AI in HR and Recruiting: Value, Limits and Fair Implementation

What AI can and cannot do in HR recruiting

In practice, AI in recruiting means two things: classic machine learning that ranks applications by learned patterns, and generative language models that draft or summarise text. Both are assistive tools. They speed up repetitive work but do not make sound suitability judgements and lack the context an experienced recruiter brings.

The core principle: AI proposes, humans decide. Deriving a rejection or an invitation from an algorithm alone is legally sensitive and professionally risky. Value emerges where AI handles the groundwork so specialists gain time for interviews, assessment and relationship-building.

The three most common use cases

• Drafting job ads: language models produce fast first drafts, keep tone consistent and help write in gender-fair, non-discriminatory language - final polish and factual accuracy stay with HR.
• Screening and matching: AI can structure CVs, sort by defined competencies and suggest longlists. The key is competency-based, explainable matching rather than opaque overall scores.
• Scheduling and communication: assistants coordinate interview slots, send acknowledgements and answer standard questions via chatbot - a relief at high application volumes, provided answers are accurate and auditable.

Bias and fairness: the central risk

AI models learn from historical data - and inherit the biases of the past. If a team has mostly hired certain profiles, a model can perpetuate exactly that pattern and systematically disadvantage other groups. A publicly known example is a large technology company that had to scrap an internal recruiting tool because it disadvantaged women.

Fairness does not arise automatically; it must be measured and actively engineered. This includes removing sensitive attributes, checking for proxy variables (such as postcodes or CV gaps), running regular fairness tests across groups, and asking whether an attribute is genuinely job-relevant.

revDSG, the EU AI Act and legal guardrails

In Switzerland the revised Data Protection Act (revDSG) has applied since 1 September 2023. Application data is personal data and partly sensitive. Article 21 revDSG on automated individual decisions is central: where a purely automated decision has a legal effect or significantly affects the person, they must be informed and may request a review by a human. The supervisory authority is the EDOEB.

The EU AI Act is also relevant: it classifies AI systems for personnel selection and recruitment as high-risk. Swiss companies are not directly bound but may be affected when hiring for the EU market or using EU tools. A data protection impact assessment is advisable where AI evaluates applicants systematically and at scale.

Introducing AI responsibly: a practical guide

• Clarify the purpose: for each use case, define which problem AI should solve and how success is measured - time saved is no substitute for quality.
• Human in the loop: establish that final selection and rejection decisions are made by people.
• Transparency to applicants: clearly disclose whether and how AI is used, and offer a contact point for questions.
• Vet data and vendors: check server location, data-processing agreements and whether inputs are used for training - never put sensitive application data into unvetted public tools.
• Document and monitor: record criteria, model versions and fairness tests in writing, and review outcomes regularly for bias.
• Involve the team: train HR in working with AI and engage employee representation early where it exists.

Swiss context: multilingualism and proportionality

The Swiss labour market is multilingual. AI can help publish ads consistently in German, French, Italian and English and structure applications from different language regions uniformly. Ensure quality and fairness are equally high across all languages - depending on training data, models are not equally strong in every language.

The principle of proportionality applies throughout: collect and process only what the role requires. Less but clean, relevant data yields fairer outcomes than maximal profiles - and lowers legal risk at the same time.