AI in CV screening under the EU AI Act: from parsing to ranking, and where the high-risk line runs

CV screening is for most employers the first place AI touches the recruitment process. What was once simple field extraction — name, education, experience — has grown in modern ATSes and sourcing tools into a layer of inferred skills, match scores and pipeline recommendations. For the EU AI Act this makes the difference between "we use a handy tool" and "we deploy an Annex III point 4(a) high-risk system".

This post explains where the classification line runs, how to assess your CV screening setup, and which questions you need to answer before every vendor choice.

What AI does in CV screening (in 2026)

Modern CV screening combines four layers that were previously separated:

1. Document parsing — extraction of fields from a CV (name, work experience, education, skills as literally mentioned)
2. Skills inference — inferring skills, experience or seniority not explicitly on the CV, based on patterns in text and context
3. Matching against vacancies — score per candidate per requisition based on match between inferred profiles and job requirements
4. Pipeline recommendations — automatic suggestions for shortlisting, follow-up actions or communication

Under the EU AI Act these four layers do not all fall under Annex III in the same way. Understanding where the line runs is crucial for recruiters and HR leaders.

Where the classification line runs

The EU Commission has refined the scope of Annex III point 4(a) via guidelines and the AI Act itself. The practical reading in 2026:

• Pure document parsing (only field extraction without interpretation) — generally outside 4(a). It is administrative support.
• Skills inference for administration — if inferred skills are only used to order CVs or make them searchable, it usually stays outside 4(a).
• Skills inference for matching or ranking — if the inferred skills are then used to score candidates against vacancies, it tips into 4(a).
• Match scores per candidate — by definition ranking. Annex III point 4(a) high-risk.
• Automated pipeline actions based on AI output — if AI decides which candidates advance to next stages, you are in 4(a) territory.

It is a layered analysis, not "all CV screening is high-risk". Many employers use only parsing and stay outside 4(a) with that. Others use full AI matching and are firmly in.

The practical differences between vendors

Different HR tools position themselves differently on this ladder. For a detailed feature-by-feature analysis see the separate vendor posts:

• SAP SuccessFactors (Joule, Talent Intelligence Hub) — full stack, 4(a) + 4(b)
• Recruitee (Hire AI, Smart Capture) — configurable scale
• Greenhouse (match scores in Structured Hiring framework) — depends on active features
• Personio (Recruiting AI) — SME context
• AFAS and Visma — mostly outside 4(a) in baseline
• LinkedIn Recruiter (Recommended Matches, Hiring Assistant) — universal blind spot
• HireVue — beyond CV screening, in assessment 4(a)
• Bullhorn — agency context with dual-deployer impact

The difference sits not only in the vendor, but also in how you configure the tool.

Step-by-step: assessing your CV screening setup

What to do now

CV screening is a good starting point for your HR AI Act trajectory. It touches virtually every employer, the classification is layered per feature, and you immediately have something usable for your AI register. Start with the tool inventory this week, use the HR AI hub for the route map and the HR AI Evidence Pack for documentation.