14

Article 6(3) first subparagraph provides that an AI system falling within one of the Annex III areas “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.” This overarching condition — the absence of significant risk, including the absence of material influence on decision-making — establishes a general threshold that is both abstract and subjective.54 The “materiality” qualifier is also inherently subjective, and may allow providers to claim that their system's influence on decisions is immaterial even where it significantly shapes the informational basis on which those decisions are made.55 If the overarching condition is too indeterminate to constrain the self-assessment effectively, the specified conditions must serve as the primary contributor to legal certainty.

3.1.2 Article 6(3) as a closed list of carve-outs#s. 14

The second subparagraph sets out four specific conditions under which the first subparagraph “shall apply”. They read as follows, with the conditions that are the focus of this thesis emphasised in bold text:

(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.**56

The list is connected by the disjunctive “or”, indicating that the conditions are alternative rather than cumulative: satisfying any one of the four is sufficient, provided that the overarching condition of no significant risk of harm is also met.

The question arises whether the four conditions are exhaustive: Can a provider invoke Article 6(3) on a basis not enumerated in the list? Do the four conditions serve as safe harbours, sufficient but not necessary, rather than as an exclusive gateway? The enumerative reading finds some textual support in the structure of the provision itself: the first subparagraph states the “no significant risk” standard as an apparently self-standing rule, and the second subparagraph opens with “the first subparagraph shall apply” where one of the conditions is met, which could be read as triggering a presumption rather than defining the only permissible path.57

Textual and systematic considerations support the conclusion that the list is closed: First, the legislative technique of enumerating specific derogation conditions following a general prohibition clause is a well-established mechanism in EU law for delimiting the scope of a derogation.58 Had the legislator intended the conditions to be illustrative, one would expect the use of qualifying language such as “in particular” or “for example” – language that is absent here.59 Second, Article 6(6) empowers the Commission to adopt delegated acts amending the conditions set out in Article 6(3), “adding new conditions to those laid down in paragraph 3, or modifying them.” The express conferral of a power to add conditions would be superfluous if the existing list were already open-ended; the provision implies that the Article 6(3) conditions can only be expanded through formal legislative amendment.60 Third, reading the conditions as a closed list is consistent with the principle of strict interpretation of derogations in EU law, as established in the ECJ’s jurisprudence. This ensures that the scope of the carve-out is legally determinate and that providers cannot invoke unenumerated grounds to escape high-risk classification.61 The consequence of this reading is that an AI system deployed within an Annex III area that does not satisfy any of the four enumerated conditions remains classified as high-risk, even if the provider considers the system’s specific function to pose no significant risk of harm. The question is not beyond doubt, and future guidance from the Commission or the ECJ may clarify whether the overarching condition has any independent operative function beyond its role as an interpretive constraint on the four conditions.

A second interpretive question is whether satisfying one of the second-subparagraph conditions deems the overarching “no significant risk of harm” threshold met (the automatic reading), or whether the overarching condition is a separate test the provider must also pass (the cumulative reading). The automatic reading reduces “no significant risk” to an interpretive constraint on the conditions. The cumulative reading is more protective, adding a case-by-case assessment on top of the condition-based classification, at the cost of legal certainty.

The automatic reading is, in my view, the better interpretation. Recital (53) describes the carve-out as covering “specific cases in which AI systems … do not lead to a significant risk of harm … because they do not materially influence the decision-making…” (my emphasis), indicating the significance/materiality threshold is integrated in the specific case conditions. The Article 6(7) power to delete conditions by delegated act “to maintain the protection of health, safety and fundamental rights” suggests the legislator already viewed the conditions as alone determining the boundary of the carve-out. The automatic reading is also more consistent with the Act’s ambition to provide clear, ex ante classification rules. The cumulative reading’s advantage is preserving the overarching condition as a safety net for edge cases the legislator did not foresee. The question would benefit from clarification.62

3.1.3 The “profiling of natural persons” carve-out from the carve-outs#s. 16

Article 6(3) third subparagraph introduces an exception from the exceptions: AI systems “shall always be considered to be high-risk where the AI system performs profiling of natural persons.”63 This contributes to the systematic interpretation of the derogation condition interpretations, and will play an important role in the operationalised criteria developed in Chapter 4.

