Document Intelligence

Intro & scope

Use cases that turn documents (PDF, scan, photo, Office, e-mail) into data, understanding or cleaned output. Patterns: DOC.CAPTURE (extract fields from recurring document types, e.g. invoices, IDs, forms), DOC.UNDERSTAND (read, compare and reason over long documents, e.g. contracts, tenders, drawings, rules checks) and DOC.TRANSFORM (convert, redact, verify or remediate documents).

Scope boundaries: - DOC-014 is the generic parsing layer (layout, tables, Markdown/JSON for RAG); the capture cases build on it and add domain schemas and validation. - DOC-008 routes and classifies before extraction; DOC-001..006, 017, 023, 024 extract after routing. - DOC-009 (portfolio extraction and comparison) vs DOC-010 (single-contract review, drafting, redlining) vs DOC-011 (data-room scale) vs DOC-015 (tenders) vs DOC-022 (LC presentations). - DOC-016 (rulebook checks on dossiers) vs DOC-027 (label proofing) vs DOC-019 (certificate verification). - DOC-018 (redaction) is a cross-cutting step before sharing or AI processing; DOC-021 uses it for DSAR and eDiscovery. - DOC-020 (courts) is the only family member with high AI Act risk (Annex III administration of justice), path C. - Gaps in numbering (DOC-025, 028, 029) are intentionally unused.

Overview

ID	Title	Pattern	Maturity	Path	HW tier	AI Act risk
DOC-001	Invoice processing & accounts-payable automation	DOC.CAPTURE	mainstream	A	HW2	minimal
DOC-002	Receipt & expense capture	DOC.CAPTURE	mainstream	A	HW2	minimal
DOC-003	Form & application processing (claims, loans, public forms)	DOC.CAPTURE	mainstream	A	HW2	depends
DOC-004	ID document data extraction for onboarding	DOC.CAPTURE	mainstream	A	HW2	depends
DOC-005	Handwriting recognition & historical document transcription	DOC.CAPTURE	early-adopter	A	HW2	minimal
DOC-006	Logistics documents: delivery notes, bills of lading, customs	DOC.CAPTURE	mainstream	A	HW2	minimal
DOC-007	Medical record & lab report extraction	DOC.CAPTURE	early-adopter	A	HW3	depends
DOC-008	Document classification & mailroom / input management	DOC.CAPTURE	mainstream	A	HW2	minimal
DOC-009	Contract analysis & clause extraction	DOC.UNDERSTAND	mainstream	A	HW3	minimal
DOC-010	Contract review, drafting & redlining assistant	DOC.UNDERSTAND	mainstream	A	HW3	minimal
DOC-011	Due diligence & data-room review	DOC.UNDERSTAND	early-adopter	A	HW3	minimal
DOC-012	Financial statement & annual report analysis (spreading)	DOC.UNDERSTAND	early-adopter	A	HW3	depends
DOC-013	Technical drawing & engineering document understanding (P&ID, CAD drawings)	DOC.UNDERSTAND	early-adopter	A	HW3	minimal
DOC-014	PDF-to-structured data conversion & table extraction (document parsing for RAG)	DOC.TRANSFORM	mainstream	B	HW2	minimal
DOC-015	Tender / RFQ document analysis & requirements extraction	DOC.UNDERSTAND	early-adopter	A	HW3	minimal
DOC-016	Automated compliance checking of documents against rules (e.g. building permits)	DOC.UNDERSTAND	early-adopter	B	HW3	depends
DOC-017	Bank statement & payslip analysis for lending	DOC.CAPTURE	mainstream	A	HW2	depends
DOC-018	Document anonymisation & redaction (GDPR)	DOC.TRANSFORM	mainstream	B	HW2	minimal
DOC-019	Certificate & quality document verification (material certificates, CoC)	DOC.TRANSFORM	early-adopter	B	HW2	minimal
DOC-020	Judicial mass-proceedings assistant (court-side case-file processing)	DOC.UNDERSTAND	early-adopter	C	HW3	high
DOC-021	eDiscovery & technology-assisted review (litigation, investigations, DSAR)	DOC.UNDERSTAND	early-adopter	A	HW3	depends
DOC-022	Trade finance document examination (letters of credit, UCP 600 / ISBP checks)	DOC.UNDERSTAND	early-adopter	A	HW3	minimal
DOC-023	Safety data sheet (SDS) extraction & hazardous-substance register maintenance (REACH/CLP)	DOC.CAPTURE	early-adopter	A	HW2	minimal
DOC-024	Commercial insurance submission intake for underwriting (broker submissions, SOV, loss runs)	DOC.CAPTURE	early-adopter	A	HW3	minimal
DOC-026	Accessible document remediation (PDF/UA tagging, alt text, EAA/BITV)	DOC.TRANSFORM	early-adopter	A	HW2	minimal
DOC-027	Packaging artwork & label compliance proofing (food, pharma, cosmetics)	DOC.UNDERSTAND	early-adopter	A	HW2	minimal
DOC-030	Construction quantity takeoff & cost estimation from plans	DOC.UNDERSTAND	early-adopter	A	HW3	minimal
DOC-031	Grant & research proposal screening and reviewer assignment (funder side)	DOC.UNDERSTAND	early-adopter	B	HW2	depends

AI Act note: "depends" means the use case is minimal risk as a pure extraction aid but becomes Annex III (creditworthiness, essential services, public benefits, justice, employment) when the output feeds decisions about natural persons; keep a human decision step and document intended purpose.

Shared stack

Common layers for all DOC use cases (and reusable by KNOW, LANG, AGENT, ANOM): one document-parsing service (DOC-014 pipeline: Docling / Unstructured / Marker, Tesseract or PaddleOCR, VLM fallback for hard pages), one OpenAI-compatible inference gateway (vLLM or llama.cpp behind LiteLLM), one embedding + vector index per corpus (bge-m3 or multilingual-e5, pgvector or Qdrant), a PII/redaction and guardrail layer (DOC-018 / Presidio), a workflow engine (n8n, Temporal or Camunda) with human-review queue and audit log, and schema validation (JSON schema, Pydantic) with source-citation (page and bounding box) on every extracted value.

HW1 (CPU/laptop): Tesseract/PaddleOCR, Docling, 3-8B quantised model via llama.cpp/Ollama, pgvector, n8n. Enough for receipts, simple invoices and classification pilots.
HW2 (single 24-48 GB GPU): 7-32B instruct model plus a 7-32B VLM (Qwen2.5-VL / Mistral Small class) via vLLM, bge-m3 + reranker, Qdrant or pgvector, Presidio, n8n/Temporal. Covers invoices, forms, IDs, handwriting, logistics, SDS, accessibility remediation, anonymisation, label proofing, grant screening (DOC-001..006, 008, 014, 017, 018, 019, 023, 026, 027, 031).
HW3 (multi-GPU 80 GB+ or EU-hosted API with DPA): 70B-class or reasoning model, large VLM for drawings, hybrid search (OpenSearch + reranker), knowledge graph (DOC-013), agent framework with approval gates, per-matter tenant isolation. Needed for long-context contract, data-room, tender, eDiscovery, trade-finance, medical, underwriting and court use cases (DOC-007, 009-013, 015, 016, 020-022, 024, 030).
Path guidance: A = managed EU API or SaaS first; B = self-host/build when data sensitivity or volume requires it; C = only with public-body governance, DPIA and conformity assessment. Share one gateway, one parsing service, one vector index per corpus and one guardrail layer across use cases; keep HW3 workloads (legal, medical, court) in separate tenants with access control.