Automate document processes with AI - sovereign and GDPR-compliant? WZ-IT builds and operates local AI systems for document processing - model, extraction and filing on your own infrastructure, from one team. See the internal assistant
Capturing invoices, sorting contracts, reading forms - document processes tie up time in every company. AI lifts them to a new level: away from pure text recognition, toward contextual understanding. This article explains the difference between OCR and intelligent document processing, the typical use cases and the sovereign, self-hosted path. As of October 2026.
Table of contents
- From OCR to intelligent document processing
- How LLM-based processing works
- Typical use cases
- Sovereign and self-hosted with Paperless-ngx
- When building it pays off
From OCR to intelligent document processing
Classic OCR (Optical Character Recognition) recognizes the text on a scanned document and converts it into machine-readable characters - no more. It does not know whether a number is an invoice amount or a customer number.
Intelligent document processing (IDP) goes a step further: it interprets the content in context. Instead of just reading characters, it recognizes the document type, understands the meaning of the fields and extracts exactly the data the downstream process needs. The leap from OCR to IDP is the leap from "reading text" to "understanding the document".
How LLM-based processing works
Language models have simplified IDP considerably. Previously every document type needed its own hard-coded template - a new rule for every new invoice structure. An LLM, by contrast, reads a document without a fixed template: it captures the context and extracts the requested fields, even when the layout is unknown.
The classic flow: OCR converts the document into text, the language model classifies it (invoice, contract, form), extracts the relevant fields and passes the result in structured form to the follow-up process. This way even unstructured documents can be processed on which rule-based systems fail.
Today it also works without the OCR step: vision-language models (VLM) read the page directly as an image, capture tables, stamps and layout and output structured text or JSON. Docling implements this with its VLM pipeline, which converts documents end to end with a vision model such as GraniteDocling. Classic OCR remains useful when a searchable PDF is needed or no GPU is available for a vision model. Which models are suitable locally is compared in vision-language models for OCR; how invoices and delivery notes are captured in structured form with self-hosted AI is shown in self-hosted AI document processing.
Typical use cases
The areas of use are everywhere documents feed into processes:
- Invoice and receipt processing - read out amount, date, tax rates and supplier and pass them to accounting.
- Contract and form analysis - recognize deadlines, parties and key clauses.
- Automatic tagging and filing - classify documents and file them in the right place.
- Incoming mail - sort incoming correspondence and assign it to the right case.
The common denominator: less manual capture, fewer media breaks, fewer errors.
Sovereign and self-hosted with Paperless-ngx
Precisely because documents are often sensitive - contracts, personnel files, client or patient records - the complete data path needs review. Cloud services are not categorically excluded for professional secrecy holders, but the service, configuration, participating people, provider chain and technical access must fit the intended use.
A sovereign path combines a self-hosted document solution with a local language model. Since version 3.0.0, the open-source Paperless-ngx ships its own AI features: suggestions for title, tags, correspondent and document type, a document chat and a workflow action that applies suggestions automatically. The backend is a local model via Ollama or an OpenAI-compatible server (configuration, AI section). With a local backend, processing stays inside the controlled environment. The setup is covered in the guide Paperless-ngx AI setup.
If you need custom prompts per document type or filling custom fields, the actively maintained paperless-gpt (MIT licence) offers both. The former add-on Paperless-AI has not been maintained since 31 March 2026 according to its repository and is no longer suitable for new installations. Which LLM to self-host? covers model selection.
If the goal is not just filing but reading a specific document class such as incoming invoices with fixed target fields and validation rules, WZ-IT implements this as AI document processing on your own infrastructure.
When building it pays off
AI document processing pays off as soon as many documents regularly run through a process and manual capture noticeably ties up time. The more uniform the document stream and the clearer the fields to extract, the faster the benefit.
For sensitive records the compliance dimension is added: here self-hosted operation is not only more efficient but often the only permissible variant. How such a system is conceived as an end-to-end knowledge system is shown in What is RAG? - document processing and knowledge retrieval mesh here.
