EHS incident & near-miss analysis and risk assessment drafting (Gefährdungsbeurteilung)

Safety teams collect incident and near-miss reports in free text, but little time remains to read them all, find patterns or keep risk assessments current. An LLM-based assistant classifies each report, clusters recurring causes, flags high-potential events and drafts risk-assessment entries, instruction material and checklists aligned with ArbSchG and DGUV rules. A qualified person (Fachkraft für Arbeitssicherheit, EHS manager) reviews and signs off everything; the AI never decides on protective measures.

Problem & value

• Today: reports arrive as paper forms, email, app entries or chat messages in varied quality and languages. They are filed under a few coarse categories, rarely analysed beyond the legally required accident statistics, and risk assessments are often generic Word/Excel documents that lag behind real operations. Small firms frequently have no up-to-date assessment at all.
• Legal baseline (Germany): employers must assess the hazards of work (§ 5 ArbSchG, including physical, chemical, biological hazards, work equipment, processes, qualification and psychological strain), document results, measures and effectiveness checks, and record accidents with more than three days of incapacity (§ 6 ArbSchG). Employees must be instructed at hiring, on change of duties and when new equipment is introduced (§ 12 ArbSchG). Accidents with more than three days of incapacity must be reported to the accident insurer within three days (§ 193 SGB VII) [1][2][3][4].
• AI value: consistent coding of unstructured reports, detection of weak signals (the same machine, shift or task recurring in near-misses), faster first drafts of assessments and instruction material, and better preparation for audits and ISO 45001 certification. Vendors market such features for EHS suites [5][6]; independently measured effect sizes on injury rates are not yet established in public evidence, so treat any vendor claim as unverified until piloted on your own data.

Variants

• SMB / Handwerk: voice or photo report on a phone → draft assessment for a trade (e.g. workshop, kitchen, warehouse) from a template, reviewed by the external safety specialist or BG advisory service.
• Manufacturing / chemicals: near-miss clustering by machine, substance and shift; linking to hazardous-substance registers and operating instructions (Betriebsanweisungen).
• Construction & logistics: multilingual reports from site and depot crews, with translation and coding into a common taxonomy.
• Healthcare: needle-stick, violence and patient-handling incidents; strictly separate from clinical incident systems (CIRS) which have own confidentiality rules.
• Public sector / schools: assessment drafting for offices, schools, fire brigades; works/staff council involvement is standard.
• Group level: harmonised taxonomy and leading-indicator dashboards across sites and countries (OSHA, HSE, local rules mapped as parameters).
• Not in scope: video-based PPE monitoring (VIS-007), posture/ergonomics from video (VIS-024), product operating manuals (LANG-022).

How it works

1. report intake (form / app / voice
2. ASR / email / photo)
3. normalisation & translation, PII minimisation (names
4. roles)
5. classification & extraction (event type, task, equipment, body part, causal factors, potential severity)
6. deduplication + embedding clustering over history (recurring patterns, hot spots)
7. retrieval over own assessments, instructions, legal/technical rules
8. LLM drafting: root-cause hypotheses, assessment entries (hazard
9. risk
10. TOP measures), instruction & checklist drafts
11. human review (EHS officer, safety specialist; works council informed)
12. approval & versioning
13. measure tracking, effectiveness review, dashboards, audit evidence

• Extraction/classification: an LLM with a fixed schema (e.g. a taxonomy based on the company's own categories and the DGUV accident statistics fields), optionally a fine-tuned encoder classifier for high volume.
• Clustering and trend detection: embeddings plus topic clustering find recurrences that keyword search misses; counts are checked with ordinary statistics, not LLM judgement.
• Drafting with retrieval (RAG): drafts cite the passages of the existing assessment and the regulation they rely on; the reviewer sees sources. Risk ratings follow the company's own matrix (e.g. likelihood × severity) rather than model opinion.
• Human-in-the-loop: every causal hypothesis is a suggestion; investigation (e.g. 5-Why, TOP hierarchy of substitution, technical, organisational, personal measures) remains with trained people. Assessments are signed by the responsible person.

Data preparation

Needed data: incident and near-miss reports, accident book entries, investigation reports, existing Gefährdungsbeurteilungen, inspection findings, SDS/work instructions and corrective-action records. In SMBs these live on paper forms and logbooks, in Excel lists, e-mails, file shares and sometimes in an existing EHS or HR system. Typical gaps: handwritten or scanned reports, inconsistent categories, free-text narratives with names, near-misses never reported, duplicates, and unclear access rights since injury data is health data.

Preparation steps: - Digitise and OCR paper records, and move to digital capture at source (PREP-006, PREP-008). - Consolidate Excel lists into one tidy table with consistent fields (PREP-011). - Define a category scheme for causes and severity before classification (PREP-019) and harvest labels from past closed cases (PREP-020). - Pseudonymise names and health details before analysis or model use (PREP-022). - Restrict access to injury data and set purpose limits (PREP-025). - Build a test set of 30 to 50 cases with the safety specialist (PREP-018).