3.2.1 Linguistic interpretation of “narrow”, “procedural”, and “task”#s. 16

Article 6(3)(a) provides that the carve-out applies where “the AI system is intended to perform a narrow procedural task.” The condition combines three lexical elements – “narrow”, “procedural”, and “task” – each of which contributes to delimiting the scope of the exception.

3.2.2 Systematic interpretation of “narrow procedural task”#s. 20

Article 14(4) requires that natural persons to whom human oversight is assigned are enabled “to properly monitor the operation of the high-risk AI system” and to “decide, in any particular situation, not to use the high-risk AI system or to otherwise disregard, override or reverse the output of the high-risk AI system.”68 A genuinely narrow procedural task does not produce outputs requiring such override mechanisms. The “procedural” character of a task should accordingly be assessed not only by its logical position in the workflow, but also by whether the output risks engaging automation bias in the subsequent human decision-maker. A task whose output is formally intermediate and procedural, but psychologically decisive – because it frames, narrows, or pre-structures the decision-maker’s cognitive field – is not procedural in the sense required by the provision: For instance, a case-management system preparing standardised documentation for cases judged to align with a simple “happy path” may, though the task being purely procedural, subconsciously tempt a time-pressed assessor to ignore signs that further investigation is needed, since investigating would forfeit the time-saving filing.69

GDPR Article 22, on automated individual decision-making, supplies a parallel systematic reference. Its distinction between automated processing that produces legal or similarly significant effects on a natural person and processing that does not maps onto Article 6(3)(a)’s procedural/substantive divide: language identification or date-sorting facilitates a later decision without itself producing such effects, whereas evaluative judgment of a person’s characteristics or situation engages both Article 22 and the high-risk regime.70

The ECJ’s judgment in Case C-634/21 (SCHUFA) provides an instructive illustration of this boundary in automated processing. The Court considered whether a credit reference agency’s calculation of a credit score amounted to a “decision based solely on automated processing” under GDPR Article 22, even though the formal lending decision was the bank’s. The Court rejected the agency’s argument that the score was merely preparatory, holding that where the probability value plays a “determining role” in the credit decision, describing it as preparatory would create a gap in the protection afforded by Article 22.71

The SCHUFA reasoning transposes to Article 6(3)(a) by analogy: the “procedural” character of a task should be assessed not by its formal label or its position in a workflow, but by its functional role in relation to the decision affecting the natural person. A formally intermediate task may be functionally decisive.72 Sorting job applications by date of receipt is procedural; sorting them by “relevance” to the position, even if the provider characterises this as a classification task, and the deployer is not the one making the hiring decisions, is substantive. See also the discussion below on the interpretation of “preparatory task” in Article 6(3)(d).

3.2.3 Teleological interpretation: the purpose of carving out procedural tasks within high-risk activities#s. 21

The teleological analysis of Article 6(3)(a) must be situated within the AI Act’s dual objectives: the protection of fundamental rights and the promotion of innovation through proportionate regulation.73 The purpose of carving out narrow procedural tasks is to avoid imposing the full weight of high-risk compliance on AI systems that, despite operating in a sensitive domain, perform functions that do not engage the interests the regulation seeks to protect.

Recital (53) articulates this purpose: there are “specific cases in which AI systems referred to in pre-defined areas specified in this Regulation do not lead to a significant risk of harm to the legal interests protected under those areas because they do not materially influence the decision-making or do not harm those interests substantially.” Narrow procedural tasks should accordingly be read as only those whose output does not bear on the legal interests that justified the inclusion of the area in Annex III.

At the same time, the purpose of the carve-out requires that it not be deprived of its effet utile. It is easy to construct scenarios in which even narrow procedural tasks could inadvertently cause harm: a duplication-removal algorithm wrongly drops a job applicant who shares a name and date of birth with another, or a date-parsing algorithm miscomputes the applicant’s years of experience because of inconsistent date formats. If the notion of “narrow procedural task” were read so strictly that any theoretical risk arising from such combinations of AI and human error sufficed to defeat it, no real-world AI system could satisfy the condition, the carve-out would exist on paper only, and the proportionality principle would be undermined. The ECJ has consistently held that derogations must be interpreted strictly, but not so as to deprive them of practical effect.74 Concretely, this means that the condition must be capable of covering genuinely low-risk procedural AI functions that are deployed within high-risk areas.

