HR documents: job ads, references (Arbeitszeugnis), policies

HR teams write the same three kinds of text again and again: job ads, employment references and internal policies. Each one is legally sensitive. Job ads must not discriminate and, under the EU Pay Transparency Directive, must state pay. German references must be favourable and clear yet still true. Policies must match works agreements and current law. LLMs produce good first drafts in seconds from structured inputs. The value comes from combining drafting with compliance checks (AGG wording, gender-neutral titles, pay range, reference grade wording) and a mandatory human sign-off. The AI writes the text. It does not make or evaluate the HR decision.

Problem & value

• Today: job ads are copied from old postings and edited by hand. References are written by HR from manager questionnaires using phrase libraries or templates. Policies are drafted by HR or legal in Word, often with no clear link to the works agreements they depend on. The result is slow turnaround, inconsistent tone and legal risk.
• Legal pressure on job ads: § 11 AGG forbids advertising a position in breach of the anti-discrimination rule in § 7(1) AGG [5]. Wording such as "Deutsch als Muttersprache", "junges Team" or gendered titles ("Sekretärin") can create a presumption of discrimination under § 22 AGG and lead to compensation claims under § 15 AGG [6]. From 7 June 2026, the EU Pay Transparency Directive (2023/970) requires the initial pay or a pay range to be given in the vacancy notice or before the interview, and bans questions about pay history [7].
• Legal pressure on references: § 109 GewO gives employees a right to a written reference (simple or qualified) that is clear and contains no hidden coded messages. Electronic form is allowed only with the employee's consent [4].
• Measured benefit (vendor data): Textio reports that its customers' applicant pools are on average 24% more qualified and 12% more diverse, and that positions are filled 17% faster. In an A/B test at Square, a rewritten job ad received twice as many applications [3]. These are vendor-published case-study figures, not independent studies. The research behind gendered-wording checks found that masculine-coded job ad wording reduces how appealing the job is to women [9].

Variants

• Job ads: posting generation from a role profile, multi-channel variants, translation for cross-border hiring (→ LANG-009), plain-language versions for blue-collar roles (→ LANG-010), and bias/inclusivity scoring of existing ads.
• References (Arbeitszeugnis): final and interim references, training certificates (Ausbildungszeugnis), and English "reference letters" for international staff. The draft is written from the manager's ratings per criterion in the German grading wording.
• Policies & HR communications: remote-work, travel, IT-use and AI-use policies, onboarding handbooks, works-agreement drafts for negotiation, and letters (e.g. contract change notices). The Q&A on top of these policies belongs to KNOW-003.
• Public sector: job ads that follow civil-service tariff and announcement rules (TVöD/TV-L, Bestenauslese under Art. 33(2) GG), with mandatory hints for severely disabled applicants.
• Private use: employees checking their own reference wording, or writing applications (→ PERS-012).

How it works

1. structured input (role profile + pay band | manager rating sheet | policy brief)
2. retrieval of house templates, approved phrase library, existing policies & works agreements
3. LLM drafting with a fixed system prompt (tone, structure, legal must-haves)
4. rule + LLM checks: AGG-sensitive terms, m/w/d / gender-neutral title, pay range present, age/origin/disability proxies, reference grade consistency, no invented facts
5. human review in editor (HR + manager; legal/works council for policies)
6. approval & publication (ATS / career site) or signature (reference)
7. archive with version

• Drafting: a general LLM is enough. Quality depends on structured inputs and examples of the company's own good texts.
• Reference grading: the manager rates criteria such as expertise, work quality and conduct on a scale. The LLM maps each rating to the established wording and must keep the overall grade consistent. A deterministic phrase library is often kept as the source of truth, with the LLM only joining the phrases smoothly. This avoids accidental downgrades through nuance.
• Compliance checks: a lexicon (e.g. a gender-decoder word list, AGG term list) plus an LLM reviewer prompt. Pay-range and mandatory-field checks are plain rules.
• Human-in-the-loop: every text is approved by a named person. For references, the manager confirms the facts and the rating. For policies, HR/legal approve and the works council is involved where co-determination applies.

Data preparation

Needed: role profiles and pay bands (HR system, Excel), manager notes, for references the employee master data, task lists and criterion ratings (HR system, personnel file, often paper), plus existing policies, works agreements, collective agreements and legal texts (file shares, intranet, PDFs).

