The article details building robust document processing pipelines beyond basic API calls in Azure AI Document Intelligence. It emphasizes that while initial extraction is easy, handling real-world complexities constitutes the core challenge. The proposed pipeline involves ingestion, classification, extraction, routing based on confidence, and posting to a system of record. Choosing the right model, whether prebuilt, custom extraction, or a custom classifier, is crucial for accuracy. The author highlights that deprecated connector actions should be avoided in favor of the Analyze Document for Prebuilt or Custom models (v4.x API). Understanding the difference between accuracy and confidence is vital; confidence scores, which are returned per field, should be used for routing decisions.The system gates documents based on per-field confidence thresholds, with stricter requirements for financially sensitive fields like InvoiceTotal. Arithmetic checks are recommended to catch errors missed by confidence scores. Implementing this logic can be done within Power Automate or an Azure Function. The article also addresses common failure points, including duplicate processing, multi-invoice PDFs, poor line item quality, and currency/locale issues. The key metric for success is the straight-through processing rate, not just model accuracy. Tracking metrics like review rate by reason and reviewer override rate provides insights for improvement. Finally, intelligent document processing uses AI to convert unstructured documents into structured, validated data with confidence scores, enabling automated routing.