# AI Document Processing

Author: Marin Niehues, Niehues Consulting · Published: 2025-06-09 · Updated: 2026-11-16 · AI Adoption & Operations
Source: https://www.niehuesconsulting.com/blog/ai-document-processing/

AI document processing, also known as intelligent document processing (IDP), reads incoming documents, validates the data and passes it to the core system without manual keying. A document that runs through without manual intervention is called straight-through processing. The effort sits elsewhere than expected: algorithm and technology are the smaller part, process and handovers the larger. This article describes a delivery-oriented approach that lets executive and divisional leadership steer selection and operation.

This article provides an overview and does not constitute legal advice.

**The value of document processing arises along the chain from the inbox to the posting in the core system. The model's extraction quality is one part of that chain, the design of the handovers the larger one.**

## What is the difference between OCR and AI document processing?

OCR converts an image into text, while AI document processing understands the document and delivers structured, validated data for the next step. Classic optical character recognition recognizes characters but does not know which figure is the invoice total and which the order number. That used to require templates and rules for each layout.

IDP adds four capabilities to text recognition. It classifies the document, extracts the relevant fields, checks them against rules and master data, and hands them to the core system. Modern language and vision models also handle layouts they have never seen before. For clearly legible digital documents they deliver the fields with high accuracy. For handwriting and poor scans, error rates are markedly higher, as they are for multi-page tables. Measure accuracy per field on a sample of your own documents before relying on a vendor's claim.

## Where document processing pays off

Four criteria, assessed before any technology selection, determine whether a document stream is suited to AI.

- Volume: how many documents pass through the process each month, and how much working time does manual handling tie up?
- Variability: how much do layout and language differ? A stream of a few formats behaves differently from one with hundreds of supplier templates.
- Cost of errors: what does a misread figure cost before it is noticed?
- Lead time: which business transaction is waiting for the document, such as a payment or an approval?

Documents with high volume and medium variability deliver the fastest proof. Lasting candidates are delivery notes and freight documents, which no e-invoice replaces. Contracts require more review effort because error costs and room for interpretation rise. The overall frame for rollout is set by the article [From pilot to scale](/blog/from-pilot-to-scale-ai-operations/).

## Do you still need AI once e-invoicing is mandatory?

For domestic B2B incoming invoices, the need for AI extraction falls once e-invoices arrive across the board, while for all other documents it remains. An e-invoice in structured formats such as XRechnung or ZUGFeRD already contains the data in machine-readable form and flows into the system without extraction. Both formats implement the European standard EN 16931, which defines the semantic data model of an electronic invoice. In Germany it is becoming mandatory in stages for domestic B2B business, and paper and PDF invoices continue to arrive during the transition.

Three rules follow for the investment decision. Do not build a new AI solution solely for domestic incoming invoices. Use it for the transition only if it is being built for other documents anyway. And plan AI where e-invoicing does not apply: invoices from foreign suppliers, small-amount invoices and receipts, and above all delivery notes, order confirmations and correspondence. Structured e-invoices still need matching against purchase order and goods receipt, and rules and models help resolve discrepancies there.

## The document chain in four stages

Every document process passes through the same four stages, and each stage involves its own decisions.

- Intake: through which channels do documents arrive, and who is responsible for their completeness? Several intake channels are merged before a model sees the documents.
- Recognition: the model classifies the document and extracts the relevant fields. Each field receives a confidence score. With language models it is derived from plausibility rules, matching against master and order data and a second extraction, because the model's self-assessment alone does not provide a reliable threshold.
- Review: documents below a defined confidence threshold go to a specialist. The threshold is a management decision because it trades speed against error risk.
- Handover: the reviewed data reaches the core system and the next step starts. Interfaces to ERP and archive determine effort and stability here.

Most delays arise in the third and fourth stages. A review without clear ownership piles documents up in a queue. A handover without a stable interface merely shifts manual work from reading to re-keying, and the time saved in recognition is lost.

**Setting the confidence threshold**
The threshold determines what share of documents continues automatically. A high threshold lowers error risk and increases manual effort, while a low threshold does the opposite. Set it per document type and per field, because a wrong amount costs more than a misspelt place name. The head of the business unit decides, IT implements.

## How much does AI document processing cost per document?

At the list prices of major cloud services for pre-trained invoice models, extraction alone costs in the region of one cent per page, but the total cost per document is driven by review and integration. Whoever compares only the page price systematically underestimates the cost of a solution. Specialized IDP platforms usually charge through volume bundles or licenses, plus implementation services.

**Calculation path: cost per document**
1. Extraction cost: the model's page price or license cost, divided by monthly volume.
2. Review cost: share of documents below the confidence threshold times handling time per exception times the specialist's hourly rate.
3. Operating cost: monitoring, maintenance of rules and interfaces, allocated per document.
4. Implementation cost: prototype, integration and enablement, spread over the planned period of use.
5. Comparison: the sum of 1 to 4 against today's cost per document from handling time times hourly rate, plus the cost of errors and late payments.

