AI in HR: The Seven-Step AI Act Test under Annex III(4)

Few business functions are as saturated with AI features as HR: applicant tracking, shift planning, time recording, performance reviews. And few are named as explicitly high-risk in the AI Act. Classifying them properly does not take gut feeling but a fixed sequence of questions. This article provides it, with the legal text, a case matrix and an evidence list.

In short

Under Annex III(4) of the AI Act, HR AI is high-risk when it is intended to screen applications, evaluate people, allocate tasks based on behaviour or personal traits, or monitor performance. First check whether it is an AI system at all, then whether an Article 6(3) exception applies. Profiling is always high-risk. High-risk duties apply from 2 December 2027.

Key points

  • Annex III(4) covers two groups: recruitment and selection (point (a)) and decisions in the employment relationship, task allocation based on behaviour or personal traits, and performance monitoring (point (b)).
  • The first question is not "high-risk?" but "an AI system at all?". According to the Commission guidelines, optimisers with fixed, predefined rules generally fall outside Article 3(1).
  • Article 6(3) exempts Annex III systems that do not materially influence the outcome of decision-making. Profiling always rules out the exception.
  • Whoever relies on the exception is the provider and must document the assessment (Article 6(4)) and register the system (Article 49(2)).
  • Employers are deployers. From 2 December 2027 the duties of Article 26 apply to them, including logs kept for at least six months (paragraph 6) and informing workers' representatives and affected workers before putting the system into service (paragraph 7).
  • Already in force today: the ban on emotion recognition in the workplace (Article 5(1)(f)), AI literacy (Article 4) and the transparency duty for interaction with AI (Article 50(1)).

Why HR is the most sensitive use case

The AI Act defines high-risk AI in two ways: products under Annex I and use cases under Annex III. In HR it is almost always Annex III(4), "Employment, workers' management and access to self-employment". The wording is broader than many product descriptions suggest.

Under point (a), AI systems are high-risk if they are intended to be used for the recruitment or selection of natural persons, in particular to place targeted job advertisements, to analyse and filter job applications, and to evaluate candidates. Under point (b), the same applies to systems intended to make decisions affecting terms of work-related relationships, promotion or termination, to allocate tasks based on individual behaviour or personal traits or characteristics, or to monitor and evaluate the performance and behaviour of persons in such relationships.

This brings not only recruiting tools within scope but potentially also shift planning, workforce dispatching and performance analytics. The key word is "intended": what counts is the intended purpose, i.e. what the provider designates the system for in the instructions for use, promotional material and technical documentation (Article 3(12)).

Deadlines under Article 113 as amended by Regulation (EU) 2026/1744: prohibitions and AI literacy apply since 2 February 2025, the Article 50 transparency duty since 2 August 2026, and high-risk requirements for Annex III systems from 2 December 2027. Anyone introducing an HR system today is already planning it for the high-risk phase.

The seven-step test

The order matters. Each step can end the test, and skipping a step leads to the wrong role or the wrong list of duties.

StepQuestionProvisionConsequence
1Is it an AI system at all?Article 3(1); guidelines paragraphs 42 to 48No: the AI Act does not apply, the test ends
2Is the practice prohibited?Article 5(1)(f)Yes: use is not permitted
3Does the purpose fall under Annex III(4)?Annex III(4)(a) and (b)No: not high-risk, Articles 4 and 50 remain
4Does an exception apply?Article 6(3), first and second subparagraphsYes: continue with step 5; provider documents and registers
5Is there profiling?Article 6(3), third subparagraphYes: always high-risk
6Who is provider, who is deployer?Article 3(3) and (4); Article 25Assign duties to each role
7Which duties and evidence?Articles 26, 86, 99, 113Set up the evidence file

The seven steps in detail

Step 1: Is it an AI system at all?

Article 3(1) requires a machine-based system that "infers, from the input it receives, how to generate outputs". The Commission guidelines on the definition of an AI system (German version linked) expressly exclude three groups in paragraphs 42 to 48: systems for improving mathematical optimisation, basic data processing systems that operate on fixed, human-programmed rules, and classical heuristics.

