EU AI Act for HR: CV Screening, Interview Scoring and Employee Monitoring Rules
Annex III point 4 puts hiring, scoring and worker monitoring squarely in the high-risk bracket. Here is what HR and People Ops owe under the AI Act, and what is banned outright.
HR is the department where the AI Act bites first. Annex III point 4 covers employment and worker management, and the tools in that category — screening, scoring, monitoring — are exactly the ones mid-market companies buy fastest.
The result is a population of organisations running at least one high-risk AI system without ever having made a decision to do so.
What falls inside Annex III point 4
Point 4 covers AI systems intended to be used for:
- Recruitment or selection of natural persons, including targeted advertising and profiling of applicants.
- Decisions affecting terms of work-related relationships — promotion, termination, task allocation, remuneration, performance evaluation.
- Monitoring and evaluating the performance and behaviour of persons in those relationships.
Every one of those is a routine feature of modern HR tech: an ATS with resume ranking, an interview-scheduling tool with “fit scoring”, a productivity analytics platform, an engagement survey with sentiment analysis.
The classification is on the system as you use it, not the vendor’s marketing. A “candidate experience platform” that ranks applicants is a high-risk AI system.
The hard prohibition that catches people out
Before any of the high-risk analysis, check Article 5. Two of the eight prohibited practices sit in HR’s territory:
- Emotion inference at work — Article 5(1)(f). Inferring emotions from a person’s face or voice in the workplace is banned. That includes “sentiment” plugins on your meeting platform that read facial affect, and any interview tool that scores emotional state.
- Biometric categorisation to infer characteristics — Article 5(1)(g). Sorting candidates by inferred protected characteristics (ethnicity, religion, sexuality, health status) is prohibited.
There is no disclosure, no consent flow and no conformity route that legitimises these. If your stack contains either, the decision is to remove it, not to document it.
Also note: social scoring under Article 5(1)(c) covers evaluation of a person’s social behaviour in a way that leads to detrimental treatment in contexts unrelated to where the behaviour occurred. A workplace analytics tool that produces a score used to deny someone a promotion can cross that line.
The deployer duties that follow from Annex III point 4
If a use case is high risk and you deploy it, Article 26 gives you a specific list:
- Use it as instructed. Follow the provider’s documented instructions of use — including the intended purpose. Using a screening tool “creatively” is your violation, not theirs.
- Assign human oversight. A named individual, trained, with the authority and resources to override or stop the system. Not a shared inbox.
- Ensure input data quality. Relevant and sufficiently representative for the purpose, and check it — historic hiring data that encodes past bias will reproduce it.
- Keep logs. Automatically generated logs retained for at least six months, so a decision can be reconstructed.
- Inform workers before deployment. Where a system is put into service at the workplace, affected workers and their representatives must be informed in advance.
- Cooperate with authorities and provide the information and access needed for an inspection.
- Respect the rights of affected persons. That includes Article 86: a right to a clear, meaningful explanation of the role the system played in a decision producing legal or similarly significant effects.
Add GDPR to the same workflow: Article 22 restricts solely automated decisions with legal or similarly significant effects, Article 35 requires a DPIA for systematic monitoring and large-scale profiling, and national works councils may hold a genuine consultation right over the deployment itself.
What this means for the buying process
Most AI Act exposure in HR is created at purchase, not at rollout. A four-question gate before the contract prevents most of it:
- Does the tool rank, score or screen people? If yes → Annex III point 4, high risk.
- Does it read emotion, tone or facial affect to make that judgement? If yes → stop; check Article 5 first.
- Can a named human reverse its output, and is that reversal recorded? If no → you cannot meet Article 14 oversight.
- What does the vendor give you for logs and instructions of use? If nothing, you cannot meet Article 26(2) and 26(6).
Contract terms matter too: obtain the provider’s technical documentation references, the Annex IV materials, the conformity assessment route and a commitment to notify you when the rule pack changes their classification.
The workplan, in order
1. Discover every HR use case. Include the free tiers and the browser extensions. A survey plus SSO logs reliably surfaces tools nobody in People Ops remembered to mention.
2. Classify each with the reasoning. Annex III point 4 for screening and evaluation; Article 5 check first for emotion and biometric categorisation; minimal for pure scheduling.
3. Kill the prohibited ones. Document the removal decision and the date.
4. Assign oversight by name. One accountable person per high-risk system, with written authority to suspend the tool.
5. Consult workers. Inform workers and their representatives before the system goes into service — in several jurisdictions this is a negotiation, not a notice, and it takes longer than the technical work.
6. Build the evidence. Register rows, input-data checks, log retention, training records for the operators, and the DPIA where GDPR applies.
7. Re-run when the rule pack moves. Classification is a versioned output, not a one-off memo.
The sequence matters most at step 5. Works-council consultation is the item that routinely slips, because it is the only one with an external clock that you do not control.
The deadline picture
Article 5 transparency and the penalty regime have applied since 2 August 2026. The Annex III high-risk deployer duties were moved to 2 December 2027 by the Digital Omnibus. Penalties for breaching the obligations that are already in force apply now, and a prohibited practice has been punishable since 2 February 2025.
Fourteen months is comfortable for the documentation and unpleasantly tight for consultation, remediation and re-procurement combined. Start with the register.
Frequently asked questions
Is AI CV screening high risk under the EU AI Act?
Yes. Recruitment and selection of natural persons sits in Annex III point 4, which makes AI-assisted screening and applicant ranking a high-risk use. The deployer takes on the full Article 26 obligations: instructions of use, human oversight, input-data quality, six-month log retention, worker notification and cooperation with authorities.
Can we use AI to monitor employee productivity?
Monitoring and evaluating worker performance and behaviour is inside Annex III point 4, so a monitoring system is high risk and carries the deployer duties. It must also satisfy GDPR — systematic monitoring requires a lawful basis and normally a DPIA — and emotion inference from faces or voice at work is prohibited outright under Article 5(1)(f).
Do we need to consult our works council before deploying AI in hiring?
Where the system is put into service at the workplace, Article 26(5) requires affected workers and their representatives to be informed beforehand. In Germany, the Netherlands, France and several other member states, national law adds genuine consultation or co-determination rights. Treat this as a scheduled workstream, not a notification email.
What if the AI screening tool is only “assistive” and a human makes the final call?
Human involvement must be real, not nominal. Article 26(2) requires oversight by a person with the competence, training and authority to intervene. A recruiter who rubber-stamps a ranked list has not exercised oversight — and where GDPR Article 22 applies, meaningful human review has its own requirements.
Does the AI Act apply to employee engagement surveys with sentiment analysis?
Text-based sentiment summary on aggregated survey responses is generally minimal risk. Inferring emotion from faces or voice — for example a webcam-based “engagement score” — is prohibited in the workplace. Classify on what the system actually does to the individual, not on the survey’s branding.
Put this guide to work on your own estate.
Run the free exposure scan to see which of these articles actually apply to your company — 12 questions, four minutes, no card. Confirm your email and the full report unlocks.
- Free forever plan
- No credit card
- EU-hosted · GDPR-native
Build a live AI register in minutes — upgrade only when you need the audit pack.
Keep reading
AI in healthcare under the EU AI Act: clinical decision support and high-risk duties
11 min
Risk & ClassificationThe Annex III High-Risk AI Systems List: All Eight Use Cases Explained with Examples
11 min
Risk & ClassificationThe Eight Prohibited AI Practices Under Article 5: A Compliance Checklist
10 min