The Act’s dual purposes thus impose a two-sided constraint: Article 6(3)(a) must be read strictly enough to prevent substantive decision-support functions from being relabelled as procedural, yet broadly enough to permit genuinely low-risk procedural functions to escape full high-risk compliance.

3.3.1 Linguistic interpretation of “preparatory” and “to an assessment”#s. 22

Article 6(3)(d) provides that the carve-out applies where “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.” This condition is linguistically more complex than condition (a), and its interpretation requires attention to each of its constituent elements.

The term “task” must be interpreted as in condition (a) – it anchors the assessment to a specific delimited function performed by the AI system in its intended context of use.

The adjective “preparatory” denotes that this is a task “that prepares or serves to prepare for something following; preliminary, introductory.”75 In legal parlance, a preparatory act is one that precedes and contributes to a final determination without itself constituting that determination. The distinction between preparatory and definitive acts is a staple of EU administrative law.76 The distinction was central in the above-mentioned SCHUFA case, in which the Court found that classifying third party credit scoring as (only) a preparatory act to a subsequent credit decision would cause “a lacuna in legal protection”77 where the credit decision-maker would be unable to provide information on, and review the merits of, the automated credit scoring process.

The prepositional phrase “to an assessment” establishes that the preparatory task must be directed towards, and in service of, an “assessment.” This implies a temporal and functional sequence: first, the AI system performs a preparatory task; then, separately, an assessment takes place, using the output of the preparatory task. The use of the indefinite article “an” suggests the preparatory task must be tied to a specific assessment act, not a vague or diffuse process.

The term “assessment” itself is not defined in the AI Act, but its ordinary general meaning denotes an act of estimation or evaluation.78 In the context of the Annex III areas, “assessment” would refer to estimations and evaluations such as, but not limited to, a natural person’s state of mind, creditworthiness, insurability, suitability for employment, eligibility for social benefits, criminal risk, or medical condition, the state of critical infrastructure: water quality, highway traffic, power supply, the relevance of facts and laws in justice processes. The assessment task output would then in turn be input to the decision-making function that the high-risk regime seeks to regulate. Interestingly, unlike Article 6(3)(c), there is no specification that the assessment be a “human assessment”.79 An AI system could perform a preparatory task to an assessment by a (separate and different) high-risk AI system, and the former system would still qualify for the Article 6(3)(d) derogation condition.

The combined effect of these elements is that Article 6(3)(d) covers AI tasks whose outputs serve as inputs to functionally and temporally separate human- or AI-performed estimative or evaluative acts, without themselves constituting or pre-empting those acts.

3.3.2 Systematic interpretation of “preparatory task to an assessment”#s. 24

3.3.3 Teleological interpretation: the purpose and limits of excluding preparatory tasks from high-risk classification#s. 25

The purpose of Article 6(3)(d) parallels that of condition (a): to keep the high-risk classification from extending to AI systems whose function does not engage the interests the regulation seeks to protect. Its specific rationale is that many Annex III use cases involve multi-stage processes in which AI is deployed at the preparatory stage but not at the stage of evaluative judgment – for example, an AI database search for candidates matching formal recruitment criteria while a human interviewer performs the substantive suitability assessment, or AI-assisted retrieval and translation of case law while the judicial assessment remains the prerogative of the judge.

The purpose of the carve-out is thus to recognise that the high-risk regime is concerned with AI functions that bear on the substance of assessments and decisions affecting fundamental rights, and that functions which merely facilitate the administrative preconditions for such decisions – making information available, organising files, translating documents – do not require the same intensity of regulation. At the same time, the carve-out must not be allowed to become a vehicle for exempting decision-support systems that, in substance, perform or pre-empt the assessment. In any case the SCHUFA doctrine of GDPR Article 22, discussed in Section 3.2.2 above, provides an upper boundary by analogy: A task should not be considered genuinely preparatory in the context of Article 6(3)(d) if its output plays a “determining role” in the following assessment.86