For shift planning this means: an optimiser that allocates shifts based on fixed rules for availability, maximum hours, rest periods and qualifications is generally not an AI system. A language model that translates free-text instructions into planning actions is one. Many products combine both, so each function is assessed separately.

Step 2: Is the practice prohibited?

Since 2 February 2025, placing on the market, putting into service and using AI systems to infer emotions of a natural person in the workplace is prohibited, unless the use is intended for medical or safety reasons (Article 5(1)(f)). This affects, for example, video interview tools that score enthusiasm or nervousness and call-centre analytics that measure employee mood. Fines reach up to EUR 35 million or 7% of worldwide annual turnover (Article 99(3)).

Step 3: Does the purpose fall under Annex III(4)?

Here the intended purpose is checked against points (a) and (b). Typical hits: ranking or filtering applications, targeted job advertising, performance and behaviour analytics, allocating shifts or tasks based on behaviour or personal traits. Typically no hit: general knowledge chatbots for employees, or drafting job adverts without audience targeting.

Step 4: Does the Article 6(3) exception apply?

An Annex III system is not considered 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. In addition, one of four conditions must be met: (a) a narrow procedural task, (b) improving the result of a previously completed human activity, (c) detecting decision-making patterns without replacing human assessment, (d) a preparatory task to an assessment.

A confirmation click alone does not meet this. The test is whether the output actually shapes the human decision. Whoever relies on the exception documents the assessment before placing on the market (Article 6(4)) and registers the system in the EU database (Article 49(2)). These duties fall on the provider, not the employer.

Step 5: Is there profiling?

The third subparagraph of Article 6(3) is the hard limit: an Annex III system that performs profiling of natural persons is always high-risk. The term follows Article 4(4) GDPR (Article 3(52) AI Act): any automated processing that evaluates personal aspects such as work performance, reliability or behaviour. Performance, punctuality or absence scores are therefore clearly profiling. Whether merely taking self-reported availability into account is profiling remains open; the classification should address this point explicitly.

Step 6: Who is provider, who is deployer?

The software vendor is the provider; the employer using the system under its own authority is the deployer. Roles can flip: under Article 25(1)(c), anyone who modifies the intended purpose of a system not classified as high-risk so that it becomes high-risk becomes a provider. One example is an employer feeding its own performance metrics into the allocation criteria. The original provider then no longer counts as the provider of that system (Article 25(2)).

Step 7: Duties and evidence

Only now is it clear which list of duties applies. The overview in the next section sorts them by date and role.

Case matrix: ten typical HR applications

The matrix shows the usual classification for a typical intended purpose. It does not replace a case-by-case assessment, but it shows at which step the decision is made.

ApplicationClassificationDecisive step and provision
Emotion analysis in video interviewsProhibitedStep 2: Article 5(1)(f)
Ranking or filtering applicationsHigh-riskSteps 3 and 5: Annex III(4)(a), usually profiling
Targeted job advertisingHigh-riskStep 3: Annex III(4)(a) ("targeted job advertisements")
Performance or behaviour scoresHigh-riskSteps 3 and 5: Annex III(4)(b), profiling
Shift allocation by performance, punctuality or sick daysHigh-riskStep 5: profiling under Article 6(3), third subparagraph
Shift optimiser with fixed rules (availability, rest periods, qualifications)Generally not an AI systemStep 1: Article 3(1), guidelines paragraphs 42 to 48
Voice assistant executes the planner's instructions, human confirmsException possibleStep 4: Article 6(3)(a); provider documents and registers; Article 50(1)
Summarising application documents without evaluationException possibleStep 4: Article 6(3)(d), provided there is no profiling
Knowledge chatbot for employees (leave, policies)No Annex III purposeStep 3; Articles 50(1) and 4 apply
Writing aid for job adverts without audience targetingNo Annex III purposeStep 3; Article 4 applies

Duties by role and date

High-risk duties start on 2 December 2027. Three blocks of duties, however, already apply today and affect every company using AI in HR.

