Common Annex III misclassifications in HR AI — and how to correct them

By PowerQuant | Updated June 2026 | Reading time: ~9 minutes

The classification framework: Annex III point 4 and Article 6

Annex III of Regulation (EU) 2024/1689 lists the categories of AI system that are presumed high-risk. For HR and employment contexts, point 4 is the relevant entry:

“Employment, workers' management and access to self-employment: (a) AI systems intended to be used for recruitment or selection of natural persons, in particular to place targeted job advertisements, to analyse and filter applications, and to evaluate candidates in the course of interviews or tests; (b) AI systems intended to be used to make decisions affecting the terms of work-related relationships, the promotion or termination of work-related contractual relationships, to allocate tasks based on individual behaviour or personal traits or characteristics, or to monitor and evaluate performance and behaviour of persons in such relationships.”

However, Article 6(2) provides a derogation: a system listed in Annex III is not considered high-risk if it does not pose a significant risk of harm to health, safety, or fundamental rights, in particular by not materially influencing the outcome of decision-making. This exception applies when one or more of four conditions is met:

• (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;
• (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.

Article 6(3) further provides that a provider who believes their Annex III system is not high-risk must document this assessment and notify the relevant market-surveillance authority before placing the system on the market. The documentation burden falls on the party claiming the exception — not on anyone challenging it.

Part 1: Under-classification errors (treating a high-risk system as not high-risk)

“Our system only makes recommendations — a human decides”

This is the most common misclassification error. The argument runs: the AI produces a score or ranking, but the recruiter or manager makes the final call; therefore the system does not “decide” anything and Article 6(2)(a) applies.

The Act does not require the AI to make a binary decision. Annex III point 4 covers systems that evaluate candidates and systems that monitor and evaluate performance — both of which produce inputs to human decisions. The Article 6(2) exception requires that the system does not materially influence the outcome. If a recruiter sees only the top 20 candidates surfaced by an AI and never reviews the other 500 applications, the AI has materially influenced the outcome regardless of who clicked “reject”.

Correction: The test is not whether a human has the last click. It is whether the AI output materially shapes the decision set. If the AI determines who a human ever considers, it materially influences the outcome.

“We only use it for internal efficiency — not for decisions about individuals”

Workforce-planning tools that predict attrition, flag flight-risk employees, or recommend allocation of assignments based on behavioural inference are often framed as “operational analytics” rather than systems that affect individuals. The framing does not change the legal analysis. Annex III point 4(b) explicitly covers AI systems used to allocate tasks based on individual behaviour or personal traits.

Correction: If the output of the system — whether surfaced directly or fed into another process — affects what work an individual receives, whether they are promoted, or whether their contract is at risk, it falls within point 4(b).

“It uses a foundation model (GPAI), so the GPAI rules apply, not Annex III”

General-purpose AI model obligations (Articles 53–55) sit with the model provider. A deployer who builds an HR application on top of a GPAI model is not using a GPAI system for classification purposes — they are deploying an AI system that uses a GPAI component. If the combined system's intended purpose falls within Annex III point 4, the high-risk classification applies to the full system, not just the underlying model.

Correction: Classification is based on the system's intended purpose, not its underlying architecture. A GPT-4-powered candidate scoring tool is an Annex III high-risk system if it evaluates candidates.

“Our vendor says it's not high-risk”

Vendors have an interest in keeping their compliance burden low. A vendor assessment that a product is “not high-risk” under Article 6(2) is self-serving and is not binding on the deployer or on any market-surveillance authority. Article 80 of the Act gives authorities the power to investigate where a system has been classified as non-high-risk and imposes fines where the classification is found to have been used to circumvent Chapter III requirements.

Correction: Request the vendor's Article 6(3) documented assessment — not a sales assertion. Read it. If the reasoning does not address “materially influencing the outcome of decision-making”, treat it as unsubstantiated.