Effort share: about 30 to 40 percent, mostly digitising and category cleanup. See GUIDE-003 and DOC-014.

Implementation paths

Which path for whom

• SMB: Path A. EHS suites with German hosting (e.g. Quentic) cover incident management and risk assessment as a standard process, with no AI team needed.
• Enterprise: Path B. Incident and near-miss data is sensitive and strategic, and linking it with own plant, HR and shift data favours an integrated build with controlled models over a closed suite.
• Developer: Path C. Incident narratives contain health data; open-weight models on own or sovereign infrastructure avoid sending them to third parties.
• Public sector: Path C. Sovereignty and procurement rules, plus health-related content, favour open-weight models in German or EU sovereign cloud.

Path A — Buy (off-the-shelf)

• EHS suites with AI features: Quentic (German vendor, software hosted in Germany; incident management, root-cause analysis, risk assessment modules, AI features advertised) [5]; Cority (Cortex AI, incident and risk management, AI motion capture for ergonomics) [6]; also Gensuite (Genny AI), Intelex, Sphera, VelocityEHS, Enablon and others — verify AI scope and EU hosting per vendor.
• Local German options: suites from smaller vendors and BG-linked tools for Gefährdungsbeurteilung templates; many safety-specialist service providers offer assessment software without AI.
• Check: EU data residency, data processing agreement, works-council agreement, export of data, German-language support, audit trail and versioning, traceability of AI drafts.
• Choose A when: standard EHS processes, no in-house AI team, desire for integrated measure tracking.

Path B — Build on cloud APIs

• Models: general LLMs with structured outputs from EU-hostable providers (Mistral, or frontier models on Azure/AWS/Google EU regions); embeddings for clustering; a reasoning model for drafting when rules are complex.
• Stack: intake form or Teams/WhatsApp-style bot → n8n or Power Automate → LLM extraction → database (Postgres + pgvector) → review UI; legal texts and own assessments indexed for RAG.
• Choose B when: you already run a safety database and want AI enrichment or custom taxonomies, or need multilingual intake without buying a suite.

Path C — Self-host (open source)

• Models: open-weight LLMs such as Qwen3-30B-A3B (Apache-2.0, 30.5B total / 3.3B active parameters, long context) [7], Mistral or Llama-family models; embeddings such as BGE-M3 (MIT, 100+ languages) [8]; Whisper-class ASR for voice reports.
• Stack: vLLM or Ollama, Qdrant/pgvector, a local classification service, n8n; documents stay on-prem.
• Hardware: HW2 is enough for report volumes of a mid-sized company and interactive drafting; HW3 for larger models or many concurrent users. Fits plants with confidentiality concerns (accident details, trade secrets) and for works councils that prefer on-prem.

Path D — Custom training

Rarely required. Worthwhile to fine-tune an encoder classifier or small LLM when you have thousands of labelled reports in one domain (e.g. a large logistics group with a proprietary taxonomy), or to reach low-cost multilingual coding. Labels come from the EHS team's corrections; keep evaluation sets per site and language.

State of the art (2026)

• LLMs code free-text safety reports into structured fields at useful quality and handle multilingual input; they are strongest for summarising, normalising and drafting, weaker for causal reasoning over incomplete facts.
• Retrieval-grounded drafting with citations to own documents and regulation texts is the standard pattern; suites ship it as assistants inside the EHS system [5][6].
• Leading-indicator thinking (near-miss reporting rates, closure times, high-potential events) pairs naturally with text analytics, but published, independent evidence that AI analysis reduces injuries is thin; claims should be tested in a pilot.
• Limits and failure modes: hallucinated legal citations or limit values; generic "boilerplate" assessments that look complete but ignore the real workplace; under-reporting bias (people stop reporting if they fear blame or surveillance); mis-coding rare serious events as routine; stale drafts when the process changes. Never use an LLM to determine exposure limits or protective measures without verification against current TRGS/DGUV sources.

Shared resources & combinations

• Same stack (llm-general + embedding-text + vector-db + text-classifier): LANG-012 (routing/classification), LANG-015 (information extraction from free text), LANG-016 (topic modelling), KNOW-002 (RAG over company documents). One embedding service and one vector store can serve all of them on a single HW2 server.
• Documentation siblings: LANG-022 (technical documentation and operating manuals) for Betriebsanweisungen; LANG-004 (HR documents) for instruction records.
• Signals upstream: VIS-007 (PPE/safety video monitoring) and VIS-024 (ergonomics from pose analysis) generate events that can feed the same incident database; PRED-006 and ANOM-005 (machine health and process anomalies) correlate with near-miss clusters on specific equipment.
• Shared building blocks: ASR for voice reports, translation, review UI and workflow engine (n8n, Camunda), measure-tracking tickets.

