High-Risk AI Classification Should Follow the Decision Chain
A response to the European Commission’s consultation on its draft Article 6 guidelines, examining human review, connected systems and the evidence behind classification in employment and platform work.
Maja Kurek, AIGP, CIPM, CIPP/E · RuleBridge Advisory
Submitted to the European Commission on 23 July 2026, in a personal capacity as an independent expert.
AI can influence access to work before a person confirms the result, and its outputs can affect decisions made later. A classification assessment needs to account for that intended data and decision chain, including connections between systems and the reuse of their signals.
The five comments summarised here asked the Commission to clarify how those connections should be assessed and what a provider relying on the Article 6(3) filter should record. The response focused on employment, workers’ management and access to self-employment within Section IV of the draft guidelines published on 19 May 2026.
What a sign-off shows.
The draft already recognises that human involvement does not change the intended purpose that brings a system within an Annex III use case. Human involvement may help establish whether an Article 6(3) condition is met, but adding a reviewer does not itself establish that condition.
Article 6(3) provides four specific, alternative conditions for an exception. The draft interprets them narrowly, in light of whether the system materially influences the outcome of a decision. An Annex III system that profiles natural persons cannot benefit from the exception.
The comment on paragraph 71 asks the guidelines to clarify that nominal sign-off does not, by itself, show that an output lacks material influence. The relevant intended conditions of use include access to the underlying record, sufficient time and competence for an independent assessment, real authority to disregard or reverse the output, and whether downstream steps treat the output as the default.
A recruiter may approve a ranked shortlist without reviewing the applications behind it; the approval alone does not establish an independent assessment.
A system split between providers.
Paragraph 75 addresses configurations in which several AI systems form part of a more complex system and their combined intended purpose or joint outputs materially influence an individual decision. Where the configuration serves an Annex III use case, the draft assesses the combined system as a whole. Paragraph 76 preserves the possibility of an Article 6(3) exception for genuinely separable procedural or preparatory functions that do not structure or feed outputs materially influencing that decision.
The submitted worked example applies this distinction to employment. Provider A structures work records and customer feedback. Provider B converts those indicators into a score or priority band. A platform uses the joint output to rank people or allocate assignments, and an operator confirms the result. In this configuration, the components together perform a substantive assessment and materially influence a listed employment decision. The response proposes treating the combined system as high-risk.
The accompanying boundary case concerns a genuinely separable system that only normalises previously verified information into predefined fields. It does not rank, hide, prioritise, recommend or infer. Its output is not intended to be reused by another component to affect assignments, pay, account status or termination, and the decision-maker reviews the full underlying record. Such a system may qualify for the Article 6(3) filter.
Access to work after onboarding.
Annex III, point 4(a) covers recruitment and selection. Point 4(b) covers specified decisions affecting work-related terms, promotion or termination, task allocation based on individual behaviour or characteristics, or the monitoring and evaluation of performance and behaviour in work-related relationships.
The Commission’s employment examples illustrate both categories. A platform that ranks self-employed service providers and determines which providers consumers see falls under point 4(a). A scheduler that allocates shifts to existing workers using behavioural indicators, including acceptance and decline rates, falls under point 4(b).
Recurring platform matching can involve either function or both. A person may be onboarded once while each assignment may involve a new contract. The comments on paragraphs 249 and 256 ask for the distinction to turn on whether the AI selects a person for access to a new contract or assignment, manages tasks or terms within an existing relationship, or performs both functions. Whether onboarding has occurred should not, by itself, decide the question.
Signals reused downstream.
On paragraph 278, the response asks the Commission to clarify that a function is not merely supportive where the intended configuration retains or reuses its signals to rank a person, affect future task access, pay or account status, or feed another component doing so. Classification should follow the intended data and decision chain, including uses beyond the immediate user-facing function.
The Commission’s employment examples include a delivery tool that suggests areas of higher demand. It falls outside point 4(b) because its suggestions are optional, declining them carries no penalty, acceptance and refusal are not recorded, and future task access is unaffected.
If the intended configuration began recording refusals and using them to allocate future work, the basis for treating the function as outside point 4(b) would need to be reassessed. The interface could remain unchanged while the role of the information in work decisions changed.
A record that says when to look again.
Under Article 6(4), a provider that considers an Annex III system not to be high-risk must document its assessment before placing the system on the market or putting it into service. The provider is also subject to registration under Article 49(2) and must supply the assessment documentation to national competent authorities on request.
Paragraph 115 of the draft already calls for an account of the intended purpose, why the system falls within Article 6(2), which Article 6(3) condition or conditions apply and why, and why the system does not perform profiling.
The comment asks for a record that also identifies the decision stage, the system’s inputs and outputs, the action expected from the user, dependencies on other components, foreseeable downstream use, and the evidence supporting the conclusion that the output does not materially influence the relevant Annex III outcome.
The response also asks the record to identify changes in configuration, data flows or intended use that trigger reassessment.
The normalisation example concerns a possible Article 6(3) exception; the courier example concerns whether point 4(b) applies at all. Both conclusions depend on the configuration described. If the normalised information later feeds a ranking, or recorded refusals are used to allocate future work, the original assessment needs to be revisited. Naming those changes helps the provider and competent authorities recognise when reassessment is needed.