Typical gaps: role profiles missing or outdated, ratings never captured in structured form, policies in several versions and as scans, no phrase library for grade wording, and personnel data in places with wide access.

Preparation steps: - Define a structured input form for references: tasks, criterion ratings, grade (see PREP-019, PREP-016). - Collect approved job ad and reference texts as a phrase library and examples (see PREP-016, PREP-015). - Digitise and OCR policies, keep one valid version and index it (see PREP-006, PREP-013, DOC-014). - Pseudonymise personal data for tests (see PREP-022) and check the legal basis (see PREP-025); lock down access rights (see PREP-023).

Rough share: about 30 to 40 percent of the effort. See GUIDE-003.

Implementation paths

Which path for whom

• SMB: Path C. Employee data, ratings and references are very sensitive, and a simple local stack (for example Ollama with a small model and a template) runs on HW1/HW2 without sending HR data to a third party.
• Enterprise: Path C. HR data stays inside the company on a shared internal AI platform, which also serves other HR use cases.
• Developer: Path C. Open-weight models with a rule-based compliance checker give full control over reference grade wording and logging.
• Public sector: Path C. Personnel data and procurement rules call for self-hosting or a sovereign German cloud.

Path A — Buy (off-the-shelf)

• Job ads / inclusive language: Textio Recruiting (US) [3]; generative job-description features in ATS/HCM suites such as SAP SuccessFactors (EU data centres), Workday, Personio (Munich, EU), and LinkedIn Hiring Assistant (turns rough notes into job descriptions, an add-on to LinkedIn Recruiter) [8]. Check which edition includes the feature.
• References: German reference software such as Haufe Zeugnis Manager (phrase-library based, over 12,000 sample formulations; the product page names no AI feature) [1], personalwerk Zeugnisgenerator, Zeugnisfairness and verlingo. These are mainly rule/phrase-based, so they are predictable and legally safe, with LLM polishing added in some products [2].
• General assistants: Microsoft 365 Copilot (EU Data Boundary), ChatGPT Enterprise, Mistral Le Chat Enterprise (EU provider) and Langdock (Berlin, EU) with shared HR prompt templates.
• Choose A when: you are an SMB with standard needs. A reference tool plus an ATS with AI drafting covers 80% of the work.

Path B — Build on cloud APIs

• Models: Mistral (EU-hosted, e.g. Mistral Medium/Large), Claude via AWS Bedrock Frankfurt or Google Vertex AI EU, GPT via Azure OpenAI EU Data Zone, and Gemini on Vertex AI EU regions.
• Build: template service with JSON schemas per document type, RAG over the policy repository (embeddings + vector DB), and a check step as a second LLM call with a fixed rubric plus a lexicon.
• Integration: ATS (Personio, SAP SuccessFactors, Workday, d.vinci, softgarden) via API to push drafts. HR master data feeds reference drafts. DMS stores signed versions.
• Choose B when: you need a custom process across ATS, HCM and DMS, multilingual output or your own check rules.

Path C — Self-host (open source)

• Models: open-weight LLMs with strong German such as Mistral Small 3.x, Qwen3 (14B–32B), Gemma 3 27B, or gpt-oss-20b/120b. Serve them with vLLM or Ollama, and use Open WebUI as the chat front end with shared prompt "workspaces".
• RAG: multilingual embeddings (e.g. BGE-M3, multilingual-e5) with Qdrant or pgvector over policies and works agreements.
• Hardware: HW2 (one 24–32 GB GPU) is enough for a quantised 14–32B model at HR-department volumes. Use HW3 for 70B+/120B models and for serving several departments.
• Why choose it: references and policies contain personal and co-determined content. On-prem helps public bodies and works councils that refuse cloud processing of employee data.

Path D — Custom training

Rarely worth it. A LoRA fine-tune of a small open model on a few hundred approved company references or job ads can lock in house style and grade wording. However, retrieval of examples (few-shot) usually gets close. Training on past references means processing employee data for a new purpose, which needs a GDPR purpose-compatibility check and pseudonymisation first.

