HR & policy self-service assistant for employees

HR teams spend much of their day answering the same questions: how many leave days are left, which travel expense rules apply, what does the works agreement say about home office. An assistant grounded in the approved policies (retrieval-augmented generation) and connected read-only to the HR system answers these in the employee's language, cites the source, and opens an HR case when it is unsure. Typical benefits are shorter wait times for employees and fewer tier-1 tickets for HR; the key risks are wrong or outdated answers, leaking other people's data, and co-determination duties.

Problem & value

• Today: answers live in PDFs, intranet pages, SharePoint folders and in the heads of HR staff. Employees email or call HR; shared mailboxes and ticket queues grow, answers differ between agents. Frontline, shift and multilingual workforces often have no easy access at all.
• AI value: instant 24/7 answers in many languages, consistent wording tied to the current policy version, personalised lookups (own balances, own payslip dates), and structured hand-over of the cases that really need a human. HR gets analytics on what people ask and which policies are unclear.
• Evidence: vendors publish deflection and resolution rates, but these are self-reported and depend on scope and measurement; this file therefore gives no generic percentage. Measure your own baseline (tickets per employee, first-response time) before rollout.

Variants

• SMB (50–500 staff): assistant inside the HR/payroll tool or Microsoft Teams, over a few dozen policy documents; HR manager owns content.
• Enterprise / shared-service centre: integrated with HR case management (ServiceNow, SAP, Workday), multi-country policy sets, entitlement-aware retrieval, multilingual.
• Public sector: policies from collective agreements (TVöD/TV-L), Dienstvereinbarungen, Reisekostenrecht; Personalrat co-determination (BPersVG / state laws); on-prem or sovereign cloud preferred.
• Frontline / shift workers: mobile or WhatsApp-style channel, voice input (see AGENT-style voice bots).
• Onboarding buddy: same engine, answers for new hires in the first 90 days (IT access, benefits, forms).
• Boundary: answering from generic documents without HR-system access is KNOW-002; searching across all intranet sources is KNOW-001; customer-facing support bots are in the CUST family.

How it works

1. employee question (Teams / web / app, SSO identity)
2. guardrails (PII/abuse filter, prompt-injection check, scope check: HR topic?)
3. query rewrite + language detection
4. retrieval: hybrid search (BM25 + embeddings) over policy index, filtered by country, entity, role
5. rerank
6. top passages with version/date metadata
7. [optional tool calls: HRIS API read-only for the caller's own data only]
8. LLM answer with mandatory citations; "I don't know" if evidence is weak
9. actions: prefill form / ticket
10. human approval (manager or HR)
11. escalation to HR case with transcript; feedback buttons; analytics

• Entitlement-aware retrieval: policy documents carry metadata (country, legal entity, employee group, validity period). The filter is applied by identity before the LLM sees text, so a German employee does not get the US policy and a normal employee does not see management-only documents.
• Personalised data: tool calls run with the employee's own token (on-behalf-of), never a super-user, so the model cannot be talked into reading someone else's record.
• Human-in-the-loop: HR owns the source content and an approval step for new policy versions; low-confidence or sensitive topics (dismissal, harassment, illness, pay disputes, whistleblowing) route directly to humans; consequential actions need confirmation.
• Not a decision system: the assistant informs; it does not decide on leave approval, pay, or performance (that would change the regulatory class, see Risks).

Data preparation

The core data are HR policies: works agreements, collective agreements, travel and leave rules, onboarding guides, benefit descriptions and FAQ. They typically live on the HR file share, in the intranet, in the HR suite and in mail answers that HR has sent repeatedly. Typical gaps are several versions of the same policy, rules that differ by location or contract type, scanned works agreements, answers that exist only in HR staff's heads, and no clear "valid from" date.

Preparation steps: - Collect the valid documents and name an HR owner per topic (PREP-013, PREP-004). - Retire outdated versions and add metadata: valid from, applies to (location, group) (PREP-010). - OCR scanned agreements (PREP-006, DOC-014). - Harvest real questions from the HR mailbox and ticket history, remove personal details (PREP-020, PREP-022) and write approved answers (PREP-016). - Keep individual employee records (salary, absences, personnel files) out of the index unless a separate, access-controlled design exists (PREP-024). - Build a test set with HR (PREP-018).

Effort is moderate, about a third of the project, because the corpus is small but the content decisions are sensitive. See GUIDE-003.

Implementation paths

Which path for whom

• SMB: Path A, because HR suites and knowledge-assistant products with EU hosting offer ready-made policy Q&A, and the policy corpus itself is small and not highly sensitive.
• Enterprise: Path B, because integration with the HR system, single sign-on and country-specific rules call for a tailored build on cloud APIs.
• Developer: Path C, because a small open-source RAG stack (Open WebUI with Ollama) is enough for a policy corpus.
• Public sector: Path C on sovereign EU or German infrastructure, because staff data and collective agreements require controlled hosting and the staff council expects it.

