AI skills inference and talent intelligence under the EU AI Act: the invisible layer hitting both 4(a) and 4(b)

Behind many modern HR platforms — Workday Skills Cloud, SAP SuccessFactors Talent Intelligence Hub, Eightfold's talent intelligence, ChartHop's skills layer — sits a horizontal layer HR leaders rarely discuss explicitly: skills inference. AI infers skills, experience level, expertise or seniority from CVs, project data, learning activity, performance reviews and internal task history, building a personal skills profile that subsequently feeds every other HR decision. For the EU AI Act this makes skills inference a dangerous blind spot: one underlying AI layer touches simultaneously 4(a) recruiting and 4(b) worker management.

This post explains what skills inference is, why it is AI Act relevant, and how HR and compliance teams can address it in their classification.

What skills inference precisely does

Skills inference is the automatic deriving of skills, experience level, expertise or seniority from data not explicitly stating those skills. Input sources vary per platform but typically include:

• CVs and application profiles — for candidates
• Employee profiles and self-reported skills — for existing workers
• Project history and task execution — for workers in role
• Performance reviews and feedback data — peer and manager input
• Learning records and certifications — completion patterns
• External profiles — LinkedIn, GitHub, publications

The output is a personal skills profile — usually with confidence scores per skill — used in matching algorithms for recruiting, succession planning, project staffing, learning recommendations, performance benchmarking and compensation calibration. One skills layer, many use cases.

Why this hits both 4(a) and 4(b)

Skills inference is in itself not a "decision". It is a data layer. But under the AI Act we look at what the AI output ultimately feeds:

• For candidates — if inferred skills are used to match or rank candidates (recruiting AI, sourcing suggestions), it falls within 4(a).
• For existing workers — if inferred skills affect compensation, mobility, project allocation or assessment, it falls within 4(b).
• Cross-cutting — one Talent Intelligence Hub typically feeds both simultaneously. Classifying as one deployment or as two is a tactical choice with implications for your oversight structure.

The practical consequence: skills inference is often the heart of enterprise HR platforms, and therefore the focus of your AI Act analysis — not a fringe case.

The bias challenge in skills inference

Skills inference has a specific bias category rarely well-addressed in vendor documentation:

• Background bias — those who express skills in technical jargon versus business language get different inferences
• Language bias — non-native English or Dutch can lead to lower confidence scores
• Pattern bias — if training data comes mostly from certain demographics, the model carries those patterns through
• Self-reporting bias — workers who assert their skills assertively get higher scores than equally-skilled colleagues who are more modest

For your FRIA and bias evaluation that means: don't just ask whether the matching algorithm is bias-tested, but whether the skills inference itself is validated for your population.

Step-by-step for skills inference dossier

What to do now

For HR architecture and compliance teams working with talent intelligence platforms: treat skills inference as priority within your 4(a) and 4(b) trajectories. Inventory this month, downstream use case mapping alongside, worker self-service for inferred skills data in your 2026 roadmap. Document via the HR AI hub and the HR AI Evidence Pack.