Building an Invoice Extraction Pipeline with Azure AI Document Intelligence and Power Automate

This article walks through building an advanced invoice extraction pipeline using Azure AI Document Intelligence and Power Automate, focusing on the 80% of work that goes beyond simply receiving JSON output from the API. It covers model selection, confidence-based routing, and failure modes that can break document pipelines in production.

The pipeline consists of five stages: Ingest, Classify, Extract, Route, and Post. The hardest part is dealing with documents that have unique layouts or handwritten information, such as scanned faxes containing purchase orders. To address this, the article recommends starting with the prebuilt invoice model and only building custom models if specific fields are consistently missed.

It also highlights that the Analyze Invoice action is deprecated, and the recommended actions are Analyze Document for Prebuilt or Custom models (v4.x API) and Classify document with document classifier (v4.x API). Importantly, the article emphasizes the importance of reading confidence scores properly, as they represent the probability that a specific extracted value was detected correctly and are not the same as accuracy scores.

For financial records like invoices, Microsoft suggests targeting a confidence score of 80% or higher, with sensitive cases like medical records requiring a score close to 100%. The article provides an example response showing document-level confidence and field-level confidence, stressing the need to gate on per-field thresholds rather than a blanket confidence score to avoid automating potentially incorrect payments.