AI Delivery Note Processing System
Intelligent document extraction with self-trained AI model and human-in-the-loop.
About the Project
A company received delivery notes in highly varying formats – all as PDF, but with very different layouts and quality levels. The previous extraction algorithm was error-prone and required extensive manual post-processing that was extremely time-consuming.
The Solution
The developed solution consists of several components: An Azure Logic App workflow reads delivery notes from a dedicated Outlook mailbox. For extraction, a custom AI model was trained, specifically optimized for the heterogeneous formats and qualities of customer delivery notes.
Training Portal
A central element is the training portal based on Azure AI Document Intelligence Studio. Here, the customer can independently upload and annotate delivery notes to continuously improve the model. This makes the previously lengthy adaptation process significantly more efficient – new formats are quickly understood by the model.
Human-in-the-Loop
For human-in-the-loop validation, a WebApp was developed (React TypeScript + FastAPI backend). The interface displays extracted data next to the PDF document, with direct linking of recognized fields to the corresponding positions in the document. Employees can quickly review extraction results and correct them if necessary before the data is booked into the ERP system.
ERP Integration
Final booking is done automatically via the ERP system's REST API through Azure Function.
Highlights
- Self-service training portal for continuous model improvement
- Custom AI model optimized for heterogeneous document formats
- Human-in-the-loop WebApp with document view and field linking
- End-to-end workflow from email inbox to ERP booking
Challenges
Training a robust model for highly varying document layouts. Intuitive UI for quick validation with document context.
Results
Drastic reduction of manual post-processing time. Customer empowerment through self-service model adaptation. Higher extraction accuracy than previous algorithm.
Interested?
Are you interested in a similar project or want to learn more? Get in touch – we look forward to hearing from you!
Get in touchRelated Projects
unified-ui – Open Source AI Integration Platform
Universal integration platform for AI agents – one interface for Foundry, N8N, LangGraph and custom solutions with central tracing and Enterprise RBAC.
Azure AI Landing Zone Framework
Enterprise framework for declarative provisioning of AI infrastructure on Azure – from Microsoft Foundry to AI Search to private networks.