The lever lies in step 2. Every percentage point more straight-through processing directly lowers review cost, so work on master data and rules usually pays more than switching to a cheaper model. Whether to buy a platform or build on a model yourself is covered in the article on [make or buy](/blog/build-buy-or-adapt-ai/), and vendor selection in the article on [selecting an AI vendor](/blog/choosing-an-ai-provider/).

## Steering quality with four metrics

An AI document process is steered through metrics like any other process. Four form the core:

- Straight-through rate: share of documents that reach the core system without manual intervention.
- Error rate after review: share of incorrect values that land in the core system despite review. Samples determine it.
- Lead time per document: time from intake to posting.
- Cost per document: total effort for operation and review, divided by the number of documents processed.

The straight-through rate rises only in operation, because exceptions from practice improve the model and rules. Plan a fixed rhythm in which error cases are analyzed and rules adjusted.

## Clarify data protection and confidentiality early

Documents often contain personal data and confidential contract content. This raises questions about storage location and access rights as well as the contract with the model provider. They belong at the start because they shape the solution: whether a model runs in a European cloud or on your own infrastructure changes the effort considerably. The article on [data protection and AI](/blog/gdpr-compliant-use-of-language-models/) sets out the questions.

## Responsibilities for steady-state operation

In operation, responsibilities decide success. A proven set-up comprises three roles:

- Process owners in the business unit: they own metrics and exception rules.
- Review specialists: they handle exceptions and report recurring error patterns.
- Technical operations: responsible for model and interfaces and for monitoring their stability.

If internal expertise for model operation and interfaces is lacking, outsource technical operations to an external team and keep process ownership in the business unit. Decide this before rollout so that every role is filled on handover day.

## Niehues Consulting as a partner for document processing

[Niehues Consulting](/) supports document processing according to the Niehues Delivery Model with its phases understand, build and introduce. Marin Niehues steers the initiative as delivery manager or product owner. For document AI, three of the model's review questions matter most. What is the baseline, and how do we measure success? Without today's cost and lead time per document there is no comparison. Does the prototype prove the benefit in the real workflow with real data? That means your documents, including every poor scan. Who owns operation from day one? The review stage and thresholds need named ownership.

**Approaches at a glance**
Criterion | Classic OCR with rule sets | Language model without process design | With Niehues Consulting
Handling new layouts | New rules per template | High flexibility with fluctuating accuracy | Model selection and confidence steering per document type
Exception handling | Manual and unregulated | Randomly distributed | Defined review stage with named ownership
Operation | Maintenance effort with every change | No monitoring | Metrics and a fixed improvement rhythm
Acceptance in the business unit | Low as soon as errors occur | Low without involvement | Business unit co-designs review and thresholds

The entry point is the free initial consultation of 30 minutes. You arrange it via the [contact page](/#kontakt) and clarify which document stream suits a first prototype.

**Introducing document processing deliberately**
Start with a document stream that combines high volume with manageable variability and that e-invoicing will not replace. Set the confidence threshold and review ownership before the prototype. Measure straight-through rate and error rate from day one. Extension to further document types then follows on the basis of figures from your own operation.

**Sources**
- German Federal Legislature: Wachstumschancengesetz, amending section 14 UStG on e-invoicing, statute, 2024
- German Federal Ministry of Finance: Einführung der obligatorischen elektronischen Rechnung bei Umsätzen zwischen inländischen Unternehmern, ministry circular, 2024
- CEN: EN 16931-1, Electronic invoicing, semantic data model, standard, 2017
- Xu, Li, Cui, Huang, Wei, Zhou: LayoutLM: Pre-training of Text and Layout for Document Image Understanding, KDD, 2020
- European Parliament and Council: General Data Protection Regulation (EU) 2016/679, regulation, 2016

## Frequently asked questions

### What is AI document processing?

AI document processing is the automatic classification and extraction of documents with language and vision models, combined with validation rules and handover to core systems. It replaces rigid template rules with models that also handle unfamiliar layouts.

### Which documents are suitable first?

Documents with high volume and manageable variability are best suited, such as delivery notes or freight documents. Domestic incoming invoices are suitable only for the transition until e-invoicing is universal. Contracts follow later because review effort and error costs are higher.

### How accurate is AI at reading invoices?

For clearly legible digital invoices, today's models achieve high field accuracy, while for handwriting or poor scans it falls markedly. What matters is accuracy per field on your own documents. Measure it in the prototype on a sample with known target values.

### Is AI document processing GDPR-compliant?

Yes, with a data processing agreement, a defined processing location and access rights limited to what is necessary. Have your data protection function assess the specific set-up before production data reaches a model.

### How long does implementation take?

A prototype in the real workflow is ready in six to eight weeks, and steady-state operation for one document stream follows after three to six months. The range depends on document variability and the number of interfaces to core systems.

### What is straight-through processing?

Straight-through processing describes documents that pass from intake to posting without manual intervention. Its share, the straight-through rate, is the most important steering metric of document processing because it directly determines review costs.