State of the art (2026)

• Drafting quality is solved for mainstream languages. Frontier and mid-size open models write fluent German and English HR texts. The differentiators are workflow, phrase governance, compliance checks and integration with ATS/HCM.
• Agentic recruiting suites (LinkedIn Hiring Assistant, Workday, SAP Joule) bundle job ad drafting with sourcing and screening [8]. Those downstream steps are high-risk (AGENT-008, AGENT-033) and should be governed separately from the writing step.
• Pay transparency turns job ads into structured documents: pay range, gender-neutral title and objective criteria are now mandatory fields, which helps automated checking [7].
• Known failure modes:
• LLMs show measurable gender bias in generated reference letters, in word choice ("warm" vs. "role model") and in invented content [10].
• Subtle wording shifts can unintentionally change a German reference grade.
• Models invent tasks or achievements that are not in the input.
• Models write generic ads that are interchangeable between employers.
• Models quote outdated legal rules in policies.
• Mitigations: keep phrase libraries authoritative, run gender-swap tests on reference prompts, add grounding checks against the input and retrieve the current legal sources.

Shared resources & combinations

• Same stack (llm-general + RAG + guardrails): LANG-002 correspondence, LANG-003 reports, LANG-011 brand-voice rewriting, and KNOW-003 HR policy self-service. The policies written here are the knowledge base KNOW-003 answers from, so they share one vector index and one LLM server (HW2/HW3).
• Pipeline neighbours: DATA-025 provides the pay bands and gender-neutral job architecture for pay-transparent ads. The job ad then feeds CONV-017 (recruiting chatbot), AGENT-033 (sourcing) and AGENT-008 (screening, high-risk). Onboarding documents feed AGENT-009.
• Boundary: LANG-036 covers recurring performance reviews and 360 feedback, which are high-risk evaluation. This use case covers one-off documents. The reference uses ratings the manager has already given and does not create the evaluation itself.
• Transform add-ons: LANG-009 (translation) and LANG-010 (Leichte Sprache) use the same model.

Effort, cost & KPIs

• Path A: days. Costs are per-user SaaS or an add-on to the ATS/HCM licence. Reference tools are typically priced per user or per number of references.
• Path B: 3–8 weeks for templates, checks and ATS integration. Token costs are negligible at HR volumes.
• Path C: 4–10 weeks including GPU server setup. The ongoing effort is model updates and prompt/phrase maintenance.
• KPIs: time to publish a job ad, reference turnaround (days from request to signature), share of drafts accepted with minor edits, compliance-check findings per ad, number of applicants and share of under-represented applicants per ad (A/B tested), number of reference disputes or corrections, and policy update lead time.

People & change

Involve management, the HR lead as process owner, the DPO, the works council, legal for AGG and reference law, and IT. Hiring managers and HR staff are the end users. Include the works council from the planning stage, not at rollout.

Works council: yes. Even if the AI only drafts, employee data and ratings flow through the system, which can be suited to monitoring behaviour or performance (§87(1) no. 6 BetrVG, BAG: objective suitability suffices); §90(1) no. 3 requires information in good time, §80(3) lets the council engage an AI expert, and §95(2a) applies if AI supports selection. Public sector: staff council. A works agreement is advisable (GUIDE-004).

Data protection: employee data is personal data; check a DPIA (Art. 35 GDPR) and conclude an AVV with every processor (GUIDE-005). The AI must not rate or rank people; ratings come from the manager.

AI literacy: support HR and managers with short training on bias, reference law and review duty (Art. 4 AI Act, GUIDE-006). Pilot with two or three HR staff on job ads first, then references. Typical concern: "the AI decides on my reference." Answer: it only formulates the manager's rating; HR and the manager sign off (GUIDE-008).

