Intro & scope
AI use cases that turn raw company data into trusted, usable and reported information. Three patterns: - DATA.PREPARE: cleaning, master data, extraction, cataloguing, quality, labelling, synthetic data, schema/EDI mapping, OT contextualisation. - DATA.ANALYZE: conversational BI, spreadsheet assistants, AutoML, geospatial, FP&A, process mining, spend, experimentation, UX analytics. - DATA.REPORT: management reporting, ESG/CSRD, regulatory reporting, product carbon footprint/DPP, pay transparency.
Scope boundaries: DATA-001 (governed NL-to-SQL on warehouses) vs DATA-006 (ad-hoc spreadsheets); DATA-002 (generic recurring narratives) vs DATA-024 (planning/variance for finance) vs DATA-012/020/021/025 (regulation-driven reports); DATA-003 (master data entities) vs DATA-010 (product data) vs DATA-015 (spend lines) vs DATA-022 (partner file mapping). Predictive modelling beyond AutoML lives in PRED-, document extraction in DOC-. Gaps in numbering (DATA-014, 017-019) are intentional/unassigned.
Use cases
| ID | Title | Pattern | Maturity | Path | HW tier | AI Act risk |
|---|---|---|---|---|---|---|
| DATA-001 | Natural language to SQL & conversational BI | DATA.ANALYZE | early-adopter | A | HW3 | minimal |
| DATA-002 | Automated reporting & dashboard narratives | DATA.REPORT | mainstream | A | HW2 | minimal |
| DATA-003 | Data cleaning, deduplication & master data management | DATA.PREPARE | mainstream | A | HW1 | minimal |
| DATA-004 | AI web data extraction (intelligent scraping) | DATA.PREPARE | early-adopter | B | HW2 | minimal |
| DATA-005 | Synthetic data generation (privacy & training data) | DATA.PREPARE | early-adopter | B | HW2 | minimal |
| DATA-006 | Spreadsheet analysis assistant | DATA.ANALYZE | mainstream | A | HW2 | minimal |
| DATA-007 | Data catalogue & metadata enrichment | DATA.PREPARE | early-adopter | A | HW2 | minimal |
| DATA-008 | Data quality monitoring & observability | DATA.PREPARE | mainstream | C | HW1 | minimal |
| DATA-009 | AutoML for business analysts | DATA.ANALYZE | mainstream | A | HW1 | depends |
| DATA-010 | Product data enrichment & classification (PIM, ECLASS, ETIM) | DATA.PREPARE | early-adopter | B | HW2 | minimal |
| DATA-011 | Geospatial analytics & location intelligence | DATA.ANALYZE | early-adopter | B | HW2 | minimal |
| DATA-012 | ESG & sustainability data collection and reporting (CSRD) | DATA.REPORT | early-adopter | A | HW3 | minimal |
| DATA-013 | Process & task mining for process improvement and automation discovery | DATA.ANALYZE | mainstream | A | HW1 | depends |
| DATA-015 | Spend classification & procurement spend analytics | DATA.ANALYZE | mainstream | A | HW2 | minimal |
| DATA-016 | Industrial data contextualisation (OT tag mapping, asset models, unified namespace) | DATA.PREPARE | early-adopter | A | HW2 | minimal |
| DATA-020 | Financial & supervisory regulatory reporting (ESEF/iXBRL tagging, COREP/FINREP) | DATA.REPORT | early-adopter | A | HW3 | minimal |
| DATA-021 | Product carbon footprint, LCA & Digital Product Passport automation | DATA.REPORT | early-adopter | A | HW2 | minimal |
| DATA-022 | Partner data onboarding & schema / EDI mapping | DATA.PREPARE | early-adopter | A | HW2 | minimal |
| DATA-023 | AI-assisted data labelling & annotation (model-in-the-loop) | DATA.PREPARE | mainstream | B | HW2 | minimal |
| DATA-024 | FP&A copilot: driver-based planning, scenarios & variance commentary | DATA.ANALYZE | early-adopter | A | HW3 | minimal |
| DATA-025 | Pay transparency & gender pay-gap analysis (EU Pay Transparency Directive) | DATA.REPORT | early-adopter | A | HW2 | depends |
| DATA-026 | Digital experience analytics: session-replay summaries & UX friction detection | DATA.ANALYZE | mainstream | A | HW2 | minimal |
| DATA-027 | Online experimentation & conversion-rate optimisation (A/B testing, bandits, causal impact) | DATA.ANALYZE | mainstream | A | HW1 | minimal |
AI Act note: all are minimal risk except DATA-009 (depends on model use, e.g. credit/HR scoring), DATA-013 (task mining of employees) and DATA-025 (HR/pay decisions), which can fall into Annex III employment or credit contexts. GDPR personal data is "yes" for DATA-003, 013, 025, 026, 027.
Shared stack
Common building blocks: a governed warehouse/lakehouse with semantic layer (DATA-001/002/024), a metadata catalogue (DATA-007) that feeds every LLM use case, an embedding index (vector DB) for entity resolution, classification and retrieval (DATA-003/010/015/021), a data-quality layer (DATA-008) and an orchestration layer for workflows.
HW1 (CPU only, no GPU)
Postgres or DuckDB + dbt, pgvector, Great Expectations/Soda, Airflow/n8n, scikit-learn/LightGBM/AutoGluon, PM4Py, Splink for entity resolution; LLM via EU-hosted API (e.g. Mistral, Azure EU region) or a 3-4B model through llama.cpp. Covers DATA-003, 008, 009, 013, 027.
HW2 (single 24 GB GPU, 7-14B models)
HW1 plus vLLM or Ollama with a 7-14B instruct model (Qwen/Mistral/Llama class), a small embedding model (bge-m3 / multilingual-e5), Qdrant or pgvector, Docling or Unstructured for OCR/parsing, Playwright/Crawl4AI for extraction, Label Studio plus SAM-class model for labelling, DataHub/OpenMetadata catalogue. Covers DATA-002, 004, 005, 006, 007, 010, 011, 015, 016, 021, 022, 023, 025, 026.
HW3 (multi-GPU / 48-80 GB, 30-70B models)
HW2 plus a 30-70B reasoning/code model (quantised) on vLLM, a reranker, a text-to-SQL agent with a semantic layer (Cube/dbt MetricFlow) and read-only, row-level-secured warehouse roles, sandboxed code execution for spreadsheets/FP&A. Covers DATA-001, 012, 020, 024.
Sharing across use cases and families
- One vLLM endpoint and one embedding service serve all DATA use cases; separate by tenant/API key, not by deployment.
- One vector store with per-use-case collections; the catalogue (DATA-007) and glossary are reused by KNOW-* (RAG) and CODE-* (SQL/code) families.
- Path A (SaaS/EU API) for low-sensitivity analytics; Path B (hybrid) where web, labelling or synthetic data is involved; Path C (open-source self-host) for quality tooling.
- GDPR: pseudonymise before any LLM call for HR/pay (DATA-025), session data (DATA-026) and process mining (DATA-013); keep logs in the EU.