For whom
HR technology providers, ATS/matching/screening providers, staffing companies, recruitment organisations, employers and teams using AI in workforce processes.
Annex III point 4
AI in HR, employment and worker management
A practical hub for AI systems around recruitment, selection, employment relationships, promotion, termination, task allocation, monitoring and performance evaluation.
Annex III point 4 of the AI Act is not limited to recruitment. It has two risk routes: access to work through recruitment and selection, and decisions within employment relationships through worker management, task allocation, monitoring and evaluation. In both routes, AI systems affect equal treatment, privacy, transparency and human oversight.
Updated: May 202620 min read
Two risk routes
Route 4a covers recruitment and selection. Route 4b covers employment conditions, promotion, termination, task allocation, monitoring and performance evaluation.
Concrete output
Use-case register, risk classification, affected-person communication, human oversight playbook, training records and question-answer documentation.
The two routes within Annex III point 4
A good assessment starts by determining whether the AI system concerns access to work or decisions within an existing employment relationship. The requirements can be serious in both routes, but the facts, affected persons and evidence items differ.
Annex III point 4a
Recruitment, selection and access to work
AI systems for targeted job advertisements, analysing and filtering applications and evaluating candidates.
Examples: Job ad targeting, CV parsing, matching, shortlisting, screening questions, assessments and interview support.
Open route 4aAnnex III point 4b
Worker management and employment relationships
AI systems for decisions on employment conditions, promotion, termination, task allocation, monitoring or performance and behaviour evaluation.
Examples: Performance scoring, productivity monitoring, scheduling with individual assessment, task allocation, promotion or attrition models.
Open route 4bDraft guidelines 19 May 2026
High-risk guidelines applied to Annex III point 4
On 19 May 2026, the European Commission published draft guidelines on the classification of high-risk AI systems. They matter for this hub because they explain how Article 6 and Annex III should be read for concrete use cases.
The guidelines are draft guidelines and the examples are not exhaustive. Use them as classification support, not as a substitute for the AI Act itself or for your own analysis of purpose, context and decision impact.
1. Start with intended use
Classification is not only about the technology, but mainly about what the system is intended to be used for. A generic AI feature may fall outside Annex III, while the same feature can become high-risk in selection, evaluation or monitoring.
2. Record the route explicitly
State whether the use case falls under 4a, under 4b, or outside Annex III point 4. This prevents recruitment/selection and worker management from being assessed as one undifferentiated category.
3. Substantiate borderline cases
For assistants, dashboards or workflow tools, the question is often how much influence the output has on a decision about a person. Record whether the output is advisory, ranking, filtering or decision-making.
What does this mean for 4a and 4b?
Route 4a
Recruitment and selection
Focus on systems that target job ads, analyse or filter applications, rank candidates or assess suitability. Evidence should show how access to work is influenced and where human oversight sits.
View route 4aRoute 4b
Worker management and employment relationships
Focus on systems that influence employment conditions, promotion, termination, task allocation, monitoring or performance and behaviour evaluation. Evidence should show what consequences the output may have for employees or self-employed persons.
View route 4bWhich HR and work AI is in scope?
Not every HR system is automatically high-risk. The question is whether the system is intended to influence access to work, evaluate candidates or support decisions within employment relationships.
| Route | Use case | Why risky | Evidence needed |
|---|---|---|---|
| 4a | Targeted job advertising | Determines which groups do or do not see vacancies. | Targeting criteria, channel choices, bias check and candidate transparency note. |
| 4a | CV parsing, matching and ranking | Influences who is considered suitable and who moves forward. | System purpose, input data, ranking logic, human review and override record. |
| 4a | Screening questions and knockout flows | May automatically exclude or discourage candidates. | Decision rules, lawful basis, human reassessment and communication. |
| 4a | Assessment or interview AI | Evaluates traits, answers or candidate suitability. | Validation, bias analysis, explainability, retention periods and oversight process. |
| 4b | Promotion, evaluation or termination | May affect employment conditions, career progression, contract termination or improvement plans. | Decision influence, evaluation criteria, human reassessment, objection route and logging. |
| 4b | Task allocation based on behaviour or traits | May affect work distribution, opportunities, pressure or compensation based on individual profiles. | Purpose limitation, characteristics used, proportionality, oversight and escalation process. |
| 4b | Monitoring and performance evaluation | Affects privacy, autonomy, evaluation and potential disciplinary consequences. | Scope definition, worker representation, data minimisation and escalation process. |
Classification check
Does this HR or work AI fall under Annex III point 4?
Answer four questions for an initial direction. The result is not a legal opinion, but helps distinguish route 4a, route 4b, borderline cases and evidence items.
What is the AI system intended to do?
What role does the output play in the process?
Who is affected by the output?
Where does human oversight sit?
The HR and work AI evidence stack
The strongest approach is neither a standalone training nor a generic policy. Build a file per recruitment or workforce process so legal, compliance, product, HR, recruitment and management can rely on the same facts.
1. AI use-case register
Record purpose, owner, supplier, data flows, roles, target group and decision impact for every tool.