FromDutyProvisionWho
2 Feb 2025No emotion recognition in the workplaceArticle 5(1)(f)Everyone
2 Feb 2025AI literacy of staff operating AIArticle 4Provider and employer
2 Aug 2026Make clear that people are interacting with AIArticle 50(1)Provider
2 Dec 2027Risk management, documentation, logging, human oversight, conformity assessment, registrationArticles 9 to 17, 43, 49Provider
2 Dec 2027Use per instructions, oversight by competent persons, relevant input data, monitoringArticle 26(1), (2), (4), (5)Employer
2 Dec 2027Keep automatically generated logs for at least six monthsArticle 26(6)Employer
2 Dec 2027Inform workers' representatives and affected workers before putting into serviceArticle 26(7)Employer
2 Dec 2027Inform affected persons about the useArticle 26(11)Employer

Explanation, impact assessments and fines

Right to explanation and data protection impact assessment

Anyone affected by a decision based on the output of an Annex III high-risk system that significantly affects them can request from the deployer a clear and meaningful explanation of the role of the AI system in the decision-making procedure (Article 86(1)). For the data protection impact assessment under Article 35 GDPR, the deployer uses the provider's information under Article 13 (Article 26(9)).

A fundamental rights impact assessment under Article 27 only concerns bodies governed by public law and private entities providing public services, not the ordinary private employer.

Fines

Infringements of the deployer duties in Article 26 and the transparency duties in Article 50 can be fined up to EUR 15 million or 3% of worldwide annual turnover (Article 99(4)(e) and (g)). For SMEs and start-ups the lower amount applies (Article 99(6)).

Works council and data protection run in parallel

The AI Act does not replace German labour law, it comes on top. Four provisions of the German Works Constitution Act (BetrVG) are central for AI in HR, three of them name AI expressly:

  • Section 87(1) nos. 2 and 6 BetrVG: co-determination on working-time arrangements and on technical devices for monitoring behaviour or performance. Under settled case law, the objective capability to monitor is sufficient.
  • Section 90(1) no. 3 BetrVG: information on planned work processes and workflows "including the use of artificial intelligence".
  • Section 80(3) BetrVG: where the works council has to assess the introduction or use of AI, calling in an expert is deemed necessary.
  • Section 95(2a) BetrVG: selection guidelines remain subject to co-determination even where AI is used to draw them up.

Data protection: Article 22 GDPR and legal basis

Article 22 GDPR prohibits solely automated decisions with legal or similarly significant effects. Following the CJEU judgment C‑634/21 (SCHUFA), a score that decisively determines a third party's decision can already be covered. Since the CJEU judgment C‑34/21, Section 26 BDSG is disputed as a legal basis for employee data; Article 6(1)(b) GDPR and works agreements under Article 88 GDPR are robust alternatives.

In practice, informing workers' representatives under Article 26(7) AI Act and informing the works council under Section 90 BetrVG can be done in one step once the classification from steps 1 to 7 is available.

Practical example: an AI assistant in shift planning

How the steps interlock in practice is shown by FixPlaner's article on classifying AI-supported shift planning under the EU AI Act (in German). FixPlaner describes the actual flow of its assistant: a language model structures the planner's instruction, fixed rules check availability, qualifications and working time, nothing is saved until explicitly confirmed, and management publishes the roster. Alev-B reviewed the article for technical accuracy.

Run through the seven steps, the picture is nuanced: the rule-based optimiser already fails step 1. The voice assistant is an AI system, touches Annex III(4)(b) through shift allocation and can qualify as a narrow procedural task under Article 6(3)(a) as long as it does no profiling. The provider must then document and register the classification, and the Article 50(1) transparency duty already applies. As soon as an employer adds its own performance or absence metrics as a criterion, the classification flips under steps 5 and 6.

Six mistakes that keep coming up

  1. "A human still clicks." A confirmation click is not effective oversight if the output shapes the decision (Article 6(3), Article 14(4)(b) on automation bias).
  2. "High-risk only applies in 2027." Prohibitions, AI literacy and Article 50 already apply; anything introduced today has to be planned for 2027.
  3. "We are only the deployer." Own criteria or a changed purpose can turn the employer into a provider under Article 25.
  4. "The exception is obvious." Without a documented assessment under Article 6(4) and registration under Article 49(2), it does not hold.
  5. "The works council comes after go-live." Article 26(7) requires information before putting into service, Section 90 BetrVG timely information already at the planning stage.
  6. "We delete logs after 30 days." Article 26(6) requires at least six months unless data protection law provides otherwise. Both belong in one retention concept.