Risks, regulation & compliance

• EU AI Act — depends:
• Text drafting: a tool that only writes or polishes text that a human then approves is usually not high-risk. It performs a preparatory task or improves a previously completed human activity (Art. 6(3)(b)/(d)) [11].
• High-risk cases: Annex III(4)(a) explicitly lists AI used "to place targeted job advertisements" and to filter or evaluate candidates. If the same tool decides who sees the ad, it is high-risk [11]. A reference generator that derives the performance grade itself from performance or behaviour data falls under Annex III(4)(b) (evaluating performance). Any profiling keeps a system high-risk regardless of Art. 6(3).
• Design rule: the human sets the rating and the AI only formulates it.
• Art. 50: Art. 50(4) disclosure applies to text on matters of public interest and has an editorial-review exemption [12]. It rarely bites for HR texts, but disclosing AI use to staff is good practice. AI literacy (Art. 4) applies to HR users.
• GDPR: references and policy cases contain employee data. You need a lawful basis (Art. 6(1)(b)/(c) and § 26 BDSG), a DPA with the vendor, EU hosting, no training on customer data and deletion of prompts. Never put health, union or disciplinary details into a prompt without a clear need.
• German labour law:
• References: § 109 GewO requires truth, goodwill, clarity and no coded phrasing [4].
• Job ads: AGG §§ 1, 7, 11, 15 and 22 [5][6], and § 164 SGB IX (check whether severely disabled applicants can fill the post).
• Works council: introducing AI tools usually triggers co-determination under § 87(1) no. 6 BetrVG when the tool can monitor behaviour or performance [13], plus information rights under § 90 BetrVG.
• Pay transparency: the national transposition of Directive 2023/970 (deadline 7 June 2026, with several Member States late) brings mandatory pay information and gender-neutral recruitment [7].
• Mitigations:
• Keep lexicon-based AGG and gender-decoder checks.
• Run gender-swap tests on reference prompts.
• Have the employee confirm the reference facts.
• Keep a version history of policies.
• Keep an audit trail of AI draft vs. final text.
• Never let the tool decide grades or targeting.

Getting started

1. Collect your 20 best job ads, references and policies as reference examples. Define a style guide and an approved reference phrase/grade table.
2. Set up an EU-hosted assistant (Path A or B) with three fixed templates: job ad, reference from rating sheet, and policy draft/update.
3. Add a compliance checklist: AGG term list, m/w/d or gender-neutral title, pay range, SGB IX notice and reference-grade consistency. Run it automatically on every draft.
4. Agree on the use with the works council and the DPO, and document that humans set ratings and approve all texts.
5. Measure turnaround and applicant response before and after. A/B test rewritten job ads.

Sources

1. Haufe — Zeugnis Manager / Arbeitszeugnisgenerator — https://www.haufe.de/hr/arbeitszeugnisgenerator
2. OMR Reviews — Haufe Zeugnis Manager reviews and alternatives — https://omr.com/de/reviews/product/haufe-zeugnismanager/all/2
3. Textio — "You can hire smarter" (customer results, products) — https://textio.com/blog/you-can-hire-smarter
4. § 109 Gewerbeordnung (Zeugnis) — https://www.gesetze-im-internet.de/gewo/__109.html
5. § 11 AGG (Ausschreibung) — https://www.gesetze-im-internet.de/agg/__11.html
6. ADVANT Beiten — Gestaltung von rechtssicheren Stellenanzeigen (AGG case law) — https://www.advant-beiten.com/aktuelles/gestaltung-von-rechtssicheren-stellenanzeigen
7. EU Pay Transparency Directive — what employers hiring in Europe need to know (Art. 5, transposition status) — https://euremotejobs.com/eu-pay-transparency-directive-what-employers-need-to-know/ ; legal text: https://eur-lex.europa.eu/eli/dir/2023/970/oj
8. TechCrunch — LinkedIn launches Hiring Assistant — https://techcrunch.com/2024/10/29/linkedin-launches-its-first-ai-agent-to-take-on-the-role-of-job-recruiters
9. Gaucher, Friesen & Kay (2011), Evidence that gendered wording in job advertisements exists and sustains gender inequality, J. Personality and Social Psychology — https://doi.org/10.1037/a0022530
10. Wan et al. (2023), "Kelly is a Warm Person, Joseph is a Role Model": Gender Biases in LLM-Generated Reference Letters, EMNLP Findings — https://arxiv.org/abs/2310.09219
11. EU AI Act — Annex III and Article 6 — https://artificialintelligenceact.eu/annex/3/ ; https://artificialintelligenceact.eu/article/6/
12. EU AI Act — Article 50 transparency obligations — https://artificialintelligenceact.eu/article/50/
13. § 87 BetrVG (Mitbestimmungsrechte) — https://www.gesetze-im-internet.de/betrvg/__87.html
14. Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj