Section 7: HR, finance-accounting, legal-compliance, procurement and operations

Scenario 64: Extracting invoice data with confidence checks

Level: Intermediate · Tools: Power Automate and AI Builder prebuilt **Invoice processing**; capacity and a Power Automate Premium license are required [10] [124] [125].

Who and in what situation

Derya is an accounting assistant at a logistics SME in Gebze. She prepares invoice number, vendor, date, purchase order number, tax and total fields from supplier PDFs for accounting entry.

How it is done today

The PDF is opened, the fields are entered manually into the ERP, and the total is checked against the invoice again. Scanned documents or documents with different layouts require additional checks.

How to do it with AI

  1. She selects My flows > New flow > Instant cloud flow; she adds a file field using Manually trigger a flow > +Add an input > File [124].
  2. She selects +New step > AI Builder > Extract information from invoices and links the file to the Invoice file field [124].
  3. She takes the invoice number, vendor, date, due date, purchase order number, tax and total from the output; if necessary, she selects the relevant page using Pages.
  4. She writes the confidence score and the original PDF value to the review queue; she does not automatically approve low-confidence records. She tries a sample invoice using Save > Test [124].

Example prompt

Extract only the clearly visible invoice number, vendor, invoice date, due date, purchase order number, subtotal, tax, grand total and currency from this invoice. Provide the original text, normalized value and confidence score. Leave any field you cannot find blank; do not estimate the total. Do not create an accounting entry.

Evidence and real-world example

Microsoft’s Natixis CIB New York story describes invoice extraction using Power Automate, Power Apps and AI Builder; it reports 90% accuracy and a scale of 3,000-5,000 invoices/year in the customer’s implementation [126].

Points to consider

Turkish does not appear in AI Builder’s list of supported languages; pilot Turkish invoices with real examples [10]. A confidence score is not a guarantee. AI Builder actions are associated with premium flows and credit consumption; check capacity with your administrator [125].

Sources for this scenario

  1. Invoice processing prebuilt AI model (learn.microsoft.com)
  2. Use the invoice processing prebuilt model in Power Automate (learn.microsoft.com)
  3. Overview of AI Builder licensing (learn.microsoft.com)
  4. Natixis CIB in New York achieves invoice processing accuracy with Power Platform (microsoft.com)