Many SMBs keep decades of customer files, drawings, quality records and contracts on paper. An AI assistant cannot search what it cannot read. A one-time digitisation project (scan, OCR, split into documents, add a few index fields) turns the cellar into a searchable corpus. The AI part is small; the real work is selecting, sorting, scanning and checking, and the result feeds every later search and RAG use case.
Problem & value
- Today: knowledge sits in folders in a cellar or archive room. Finding an old customer contract, a calibration record or a drawing means asking the one colleague who knows the shelf. Some paper is also kept only because nobody dared to throw it away.
- AI value: once text is machine-readable, full-text and semantic search work across the whole history, and assistants can answer questions from it (PREP-013, DOC-014). OCR and layout models also split stacks of pages into single documents and suggest type, date and reference.
- Honest scope: only scan paper still needed; part of the archive should be destroyed. No benefit figure is claimed; measure your own search time.
Variants
- Customer, project, quality and maintenance files; technical drawings (wide-format scanner or service provider); contracts (high sensitivity; HR files usually stay out of scope).
- Public sector: registry and building files under archive law.
- Out of scope: ongoing incoming mail (DOC-008) and historical handwriting (DOC-005).
How it works
- select & sort paper (retention check)
- scan (duplex, 300 dpi, separator sheets)
- clean-up (deskew)
- OCR + layout parsing
- split & classify (type, date, reference)
- index fields
- human sample check
- DMS / file share
- search and RAG
- OCR: Tesseract via OCRmyPDF (inside Paperless-ngx) handles clean print; Docling or a vision-language model helps with poor scans and tables. Handwriting stays weak (DOC-005).
- Splitting: barcode or blank-page separator sheets are reliable; an LLM can suggest title, type and date.
- Human-in-the-loop: sample checks per batch, review of low-confidence fields, decision on destruction.
Data preparation
Which data: the paper itself plus existing lists (folder registers, customer numbers). It lives in archive rooms and cellars, sometimes as half-scanned PDFs without text layer.
Typical gaps: mixed formats (A4, A3, drawings), staples, fading thermal paper, no folder-to-customer mapping, duplicates, personal data without a retention decision, and no owner who says what must be kept.
Preparation steps: 1. Inventory the archive: shelf metres, types, age, sensitivity (PREP-001). 2. Apply retention rules first: what must be kept (tax and commercial law, product liability), what may be destroyed, what has no value. Involve the DPO for personal data (PREP-025). 3. Decide the target structure and a handful of index fields: type, date, customer or project number, retention end (PREP-009, PREP-010). 4. Pilot with 2-3 folders to fix scan settings and separator method. 5. Abgabenordnung (AO) section 147, retention and replacing scanning - https://www.gesetze-im-internet.de/ao_1977/__147.html 6. Run OCR, sample-check 5 percent of pages per batch, re-scan failures. 7. Set access rights before indexing for AI so that assistants do not expose contract or personnel data to everyone (PREP-023).
Tools: the document scanner with feeder you probably own, Paperless-ngx, OCRmyPDF, Nextcloud, SharePoint, DATEV Dokumentenablage or your DMS; a scan service for drawings and bulk.
Rough share of effort: high, about 70-80 percent of the project is sorting, scanning and checking, not AI (our estimate). Plan days of staff time per few dozen shelf metres plus one to two days of IT set-up. See GUIDE-003.
Implementation paths
Which path for whom
- SMB - Path C: the data is customer and contract files with personal and trade-secret content, and Paperless-ngx with OCRmyPDF runs on a small server (HW1) that also serves later use cases (Baukasten effect: PREP-010, DOC-008, DOC-014). Path A (scan service or DMS) is a valid alternative if no one wants to run a server.
- Enterprise - Path B: high volume, scanning usually outsourced, OCR and classification via cloud document APIs integrated with the ECM.
- Public - Path C: sensitive registry files, sovereign hosting, archive law.
Path A - Buy (off-the-shelf)
- Scan service providers in Germany and the EU (typically ISO 27001-certified, with data-processing agreement): good for large volumes, drawings and one-off peaks. Mark: pick EU-based providers and sign an AVV (Art. 28 GDPR).
- DMS with built-in OCR: DocuWare, d.velop documents, ELO, SharePoint, DATEV Dokumentenablage (EU-hosted options exist at DocuWare, d.velop and ELO).
- Choose A when: you want no own operations and a standard archive process.
Path B - Build on cloud APIs
- Document parsing APIs: Azure AI Document Intelligence (EU regions), Google Document AI (EU location), AWS Textract (Frankfurt), Mistral OCR (EU provider).
- Orchestration: n8n or Power Automate watching a scan folder, calling OCR and an LLM for type and date, writing to SharePoint, Nextcloud or the DMS.
