AGG and AI shortlisting: what German HR needs to check

Why the AGG does not disappear behind GDPR when you add AI

Vendor checklists often stop at the DPA, EU hosting, and deletion deadlines. The AGG bites earlier: the advertisement and every criterion you use to sort applications. Algorithmic shortlisting does not rewrite the prohibitions — it scales them. A biased criterion in a spreadsheet hits one person; the same criterion in a model hits the whole funnel.

• AGG § 1 protects against discrimination on grounds including ethnic origin, sex, religion, disability, age, and sexual identity
• § 3 distinguishes direct and indirect discrimination — proxies (postcode, CV gaps, “culture fit”) can have the same effect
• § 7 covers access to employment, so recruiting is in scope
• § 11 covers job advertisements: they must not breach § 7
• GDPR (Art. 22, transparency) and the AI Act (Art. 14, Annex III point 4) add duties — they do not replace the AGG

Where AI specifically raises the AGG risk

The model rarely stores the protected characteristic as a column. It stores correlates. That is why “we do not save gender” is not a defence on its own.

• Training on historical hires reproduces whom you have hired — including the gaps
• Keywords such as elite universities, sports clubs, or particular employers can proxy origin and age
• Unstructured “culture fit” scoring is legally and diagnostically weak — and hard to explain when someone asks
• Video or voice analysis is outside what most DACH teams can justify; leave it out
• A rejection the score has already made, with the recruiter only confirming, is the Art. 22 pattern the CJEU described in the SCHUFA line: a decisive score, a human as a formality

Checklist before the shortlist goes live

Run this with legal or the DPO. “The vendor is GDPR-ready” is not a row on this list.

Fix the job ad first — before anything scores

The funnel is already skewed if qualified people never apply. Phrases such as “young team”, “digital native”, “German as a mother tongue” without a demonstrable requirement, or unnecessary physical must-haves, are the classic § 11 issue. A language check before publish is cheaper than explaining a homogeneous shortlist later.

• Derive requirements from the role profile, not from the last person who held the job
• State language as a CEFR level when the role needs it — not as a proxy for origin
• Replace “team fit” with observable collaboration, or delete it
• Use the same ad internally and externally; document side doors (“we already know them”) or you create an unchecked second track

Screening: structure it, justify it, allow dissent

Skills-first matching against a written profile is the antidote to gut feel — if a human reads the rationale. A rank without “why” forces recruiters to trust the model. That is the opposite of AI Act Art. 14.

• Same core interview questions and the same rating scale — structured screening does not rescue an unstructured conversation
• Blind early screening (name, photo, age later) reduces a known anchor; it does not anonymise later rounds
• Overrides are a feature, not a bug: if nobody disagrees, oversight is dead
• Rejections short and factual, without a fake rationale that opens a new AGG surface — keep the file internally

What software can carry — and what it does not certify

An ATS can enforce form fields, language flags, explainable scores, override logs, and deletion deadlines. It cannot find that your process complies with the AGG. Compliance stays with you: configuration, training, works-council agreement, a real last decision.

• Job-ad checks and structured criteria reduce typical breaks — they are not an anti-discrimination certificate
• EU hosting and a DPA answer privacy questions, not what the model sorts on
• “Bias-free AI” as a vendor claim is a warning; serious vendors describe limits and human oversight
• Counsel, the works council, and equal-treatment officers remain part of the rollout — not a footer link

Bottom line

AGG and AI shortlisting are the same question in two languages: what are we sorting on, and could we explain it to a supervisor, a court, and the person we rejected? Document the ad, the criteria, the rationale, and the human decision, and you cut risk and time-to-hire together. Buy only a score, and you buy speed — and the record that comes with it.