Intro & scope
Text-in/text-out use cases (EU/GDPR focus): WRITE (drafting), TRANSFORM (summarise, translate, simplify, proofread), CLASSIFY (routing, sentiment, moderation, extraction) and ANALYZE (corpus analysis, monitoring, research). Scope boundaries: LANG-008 summarises documents in general, KNOW-* covers retrieval/Q&A over knowledge bases, DOC-* covers layout-heavy document processing, SPEECH-* covers audio. LANG-012 is generic routing while LANG-015 extracts entities; LANG-013 covers sentiment on feedback while LANG-018 monitors external channels. Legal drafting (LANG-023/034/035) and HR/surveillance uses (LANG-004, 032, 036) carry higher AI Act and works-council obligations.
Use cases
| ID | Title | Pattern | Maturity | Path | HW tier | AI Act risk |
|---|---|---|---|---|---|---|
| LANG-001 | Marketing copy & product description generation | LANG.WRITE | commodity | A | HW2 | limited |
| LANG-002 | Email & business correspondence drafting | LANG.WRITE | commodity | A | HW2 | minimal |
| LANG-003 | Report, proposal & offer writing assistant | LANG.WRITE | mainstream | A | HW2 | minimal |
| LANG-004 | HR documents: job ads, references (Arbeitszeugnis), policies | LANG.WRITE | mainstream | A | HW2 | depends |
| LANG-005 | SEO content production at scale with brand voice | LANG.WRITE | mainstream | A | HW2 | limited |
| LANG-006 | Grant, tender & RFP response writing | LANG.WRITE | mainstream | A | HW3 | minimal |
| LANG-007 | Social media post generation & content repurposing | LANG.WRITE | commodity | A | HW2 | limited |
| LANG-008 | Summarisation of long documents & document sets | LANG.TRANSFORM | mainstream | A | HW2 | minimal |
| LANG-009 | Machine translation & localisation | LANG.TRANSFORM | commodity | A | HW2 | minimal |
| LANG-010 | Plain-language & easy-read simplification (Leichte Sprache) | LANG.TRANSFORM | early-adopter | A | HW2 | limited |
| LANG-011 | Proofreading, style & brand-voice rewriting | LANG.TRANSFORM | commodity | A | HW2 | minimal |
| LANG-012 | Text classification & routing (emails, tickets, requests) | LANG.CLASSIFY | mainstream | A | HW1 | depends |
| LANG-013 | Sentiment & customer feedback analysis (reviews, NPS, surveys) | LANG.CLASSIFY | mainstream | A | HW2 | minimal |
| LANG-014 | Text content moderation & toxicity detection | LANG.CLASSIFY | mainstream | A | HW2 | minimal |
| LANG-015 | Named-entity recognition & information extraction from free text | LANG.CLASSIFY | mainstream | B | HW1 | minimal |
| LANG-016 | Topic modelling & trend detection in text corpora | LANG.ANALYZE | mainstream | A | HW2 | minimal |
| LANG-017 | Competitive & market intelligence from web and news | LANG.ANALYZE | mainstream | A | HW2 | minimal |
| LANG-018 | Social listening & media monitoring | LANG.ANALYZE | mainstream | A | HW2 | minimal |
| LANG-019 | Qualitative research coding (interviews, open survey answers) | LANG.ANALYZE | mainstream | A | HW2 | minimal |
| LANG-020 | Regulatory change & policy horizon scanning | LANG.ANALYZE | early-adopter | A | HW3 | minimal |
| LANG-021 | Fact-checking & misinformation detection | LANG.ANALYZE | early-adopter | B | HW2 | minimal |
| LANG-022 | Technical documentation authoring: operating manuals & safety instructions (CCMS) | LANG.WRITE | early-adopter | A | HW2 | minimal |
| LANG-023 | Patent application drafting & office-action response | LANG.WRITE | early-adopter | A | HW3 | minimal |
