The European Commission's draft guidelines on high-risk AI, published on 19 May 2026, make one thing clear: HR and recruitment are no longer edge cases. If you use AI to find, rank, evaluate or monitor people in work contexts, you need to explain why that system does or does not fall under Annex III point 4 of the AI Act.
The document is still a consultation draft, but it is already the most concrete interpretation of high-risk AI in work and recruitment. Use this article as a practical checklist alongside the HR AI hub, the route for recruitment and selection and the route for worker management and monitoring.
Why HR Is So Sensitive
The Commission focuses on the power imbalance between employer, candidate, worker, self-employed person or platform worker. An AI score can determine who sees a job ad, who gets an interview, who gets promotion opportunities, who receives fewer shifts or who disappears from a platform. That affects income, career prospects, privacy, non-discrimination and dignity at work.
That is why the AI Act does not treat HR AI as ordinary process automation. The key question is not whether the tool "only supports" a human. The question is whether the output meaningfully influences access to work, employment terms, promotion, termination, task allocation or performance evaluation.
The Core: Point 4(a) and 4(b)
| Route | What it covers | Typical HR systems |
|---|---|---|
| 4(a) Recruitment and selection | Targeted job ads, analysis and filtering of applications, candidate evaluation | ATS ranking, CV screening, sourcing, matching, online assessments, interview scoring |
| 4(b) Worker management and work relationships | Employment terms, promotion, termination, task allocation based on behaviour or traits, monitoring and evaluation | Workforce management, performance analytics, shift allocation, platform scoring, productivity monitoring |
For the general Article 6(3) filter, see the deep-dive on the Commission guidelines. For a first triage, use the Annex III Classifier.
7 Checks for HR and Recruitment
1. Does the AI rank or score candidates?
CV ranking, fit percentages, top-N shortlists, match scores and "recommended candidates" quickly fall under point 4(a). It does not matter that a recruiter still clicks "invite". If the AI meaningfully shapes the order or shortlist, you are in high-risk territory.
2. Does the AI optimise who sees a vacancy?
Employer branding or generic job promotion is not automatically high-risk. Targeted job ads become different when the system uses personal characteristics, behaviour, profile data or similar signals to determine who sees the vacancy. Where profiling is involved, the Article 6(3) filter is effectively unavailable.
3. Is CV parsing really just parsing?
A parser that converts a PDF into fixed fields may fall within a narrow procedural task. But once the system infers skills, interprets employment gaps, weighs education level or produces a suitability score, the function changes from administration to evaluation.
4. Is the chatbot logistical or selective?
A chatbot that schedules interviews or answers procedural questions usually sits outside high-risk. A chatbot that asks knockout questions, evaluates answers, summarises motivation letters for selection or moves candidates to the next round belongs in the classification check.
5. Does the assessment tool measure behaviour, personality or performance?
Online assessments, game-based tests, video interviews, language or voice analysis and personality models are sensitive. If the output is used to assess suitability, potential, reliability or culture fit, the use case usually falls under point 4(a). Pay special attention: emotion recognition in the workplace is in principle prohibited under Article 5, except for strictly medical or safety reasons.
6. Does the system continue after hiring?
Many HR risks shift after hiring to point 4(b): schedules, work allocation, performance reviews, bonuses, promotion, non-renewal, account deactivation or productivity monitoring. If AI uses behaviour, personal traits, ratings or performance to allocate work or opportunities, high-risk is very close.
7. Can you evidence the vendor claim?
"AI Act compliant" is not evidence. Employers remain responsible as deployers for proper use, human oversight, logging, information to candidates and workers, incident handling and, where needed, DPIA or FRIA. Ask for system purpose, input data, evaluation method, bias tests, instructions for use, logging, change management and the classification rationale.
What May Fall Outside High-Risk?
The guidelines leave room. Not every HR tool with AI is automatically high-risk.
- Interview scheduling without candidate evaluation
- Automatic acknowledgement emails for applications
- Inclusive language checks on job descriptions
- Format conversion or organisation of CV information without scoring or ranking
- Onboarding support after hiring without impact on employment terms or evaluation
- Feedback shown only to the worker, not to a manager, HR file or performance process
The practical line is simple: does the AI only help with administration, or does it influence an opportunity, evaluation, right, employment term or work allocation?
Internal Route for HR Teams
Separate recruitment from worker management
Use the pages for route 4(a) and route 4(b) to assess systems by function.
Record evidence per system
Use the HR AI Evidence Pack to record use case, vendor claim, data sources, human oversight, information duties and open risks.
Test your AI against the Article 6(3) filter
Interactive self-assessment, updated for the Commission guidelines of 19 May 2026. 9 steps, personal report with reasoning, vendor questions and next steps.
Frequently Asked Questions
Practical questions about the draft guidelines for high-risk AI in HR and recruitment.