- Choose B when: layouts are varied, volume is large, and an AVV with EU data residency is in place.
Path C - Self-host (open source)
- Stack: Paperless-ngx (consume folder, OCRmyPDF/Tesseract, tags, correspondents, document types), Nextcloud or file share for storage; optional Docling or PaddleOCR for hard layouts; a small local LLM via Ollama for suggested titles and types.
- Hardware: HW1 (CPU server, 8+ GB RAM) is enough for OCR and Paperless-ngx; HW2 only if you want a local VLM for difficult scans.
- Effort: one to two days set-up, a few days for tag and rule tuning, then ongoing backups and updates by IT.
- Choose C when: data is sensitive and a simple, shared archive platform is wanted.
Path D - Custom training
Not needed; pre-trained OCR and layout models cover this case.
State of the art (2026)
- Open OCR is mature for print; layout-aware parsers (Docling, PaddleOCR-VL) improve tables and mixed layouts.
- LLMs can propose titles, types and dates, but they hallucinate plausible values: keep a check on dates and reference numbers.
- Limits: handwriting, stamps, faded thermal paper, skewed or low-resolution scans and CAD drawings stay error-prone. OCR errors silently hurt retrieval, so test with real questions (PREP-018).
- Legal archiving: a searchable scan is not automatically a replacement for the original. German tax rules (GoBD) allow replacing scanning if a documented process (Verfahrensdokumentation) exists; keep the originals until that is settled with your tax adviser.
Shared resources & combinations
- Same stack (ocr-docparse + workflow-automation): DOC-001, DOC-008, DOC-014, PREP-007; one Paperless-ngx or DMS instance serves back-file scanning and ongoing intake.
- Downstream: PREP-009, PREP-010, PREP-013, PREP-023.
Effort, cost & KPIs
- Path A: weeks; cost per page or shelf metre, quote several providers. Path B: 2-4 weeks plus API fees. Path C: 1-2 days set-up; staff scanning time dominates.
- KPIs: pages per day, share passing sample check, share with complete index fields, search time before/after.
People & change
Involve early: management (scope, budget), the department head (content owner), IT (server, backup), data owners (sort and retention), the tax adviser (GoBD) and the DPO. A DPIA is usually not needed for plain scanning; an AVV with scan providers is (GUIDE-005, GUIDE-010).
Works council: usually no co-determination for scanning itself, since it does not monitor employees. It changes if HR files are included or if the DMS logs individual user activity; then §87(1) no. 6 BetrVG applies and the council should be informed early. §90(1) no. 3 applies when work processes are planned with AI (GUIDE-004). Public bodies involve the Personalrat.
AI literacy: Art. 4 AI Act requires measures that support AI literacy; a short briefing on OCR errors and checking AI-suggested index fields is enough (GUIDE-006).
Communication and pilot: veterans know the archive best, so make them the sort team. Start with 2-3 folders and name an AI champion (GUIDE-008, GUIDE-009). Typical concerns: redundancy of paper knowledge holders, wrong destruction, legal value of scans. People decide what is kept; AI only suggests index values.
Risks, regulation & compliance
- EU AI Act: minimal risk. OCR and indexing are no Annex III use. Art. 4 AI literacy applies.
- GDPR: archives hold personal data of customers, employees and suppliers. Check legal basis and purpose when reusing old files for AI (PREP-025), delete data past retention instead of digitising it, sign an AVV with scan service providers and cloud OCR vendors (Art. 28), and restrict access.
- Bookkeeping: GoBD and retention periods; get tax-adviser sign-off before destroying originals.
- Risks: missing pages, mis-indexed documents, oversharing through AI search (PREP-023), loss of originals. Mitigate with batch protocols, sample checks and access clean-up.
Getting started
- Walk the archive with the department head; estimate shelf metres, types, sensitivity.
- Agree retention and destruction rules with tax adviser and DPO.
- Set up Paperless-ngx or your DMS; pilot 2-3 folders and measure OCR quality.
- Scan in batches with separator sheets; sample-check.
- Fix access rights, then test search with 20 real questions.
Sources
- Paperless-ngx documentation - https://docs.paperless-ngx.com/
- OCRmyPDF documentation - https://ocrmypdf.readthedocs.io/
- Docling project - https://github.com/docling-project/docling
- Tesseract OCR - https://github.com/tesseract-ocr/tesseract
- Abgabenordnung (AO) section 147, retention and replacing scanning - https://www.gesetze-im-internet.de/ao_1977/__147.html
- Regulation (EU) 2024/1689 (AI Act) - https://eur-lex.europa.eu/eli/reg/2024/1689/oj
- PaddleOCR - https://github.com/PaddlePaddle/PaddleOCR