Azure AI Document Intelligence: Extracting Structured Data from Invoices

Azure AI Document Intelligence transforms unstructured documents into actionable data. The prebuilt invoice model extracts key fields with remarkable accuracy, eliminating manual data entry in accounts payable workflows.

Setting Up Document Intelligence

First, provision an Azure AI Document Intelligence resource and configure your client.

from azure.ai.documentintelligence import DocumentIntelligenceClient
from azure.ai.documentintelligence.models import AnalyzeResult
from azure.core.credentials import AzureKeyCredential
import os

endpoint = os.environ["AZURE_DOCUMENT_INTELLIGENCE_ENDPOINT"]
key = os.environ["AZURE_DOCUMENT_INTELLIGENCE_KEY"]

client = DocumentIntelligenceClient(
    endpoint=endpoint,
    credential=AzureKeyCredential(key)
)

Processing Invoices

The prebuilt invoice model recognizes vendor details, line items, totals, and payment terms automatically.

def extract_invoice_data(invoice_url: str) -> dict:
    """Extract structured data from an invoice URL."""
    poller = client.begin_analyze_document(
        model_id="prebuilt-invoice",
        analyze_request={"urlSource": invoice_url}
    )
    result: AnalyzeResult = poller.result()

    extracted_data = {
        "invoices": []
    }

    for invoice in result.documents:
        invoice_data = {
            "vendor_name": get_field_value(invoice, "VendorName"),
            "vendor_address": get_field_value(invoice, "VendorAddress"),
            "invoice_id": get_field_value(invoice, "InvoiceId"),
            "invoice_date": get_field_value(invoice, "InvoiceDate"),
            "due_date": get_field_value(invoice, "DueDate"),
            "subtotal": get_field_value(invoice, "SubTotal"),
            "total_tax": get_field_value(invoice, "TotalTax"),
            "invoice_total": get_field_value(invoice, "InvoiceTotal"),
            "line_items": extract_line_items(invoice)
        }
        extracted_data["invoices"].append(invoice_data)

    return extracted_data

def get_field_value(document, field_name: str):
    """Safely extract field value with confidence score."""
    field = document.fields.get(field_name)
    if field:
        return {
            "value": field.content,
            "confidence": field.confidence
        }
    return None

def extract_line_items(invoice) -> list:
    """Extract individual line items from invoice."""
    items_field = invoice.fields.get("Items")
    if not items_field:
        return []

    line_items = []
    for item in items_field.value:
        line_items.append({
            "description": item.value.get("Description", {}).get("content"),
            "quantity": item.value.get("Quantity", {}).get("content"),
            "unit_price": item.value.get("UnitPrice", {}).get("content"),
            "amount": item.value.get("Amount", {}).get("content")
        })
    return line_items

Confidence Thresholds

Always implement confidence thresholds for automated processing. Fields below 0.85 confidence should trigger human review.

Document Intelligence handles multiple invoice formats without custom training, making it ideal for processing invoices from diverse vendors.