Path A — Buy (recommended for most SMBs)

• HR suites with built-in assistants: SAP SuccessFactors with Joule (SAP, Germany/EU), Workday (agents such as Sana Concierge listed on its AI page [3]), Personio (Germany) and other European HR tools, plus ServiceNow HR Service Delivery and Microsoft-based setups. Check each vendor's current feature and EU-hosting status.
• Enterprise service platforms: ServiceNow (HR Service Delivery / virtual agent), Moveworks (now part of ServiceNow), Leena AI, Microsoft 365 Copilot agents over SharePoint.
• Low-code agent builders: Microsoft Copilot Studio can build agents on SharePoint/Dataverse knowledge and publish them to Teams and other channels, with analytics and evaluation sets [4].
• Choose A when: policies are in a standard HR suite or M365, volume is moderate, and the vendor offers EU data residency, a DPA, and works-council-friendly logging controls.

Path B — Build on cloud APIs

• Models: Mistral (France) via La Plateforme with EU hosting — current catalogue includes Mistral Medium 3.5 and Mistral Small 4 (Apache 2.0) [5]; Azure OpenAI / Microsoft Foundry and AWS Bedrock in EU regions; Anthropic Claude and Google Gemini via EU-region offerings on the hyperscalers.
• Retrieval: Azure AI Search, Elastic, OpenSearch or pgvector for hybrid search; a multilingual embedding model and reranker (e.g. Cohere Rerank, Voyage, or the open BGE family).
• Orchestration: LangChain/LlamaIndex or plain code; MCP servers or REST connectors to SuccessFactors, Workday, Personio, DATEV HR; Teams bot via Bot Framework.
• Build when: you need custom flows (e.g. multiple HR systems, country logic) or tight integration.

Path C — Self-host (open source)

• Models: Mistral Small 4 (Apache 2.0) or Ministral 3 14B/8B for light loads [5]; Qwen3 (8B–32B, run via Ollama/vLLM) [6]; check each license. 7–30B models are sufficient because answers are grounded in retrieved text.
• Embeddings/rerank: BAAI bge-m3 (MIT, 100+ languages, dense + sparse + multi-vector, 8192 tokens) [7] plus a bge reranker.
• Stack: Open WebUI or Dify, LlamaIndex/Haystack, Qdrant/pgvector/OpenSearch, vLLM or Ollama, Keycloak for SSO.
• Hardware: one 24–48 GB GPU serves a few hundred users for small models; an 8-bit 24–32B model fits HW2–HW3. Effort: a platform engineer part-time plus an HR content owner. Choose C for data-sovereignty demands (public sector, works-council requirement to keep data on-prem).

Path D — Custom training

Rarely needed. Fine-tuning a small model for tone, German HR terminology or strict answer format, or training a domain reranker on logged question-passage pairs, can help after a few thousand real questions exist. Start with retrieval quality, evaluation sets and prompt design instead.

State of the art (2026)

• Grounded RAG is the baseline: hybrid search + reranking + citation-forcing prompts; agentic retrieval (the model reformulates and retries queries) is now offered in platforms such as Copilot Studio [4].
• Agents over HR systems: vendors moved from FAQ bots to agents that execute tasks (request leave, update bank details) via tools, with human approval steps.
• Multilingual: current models and embeddings (e.g. bge-m3 [7]) handle German, English, Turkish, Polish, Ukrainian etc. well enough for policy Q&A; verify with native-speaker test sets.
• Evaluation: use gold question sets per policy domain, graded for correctness, citation accuracy and correct abstention; run them on every policy or model change (Copilot Studio ships evaluation sets [4]).
• Limits / failure modes: outdated or contradictory policy versions; hallucinated entitlements (e.g. invented leave days); conflicts between law, collective agreement and works agreement (the model cannot decide hierarchy reliably); employees uploading personal data in free text; prompt injection via documents; identity/permission leaks; over-trust in confident answers. Mitigate with citations, abstention, disclaimers ("not legal advice, HR decides"), and human escalation.

Shared resources & combinations

• Shared building blocks: hybrid index + embedding + reranker + LLM gateway + SSO/ACL layer are the same as in KNOW-002 (company RAG) and KNOW-001 (enterprise search); one platform can serve all three with separate indices and access rules. Guardrails and PII filters are reusable across CUST bots.
• Hardware: HW2–HW3 inference node, shared with other RAG workloads; embeddings run on CPU or small GPU.
• Combinations: KNOW-011 (keeps FAQs and policies current from tickets), KNOW-014 (organisational memory), KNOW-012 (expert finding for questions that need a person), KNOW-004/KNOW-009 for technical or standards Q&A in the same interface, DOC-001-style document pipelines for submitting receipts or forms that the assistant starts.