2. Provider/deployer role map
Determine who is provider, deployer, importer, distributor or product integrator. Providers and users usually carry different evidence burdens.
3. AI Act classification note
Explain whether the use case is prohibited, high-risk, limited-risk or low-risk. Make uncertainty explicit.
4. GDPR, data and bias check
Connect AI Act risk with GDPR lawful basis, data minimisation, retention, special categories and bias monitoring.
5. Human oversight playbook
Describe when a recruiter, hiring manager, line manager or HR professional must review, override, escalate and record why they deviated from AI output.
6. Transparency for affected persons
Create clear texts for candidates, employees and self-employed persons explaining AI use, purpose and human oversight in plain language.
7. AI literacy and training records
Record what each role must understand: recruiter, hiring manager, product team, legal, compliance and user support.
Tools, templates and scans for HR and work AI
Use these resources as the starting point. Most organisations do not need more theory; they need a clear view of which evidence items are missing.
Whitepaper
HR-AI Evidence & Readiness
Read the explanation layer above the template pack: route 4a/4b, evidence stack and 30-day approach for HR-AI.
Open resourceCheck
Annex III point 4 classification check
Get an initial indication whether an HR or work-AI use case points to route 4a, route 4b, a borderline case or outside scope.
Open resourceTemplate pack
HR-AI Evidence Pack
Download the editable pack for use-case register, Annex III classification, human oversight, transparency, bias/data and training records.
Open resourceTool
AI Act risk classification
Determine whether an HR or work-related use case moves towards high-risk, limited-risk or low-risk.
Open resourceFRIA/DPIA
FRIA generator
Translate fundamental rights risks into a structured impact assessment.
Open resourceTemplate
AI register template
Record HR tools, suppliers, purposes, owners and risk status.
Open resourceTemplate
AI transparency notice
Make communication concrete for candidates, employees and other affected persons.
Open resourceTraining
AI literacy test
Determine which baseline knowledge is relevant for people working with AI systems.
Open resourceGuide
DPIA vs FRIA
See when privacy impact and fundamental rights impact intersect in HR and work AI.
Open resourceNext route
Where should the next step live?
Responsible AI Platform remains the knowledge and source layer. For execution, implementation and training, this hub points to the right specialised environment.
Embed AI
Route 4a: recruitment and selection
For classification, gap analysis and evidence around matching, screening, job ad targeting and candidate evaluation.
Embed AI
Route 4b: worker management and monitoring
For AI around task allocation, monitoring, performance, promotion, termination and other employment decisions.
LearnWize
AI literacy and training evidence
For role-based training, progress, certificates and Article 4 evidence for HR, recruiters, managers and product teams.
Priority by audience
Not every organisation has the same evidence problem. Providers need to explain functionality and limitations, recruitment organisations need controlled selection workflows and employers need explainable workforce processes.
HR technology providers
Document the classification, intended use, limitations, human oversight and user information for each AI functionality.
Recruitment operators
Bring existing tooling, recruiter workflows and candidate communication under control without freezing the whole operation.
Large employers
Start with the processes with the highest volume and reputational risk: matching, screening, task allocation, monitoring and performance evaluation.
Read more on Responsible AI Platform
This hub is the starting point. The articles and tools below cover the separate components: classification, FRIA/DPIA, AI literacy and transparency for affected persons.
Route 4a: recruitment and selection
Deep dive for AI in job ad targeting, matching, screening, ranking, assessments and candidate evaluation.
Route 4b: worker management and monitoring
Deep dive for AI in task allocation, monitoring, performance, promotion, termination and employment relationships.
HR & employment sector page
Broad overview of AI Act risks in HR, employment relations and recruitment.
AI literacy in recruitment
Why recruiters and hiring managers need role-based AI knowledge.
AI Act & HR recruitment
The practical impact of the AI Act on selection, matching and human oversight.
DPIA vs FRIA
When privacy impact and fundamental rights impact need to be assessed together.
FAQ about AI in HR and work
Is all AI in HR automatically high-risk?
No. Context and decision impact matter. A tool that edits text differs from a system that filters candidates, ranks employees, allocates tasks or evaluates performance. Once AI influences access to work or employment decisions, you should seriously assess the high-risk route.
What is the difference between compliance and evidence?
Compliance means meeting obligations. Evidence is the proof that you can demonstrate this in practice: registers, classifications, logs, work instructions, training records, transparency texts and decision procedures.
Must a recruiter always manually check AI output?
For systems with serious decision impact, human oversight must be meaningful. The human must understand the output, be able to intervene, deviate and record the reason for that deviation. A tick box after the fact is usually not enough.
Where do you start if you use multiple recruitment tools?
Start with a use-case register. Record whether each tool uses AI for writing, searching, matching, screening, ranking, assessing, task allocation, monitoring or communicating. Then prioritise use cases that assess, exclude, rank or steer candidates or workers.
Why is AI literacy important in HR?
Recruiters, hiring managers, HR professionals and line managers need to understand where AI output comes from, what errors may occur, how bias can arise and when they must escalate. For high-risk use cases, role-based training is part of the evidence file.
Sources
This hub links to official and primary sources where possible. Always check the current legal text and supervisory information for definitive qualification.