| LANG-024 | Promotional & marketing compliance review (MLR, financial promotions, green claims) | LANG.CLASSIFY | early-adopter | A | HW3 | minimal |
| LANG-026 | Public consultation & objection (Einwendungen) analysis | LANG.ANALYZE | early-adopter | B | HW3 | minimal |
| LANG-027 | Synthetic respondents & persona simulation for market research | LANG.ANALYZE | early-adopter | A | HW3 | minimal |
| LANG-028 | Automated data-to-text journalism & newsroom assistant | LANG.WRITE | mainstream | A | HW2 | limited |
| LANG-029 | Creative writing co-authoring & manuscript assessment | LANG.WRITE | early-adopter | A | HW2 | minimal |
| LANG-030 | Sign-language translation (text/speech to signing avatar, sign recognition) | LANG.TRANSFORM | early-adopter | A | HW2 | limited |
| LANG-031 | Customs tariff (HS/CN/TARIC) & export-control classification | LANG.CLASSIFY | early-adopter | A | HW3 | minimal |
| LANG-032 | Employee communications surveillance (eComms & voice) for conduct risk | LANG.CLASSIFY | early-adopter | A | HW3 | high |
| LANG-033 | Bank transaction enrichment, categorisation & booking suggestions (Kontierung) | LANG.CLASSIFY | mainstream | A | HW1 | minimal |
| LANG-034 | Legislative drafting & amendment consolidation assistant | LANG.WRITE | early-adopter | B | HW3 | depends |
| LANG-035 | Litigation drafting assistant (pleadings, briefs, Schriftsätze) | LANG.WRITE | early-adopter | A | HW3 | minimal |
| LANG-036 | Performance review, goal-setting & 360-feedback assistant | LANG.WRITE | early-adopter | A | HW3 | high |
| LANG-037 | Investor-side ESG, EU Taxonomy & SFDR portfolio analytics | LANG.ANALYZE | mainstream | A | HW3 | minimal |
| LANG-038 | Online review response & local reputation management | LANG.WRITE | mainstream | A | HW1 | minimal |
| LANG-039 | Outside-counsel invoice review & legal spend management | LANG.CLASSIFY | early-adopter | A | HW2 | minimal |
| LANG-040 | EHS incident & near-miss analysis and risk assessment drafting (Gefährdungsbeurteilung) | LANG.ANALYZE | early-adopter | A | HW2 | minimal |
Shared stack
Common layers: an OpenAI-compatible inference gateway (vLLM or llama.cpp server behind LiteLLM), a shared embedding service (multilingual-e5 / bge-m3), one vector store (pgvector or Qdrant), Presidio-style PII redaction plus guardrails, a workflow engine (n8n or Temporal) with human-approval steps, and audit logging. The same stack serves KNOW, DOC, AGENT, CONV and SPEECH families.
- HW1 (CPU/laptop, no GPU): 3-8B quantised model (Qwen/Llama/Mistral small via llama.cpp/Ollama), small encoder classifiers (xlm-roberta / setfit), pgvector, n8n. For routing, extraction, review replies, bookkeeping suggestions.
- HW2 (single 24-48 GB GPU): 14-32B instruct model quantised via vLLM, bge-m3 embeddings plus reranker, Qdrant or pgvector, NLLB/MADLAD or LLM for translation, optional Whisper for audio. For drafting, summarisation, translation, sentiment, topic modelling, monitoring.
- HW3 (multi-GPU 80 GB+ or EU-hosted API): 70B-class or reasoning model (or EU-sovereign hosted API with DPA), hybrid search (OpenSearch + reranker), knowledge graph, agent framework with approval gates. For tender/legal/patent drafting, regulatory scanning, compliance review, eComms surveillance (with DPIA, works council, retention limits).
- Path guidance: A = managed/EU API or SaaS first; B = self-host or build when data sensitivity requires it. Share one gateway, one embedding index per corpus and one guardrail layer across all use cases to avoid duplicate infrastructure.