Automated Processing of Complex Order Documents

Incoming orders came from multiple sources and followed no consistent layout. Although the content was similar, structural differences meant that rule based extraction approaches quickly reached their limits. Relevant data therefore had to be transferred manually into the target system.

Order documents and coordination were largely handled via email, resulting in limited transparency and high coordination effort. As volumes increased, it became clear that this approach was not sustainable in the long term.

• Different document layouts
• Manual data transfer into the target system
• High error rates
• Coordination and follow up via email
• Limited scalability as volumes increase

The goal of the project was to establish an automated process that reliably identifies relevant information from differently structured order documents and transfers it into existing target systems. Manual effort was to be significantly reduced while maintaining human oversight.

Key requirements included high recognition accuracy, reliable handling of varying formats, and an architecture that supports future adaptations. In addition, the process was intended to become more transparent and improve collaboration.

An analysis of existing processes showed that fixed rule based extraction was not sufficient. Instead, an approach was chosen that interprets document content and adapts flexibly to different layouts.

As a first step, PDF files are converted into images to make visual structure (tables, stamps, handwritten notes, unusual layouts) reliably accessible for downstream processing. The content is then analyzed using large language models (LLMs) to understand both text and context beyond simple pattern matching.

First, the document type and relevant metadata are identified and evaluated to determine whether further processing is required (e.g., routing to the correct extraction template, applying document-specific checks, or skipping unsupported inputs). 

Based on this classification, relevant data is extracted in multiple stages: initial capture of candidate fields, refinement using context across pages, and normalization into the target structure.

To ensure quality, extracted results are logically validated (completeness checks, cross-field consistency, plausible value ranges) and additionally reviewed by a second model. This significantly reduces incorrect interpretations.

Finally, in ambiguous or low-confidence cases, the process falls back to targeted human involvement, focusing review effort only where automated validation and model agreement are not sufficient.

The project started with a proof of concept to validate the approach using real documents. Early on, it became clear that most documents could be reliably identified and processed.

Implementation followed an iterative approach with short cycles and close alignment, allowing adjustments to be made early.

Computationally intensive processing steps were offloaded to asynchronous processing. Model changes and extensions could be implemented without major restructuring.

In parallel, surrounding workflows were optimized. Orders are automatically assigned, centrally managed, and coordinated directly within the system.

The automated process significantly reduced manual effort. Employees no longer need to fully enter orders manually and instead focus on reviewing the automatically extracted results.

Error rates decreased noticeably while more orders could be processed within the required timeframes. Overall transparency improved, as documents, data, and communication are now centrally available.

In addition, employee satisfaction increased due to the reduction of repetitive tasks and a stronger focus on professional review and operational control.