AI document processing: automate PDFs, invoices and emails

Start with one recurring task

Useful starting points include incoming invoices with defined mandatory fields, orders with recurring line items, or delivery notes that can be matched to an order. One document class, one input channel and one target system are easier to evaluate than an entire inbox.

At low volumes with short manual handling times, an existing accounting feature or a simple import may be more economical. Contracts requiring interpretation and documents used for consequential decisions need different controls from capturing an invoice date.

Choose parsers, OCR and vision models for the input

OCR recognises text in images. A vision-language model combines image and language information and can associate fields with their visual context. A parser reads a file format’s structure. These methods can be combined; specialised OCR models may themselves be VLMs.

A digital PDF with a text layer does not automatically need vision. Its text layer can still have broken reading order or unusable tables. Check the output before extracting business fields. For structured invoice data, format validation is usually a better first step than recognising the same values again.

Classification and extraction produce a proposal

Classification assigns a document to a predefined class. Extraction maps relevant information to a fixed schema. Preserve the original value, normalised value and source location for critical fields. Unknown classes and unreadable values need explicit states.

Valid JSON does not prove that an invoice is correct. Business validation checks required fields, amounts, currencies, rounding, tax lines and approved master data. Consistent sums cannot reliably detect numbers that were all extracted incorrectly. Tax treatment must match the actual case and may require expert review.

For example, an invoice shows a gross total of EUR 120.00 and EUR 80.00 outstanding after a partial payment. Using EUR 80.00 as the invoice total is formally plausible but wrong. The field contract must distinguish gross total from outstanding balance.

Make uncertainty a visible review path

Human-in-the-loop means a defined technical review step. Show the original beside the proposal, mark missing or conflicting fields and explain why review is required. Corrections and approval apply to a specific revision; a later change requires renewed approval.

Confidence scores are signals rather than guarantees. A model’s self-reported certainty is not a calibrated error probability. Test thresholds against your own reference cases, measuring missed errors and unnecessary escalations. Reviewing every proposal is a useful starting point for a first pilot.

New business partners, changed bank details, uncertain currencies and business discrepancies require separate checks. Automatic payments are a distinct and more consequential scope, not an implicit part of extraction.

Connect ERP, CRM and DMS through controlled interfaces

The target system remains authoritative for its master data and business rules. An adapter maps approved document data to its documented API or import contract. A JSON format alone is not a completed integration. Check available interfaces, write permissions, rate limits and error behaviour before building extraction.

A timeout may occur after the target has saved a record. Persistent workflow IDs, idempotency or reconciliation through a unique external reference prevent blind repeat writes. An ambiguous transfer needs its own reconciliation state. Business duplicates also require tenant, issuer and invoice-number checks.

A DMS manages documents and versions, an ERP processes business transactions, and RAG supports knowledge questions with sources. Each goal has different quality criteria. A correct document answer does not prove a correct ERP posting.

Define actual data flows and operating responsibilities

Documents may contain personal and confidential information. Map the entire data flow: inbox, storage, parser, model provider, review UI, target system, backups and logs. EU hosting or local inference alone does not establish GDPR compliance. Contracts, purpose, access, retention and deletion must match actual processing.

Document content must not become system instructions. Limit file sizes and formats, isolate parsers, handle unknown inputs safely and keep ERP credentials outside the model. Review and retries also need tenant and permission checks.

Compare Document AI costs per completed task

Total cost includes setup, integration, software or API usage, storage, operations, evaluation and human review. Depending on the service, billing may be per page, document, token or capacity. Check the actual tariff: OCR and complete field extraction are different services.

Use cost per completed task and active review time to decide. Include retries, difficult documents and new layouts. Released time becomes financial savings only when it reduces actual cost or creates economically useful capacity.

Illustrative calculation, not a measurement: at 1,000 documents and EUR 40 internal hourly cost, each remaining minute of review costs about EUR 667 per month. Cheap extraction can therefore be more expensive than a method needing fewer corrections. The German technical guide includes a full, explicitly hypothetical pilot calculation.

Buy, integrate or build

First check whether your accounting, ERP or DMS system already handles the task adequately. Then compare a specialised service with a narrow custom workflow using the same cases. Building a foundation model is normally unnecessary for this starting point.

Assess your document process

Record the following in a short process note. It lets you compare existing features or products before commissioning custom work. Do not send confidential original documents to providers without checking how they will be handled.