“We are too small for the obligations to apply”

There is a size-threshold exception for some SME providers on certain requirements, but it does not exempt deployers from their Article 26 obligations. The deployer obligations in Articles 26 and 27 apply regardless of the deployer's size. The Digital Omnibus provisional agreement of 7 May 2026 proposes to raise the SME threshold to 750 employees (from 250) for certain provider-side simplifications — but deployer obligations are not removed by that adjustment.

Correction: Check whether a size exception applies to your specific obligation, not to the AI Act generally. For deployers, Article 26 does not contain a size carve-out.

Part 2: Over-classification errors (treating a non-high-risk system as high-risk)

“Any AI that touches HR data must be high-risk”

Processing personal data does not make a system high-risk. GDPR and the AI Act are distinct legal instruments. An HR analytics dashboard that shows aggregated attrition rates, or a leave-management tool with a predictive absence flag, may process personal data but not trigger Annex III point 4 — if it does not materially influence employment decisions about identifiable individuals.

Correction: Ask whether the system produces an output — a score, recommendation, ranking, or flag — that influences a decision about a specific individual in a point 4(a) or 4(b) context. Aggregated reporting tools that do not reach individual employment decisions are typically not within Annex III point 4.

“Our calendar scheduling tool uses AI so it must be high-risk”

AI-assisted scheduling tools that propose meeting times or interview slots based on calendar availability are a canonical Article 6(2)(a) or (d) case — they perform a narrow procedural or preparatory task. They do not evaluate a candidate or influence an employment decision. Treating them as high-risk diverts compliance resources from systems that genuinely are.

Correction: Apply the Article 6(2) conditions methodically and document the reasoning. A scheduling tool that arranges an interview is not evaluating the candidate — the interview is. Classify clearly and document the basis.

Over-classifying GPAI-assisted drafting tools

A recruiting coordinator using an LLM to draft job descriptions or offer letters, where a human reviews and approves before anything is sent, is a legitimate Article 6(2)(b) candidate — the AI improves the result of a human-led activity; it does not filter applicants or evaluate candidates. Classify it as limited-risk (transparency obligations under Article 50(4) may still apply if AI-generated content reaches the public), but do not load it with Annex III high-risk evidence requirements.

Correction: Separate the AI drafting tool from the decision it supports. If a human independently reviews the output before it reaches a candidate or affects an employment decision, the drafting tool may not be Annex III high-risk. Document the workflow clearly — the human-review step is the pivot.

A classification workflow for HR AI systems

1. Step 1 — Define the intended purpose in plain language: what decision or output does this system support, for whom, and in what context?
2. Step 2 — Map to Annex III: does the intended purpose fall within point 4(a) or 4(b)? Be precise about whether the output reaches an individual-level decision, not just an aggregate one.
3. Step 3 — Test Article 6(2): is the system performing only a narrow procedural task, improving a completed human activity, detecting deviations without replacing human review, or performing a preparatory task? If yes, document why and classify as non-high-risk.
4. Step 4 — Check Article 6(2) materiality condition: even if a 6(2) condition superficially applies, does the system materially influence the outcome? A “preparatory” tool that in practice determines which candidates ever reach a human reviewer fails this gate.
5. Step 5 — Document and store: the classification reasoning for every system — high-risk, non-high-risk, and “under review” — should be recorded in your AI inventory. “Under review” entries need a resolution date.

What the Commission guidelines say

The European Commission published draft guidelines on the classification of high-risk AI systems (available at digital-strategy.ec.europa.eu). These guidelines provide worked examples and clarify how the Article 6(2) derogation interacts with Annex III point 4. They confirm that the key test is whether the system materially influences the decision-making outcome as seen from the perspective of the affected person — not from the technical architecture or the vendor's product marketing.

Sources

• Regulation (EU) 2024/1689, Articles 3, 6, 26, 80 and Annex III — eur-lex.europa.eu/eli/reg/2024/1689/oj
• European Commission draft guidelines on classification of high-risk AI systems — digital-strategy.ec.europa.eu
• AI Act Service Desk — Article 6 classification rules — ai-act-service-desk.ec.europa.eu
• AI Act Service Desk — Annex III — ai-act-service-desk.ec.europa.eu