The evidence file: ten documents that carry any review

Whatever the classification result, a fixed file per HR system pays off. It answers the works council, data protection and supervisory authorities with the same material.

Create the classification note and role matrix directly in the AI inventory. The free AI Act check shows in a few minutes where your company stands on the rest.

  1. Purpose statement: what the system is used for and what it is expressly not used for.
  2. Classification note covering steps 1 to 7, with date and sources.
  3. Role matrix: provider, deployer, own modifications (Article 25).
  4. AI literacy training records (Article 4).
  5. Oversight concept: who reviews, with what authority and what latitude (Article 26(2)).
  6. Logging and deletion concept (Article 26(6), GDPR).
  7. Information to workers' representatives and workers (Article 26(7), Section 90 BetrVG).
  8. Data protection impact assessment using the provider's information (Article 26(9)).
  9. Information to affected persons (Article 26(11)) and transparency notice (Article 50(1)).
  10. Process for explanation requests and objections (Article 86, Article 22 GDPR).

Sources

Regulation (EU) 2024/1689, consolidated version of 27 July 2026, in particular Articles 3, 4, 5, 6, 25, 26, 49, 50, 86, 99, 113 and Annex III(4).

Commission guidelines on the definition of an AI system (German version), paragraphs 42 to 48.

German Works Constitution Act (BetrVG), Sections 80, 87, 90, 95 as currently in force.

Deadlines and amendments by Regulation (EU) 2026/1744: Digital Omnibus 2026/1744 – the new AI Act deadlines. Overview: EU AI Act compliance guide.

As of 19 November 2026. This article does not constitute legal advice.

Key Takeaways

  • First check whether it is an AI system at all: rule-based optimisers generally fall outside Article 3(1).
  • Annex III(4) covers recruitment, decisions in the employment relationship, task allocation based on behaviour or traits, and performance monitoring.
  • Article 6(3) can provide relief, but never with profiling, and only with documentation and registration by the provider.
  • From 2 December 2027 employers carry the deployer duties of Article 26; prohibitions, AI literacy and Article 50 already apply today.
  • Labour law and the GDPR run in parallel; one evidence file per HR system serves all three reviewers.

Frequently Asked Questions

No. A system is high-risk if its intended purpose falls under Annex III(4) and no Article 6(3) exception applies. Systems with fixed rules are often not AI systems at all, and knowledge chatbots usually fall outside Annex III. Profiling, by contrast, always makes an Annex III system high-risk.

Under Article 113 as amended by Regulation (EU) 2026/1744, from 2 December 2027 for Annex III systems. The ban on emotion recognition and AI literacy apply since 2 February 2025, the Article 50 transparency duty since 2 August 2026.

Not automatically. What matters is whether the output materially influences the outcome. A confirmation without real latitude or without demonstrable deviations is not effective oversight and does not support the Article 6(3) exception.

The provider. It documents the assessment before placing on the market (Article 6(4)) and registers the system in the EU database (Article 49(2)). The employer should obtain these documents and add them to its own evidence file.

As a rule, yes. Section 90(1) no. 3 BetrVG requires information on the use of AI at the planning stage, Section 87(1) no. 6 co-determination for systems capable of monitoring behaviour or performance. For high-risk systems, the information duty under Article 26(7) AI Act is added from 2 December 2027.

At least six months, insofar as they are under the deployer's control and Union or national law, in particular data protection law, does not provide otherwise (Article 26(6)).

EU AI ActAnhang III Nr. 4KI im PersonalbereichArtikel 6 Absatz 3Artikel 26BetriebsratHochrisiko-KI

Ready for Your Assessment?

Use our interactive templates to measure your IT organization's maturity — with automatic scores, AI-powered recommendations, and professional PDF reports.