Effort, cost & KPIs

• Path A: days to weeks to configure; licensing per user or per site. Path B: 1–3 months for intake, classification, review UI and RAG. Path C: 2–4 months plus a GPU server and operations. Path D: add 1–3 months for labelling and evaluation.
• Cost drivers: report volume, multilingual needs, integration with HR/ERP/maintenance, review effort by safety specialists, works-council process.
• KPIs: near-miss reports per 100 employees (reporting culture), share of reports coded without correction, time from report to assessed risk, time to close measures, share of assessments reviewed within the planned interval, recurrence rate of similar events, lost-time injury frequency (outcome, lagging).

People & change

Involve management (safety is a legal duty of the employer), the EHS or safety-specialist lead, and the people who report and investigate incidents. Bring in IT and information security for hosting and access control, HR for employee-data questions, and the DPO early: incident reports include health data, so a DPIA is likely, plus an AVV with every processor (GUIDE-005).

Works council: yes. Analysis of incidents, patterns by person or shift is objectively suited to monitor behaviour or performance, so §87(1) no. 6 BetrVG applies (BAG), and the works council also co-determines occupational safety measures under §87(1) no. 7. §90(1) no. 3 requires timely information on AI-supported work processes, and §80(3) allows an AI expert. Aim for a works agreement that rules out individual blame ranking (GUIDE-004). Public sector: staff council.

AI literacy: support safety staff with training on verifying AI root-cause suggestions (GUIDE-006). Pilot at one site for two to three months (GUIDE-009). Typical concern: reports used to blame individuals, which would lower near-miss reporting. Address it with clear purpose limits, pseudonymised analysis and a just-culture message (GUIDE-008). The safety specialist signs off every risk assessment; AI only drafts.

Risks, regulation & compliance

• EU AI Act: drafting and analytics for safety documentation is generally minimal risk (AI literacy duty, Art. 4). It becomes high-risk if the system is used to monitor and evaluate employees' performance or behaviour or to allocate tasks based on individual behaviour (Annex III, point 4) [9]; do not turn incident analysis into individual blame scoring. Emotion recognition at the workplace is prohibited (Art. 5(1)(f)); Annex III high-risk obligations apply from 2 December 2027 after the AI Omnibus (Reg. (EU) 2026/1744), Art. 50 from 2 August 2026. Voice bots interacting with staff need disclosure (Art. 50).
• GDPR: reports contain health data (Art. 9) and names of injured persons and witnesses; apply data minimisation, role-based access, pseudonymised analytics, retention limits, DPIA where appropriate, processor contracts, no training of vendor models on your data.
• Works council (BetrVG): introduction of technical systems that can monitor behaviour or performance triggers co-determination (§ 87(1) Nr. 6), and accident-prevention rules fall under Nr. 7 [10]; conclude a Betriebsvereinbarung before rollout.
• Occupational safety law: responsibility for the assessment stays with the employer (§ 5, 6 ArbSchG); AI drafts do not discharge this duty. Statutory accident reporting (§ 193 SGB VII) and works/staff council co-signing are unaffected [4]. Outside Germany: OSHA recordkeeping, UK RIDDOR, ISO 45001.
• Mitigations: human sign-off, citation of sources in drafts, version history of AI vs. final text, blame-free reporting policy, regular sampling audits of AI coding.

Getting started

1. Export 12–24 months of incident and near-miss reports and your current assessments; measure how many are categorised and when each was last reviewed.
2. Agree the taxonomy and risk matrix with the safety specialist, occupational physician and works council; draft the Betriebsvereinbarung.
3. Pilot AI coding on 200–500 historical reports and compare with expert labels; look at false negatives on serious events first.
4. Add drafting for one area (e.g. warehouse or workshop) with retrieval over your own documents and the relevant rules; require reviewer sign-off.
5. Track the KPIs quarterly and expand sites and languages only after reporting rates stay stable.

Sources

1. Gesetze im Internet — ArbSchG § 5 Beurteilung der Arbeitsbedingungen — https://www.gesetze-im-internet.de/arbschg/__5.html
2. Gesetze im Internet — ArbSchG § 6 Dokumentation — https://www.gesetze-im-internet.de/arbschg/__6.html
3. Gesetze im Internet — ArbSchG § 12 Unterweisung — https://www.gesetze-im-internet.de/arbschg/__12.html
4. Gesetze im Internet — SGB VII § 193 Anzeige von Versicherungsfällen — https://www.gesetze-im-internet.de/sgb_7/__193.html
5. Quentic — EHS software (incidents, risk assessment, AI) — https://www.quentic.com/
6. Cority — EHS platform, Cortex AI — https://www.cority.com/
7. Qwen3-30B-A3B-Instruct-2507 model card (Hugging Face) — https://huggingface.co/Qwen/Qwen3-30B-A3B-Instruct-2507
8. BAAI BGE-M3 model card (Hugging Face) — https://huggingface.co/BAAI/bge-m3
9. EU AI Act — Annex III high-risk systems, point 4 employment — https://artificialintelligenceact.eu/annex/3/
10. Gesetze im Internet — BetrVG § 87 Mitbestimmungsrechte — https://www.gesetze-im-internet.de/betrvg/__87.html