Effort, cost & KPIs

• Effort: A: days to weeks (configuration, content clean-up is the real work). B: ~1–3 months for a production pilot with evaluation and integrations. C: similar plus infrastructure ops. D: add months.
• Cost drivers: per-user licences (A), tokens and search service (B), GPU and ops (C), content curation and evaluation by HR (all), integration with HR systems, works-council process.
• KPIs: containment/deflection rate (questions solved without human), answer accuracy and citation correctness on gold set, abstention precision, escalation hand-over quality, first-response time, HR tickets per 100 employees, employee satisfaction (thumbs up/down), share of questions that revealed a policy gap, incident count (wrong answers reported).

People & change

Involve HR from day one as process owner, together with management, the data protection officer (DPO) and, if there is one, the works council. Legal or compliance checks employment-law statements; IT handles single sign-on and hosting.

Works council: yes in practice. The assistant is used by employees and logs their questions, so it is suited to monitor behaviour: §87(1) no. 6 BetrVG co-determination applies, and §90(1) no. 3 requires timely information. The works council may engage an AI expert under §80(3). Conclude a works agreement covering logging, retention, no performance evaluation and the scope of topics. In the public sector the staff council (Personalrat) is involved (GUIDE-004).

Data protection: questions can reveal health, family or conflict situations, so a DPIA is likely advisable, an AVV is needed with the provider, and sensitive topics should be routed to a human HR contact (GUIDE-005).

Start with a pilot in one location or department. Communicate openly what is logged and who sees it. Typical concerns: "HR is being replaced" (answer: the assistant handles routine questions, HR handles cases) and "my questions are monitored" (answer: the works agreement and pseudonymous analytics). Offer a clear escalation to a person. Support AI literacy with a short briefing (GUIDE-006, GUIDE-008).

Risks, regulation & compliance

• EU AI Act: a pure information chatbot is generally not an Annex III system, but Art. 50(1) requires that people are informed they interact with an AI unless this is obvious from context; the obligation applies from 2 August 2026 [2] (check current status of any postponement proposals). Art. 4 AI literacy applies to deployers. It becomes high-risk if used for decisions affecting terms of work relationships, task allocation by behaviour/traits, or monitoring and evaluating performance and behaviour (Annex III, point 4) [1] — do not extend the assistant to such functions without a full high-risk compliance process. Emotion recognition of employees is prohibited (Art. 5) and must not be added.
• GDPR / BDSG: employee data processing needs a legal basis (Art. 6, Art. 88 GDPR, § 26 BDSG as applicable); DPIA is advisable because of the employment context and special-category data (health, union membership, religion) that employees may type; data minimisation, retention limits for chat logs, DPA with providers, EU data residency, no training on customer data, role-based access, deletion rights.
• Co-determination: in Germany the works council has co-determination rights over technical systems suited to monitor behaviour or performance (§ 87(1) no. 6 BetrVG) [8]; chat logs and analytics fall under it. Agree a works agreement early (purpose limitation, no individual evaluation, log retention, access). Public sector: Personalrat. Other countries have comparable consultation duties.
• Legal correctness: answers on labour law are not legal advice; route legally sensitive questions to HR/legal.
• Security: prompt injection through ingested documents, over-broad service accounts, leakage across entities; apply least privilege, output filtering, logging, red-team tests.

Getting started

1. Pick the top 30–50 recurring HR questions from tickets and mailboxes and name an HR content owner; clean and version the source policies, add metadata (country, entity, validity).
2. Involve the works council and data-protection officer; draft purpose, log and DPIA outline.
3. Run a pilot (Path A or a Copilot Studio / Open WebUI prototype) with 20–50 users in one country, citations mandatory, escalation button to HR.
4. Build a gold question set (including trick and out-of-scope questions) and evaluate before each change.
5. Add read-only HRIS lookups and one low-risk action (e.g. leave balance, certificate request), then review KPIs and expand scope.

Sources

1. EU AI Act, Annex III (employment, point 4) — https://artificialintelligenceact.eu/annex/3/ (accessed 2026-10)
2. EU AI Act, Article 50 transparency obligations — https://artificialintelligenceact.eu/article/50/ (accessed 2026-10)
3. Workday, AI and agents overview — https://www.workday.com/en-us/artificial-intelligence.html (accessed 2026-10)
4. Microsoft Learn, What is Copilot Studio — https://learn.microsoft.com/en-us/microsoft-copilot-studio/fundamentals-what-is-copilot-studio (accessed 2026-10)
5. Mistral AI docs, models overview — https://docs.mistral.ai/getting-started/models/models_overview/ (accessed 2026-10)
6. Ollama library, Qwen3 — https://ollama.com/library/qwen3 (accessed 2026-10)
7. BAAI bge-m3 model card — https://huggingface.co/BAAI/bge-m3 (accessed 2026-10)
8. § 87 BetrVG, Gesetze im Internet — https://www.gesetze-im-internet.de/betrvg/__87.html (accessed 2026-10)