Automatically extract invoice data with AI Builder

AI Builder processes invoices in Power Automate via the prebuilt action "Extract information from invoices": you attach a PDF or image file to a cloud flow, the AI model reads fields such as invoice number, invoice date, vendor, total amount and tax amounts, and returns a confidence value between 0 and 1 for each field. Usage is billed via AI Builder credits, using it in a flow requires a Premium license in Power Automate or Power Apps, and recognition accuracy varies significantly depending on the field and invoice layout. As of: July 2026.

How to build the flow with the invoice processing model

According to the Microsoft guide on the prebuilt invoice processing model in Power Automate you first create a cloud flow with a manual trigger and a file input field, for example "My invoice". As the next step, add the AI Builder action "Extract information from invoices" and link it to the uploaded file. In practice, the trigger more often runs via a SharePoint or email attachment trigger instead of manual input; the principle remains the same. For multi-page documents containing several invoices, the "Pages" parameter lets you specify the page range of a single invoice. Microsoft explicitly recommends this, because it both reduces the cost of the model prediction and increases the hit rate, as long as the selected range actually contains only one invoice.

What data the model delivers and how reliable it is

According to the invoice processing overview the model returns several dozen fields, including:

• Header data: Invoice ID, invoice date, due date, payment terms
• Amounts: Subtotal, total tax, invoice total, amount due
• Addresses: Billing, shipping, customer and vendor address including recipient
• Payment details: IBAN, SWIFT and bank account number, if stated on the invoice

For each individual field, the model also returns a confidence value between 0 and 1. This is the honest part of the model: it does not recognize every field equally well on every invoice; especially with unusual layouts, poor scans or rare languages, the hit rate fluctuates noticeably. For fields with persistently low confidence, AI Builder offers the option of adding your own document processing model trained on your own vendors. According to Microsoft, around 35 languages are supported, including German, for files up to 20 MB in PDF, JPEG or PNG format.

What the processing costs in AI Builder credits

Honesty pays off here rather than marketing promises: every processed invoice page costs credits, regardless of whether the model ultimately recognizes all fields correctly. According to the licensing overview the category "Receipt, invoice, ID document analysis" consumes 32 AI Builder credits per page, equivalent to 8 Copilot credits. With an annually prepaid Tier 1 add-on with a capacity of 1 million credits for 500 US dollars per month, that equals around 1.6 US cents per processed invoice page. 500 single-page invoices per month therefore consume about 16,000 credits, a fraction of this capacity; however, with multi-page batch PDFs containing several invoices, consumption rises quickly, which is why the page range parameter mentioned above directly saves money here. Important for budget planning: the 5,000 AI Builder credits included in a Power Automate Premium license will, according to Microsoft, be discontinued on November 1, 2026, and since November 2025 new customers can no longer purchase AI Builder capacity add-ons but must switch to Copilot credits. If you process beyond your quota, you are not automatically charged extra; instead, processing is blocked or redirected to Copilot credits until new capacity becomes available or the month resets.

Premium license and other requirements

As soon as an AI Builder action is used in a flow, that flow, according to Microsoft, automatically becomes a Premium flow and requires a Power Automate Premium license or a Power Apps Premium license; the free basic plans are not sufficient. In addition, Microsoft limits calls to the document processing model to 360 per environment within 60 seconds, which can become relevant with very high invoice volumes. If you want to test AI Builder: model tests in the prompt or model view are, according to Microsoft, free and do not consume credits; only actual execution in flows or apps costs credits. If you want to extend an existing Power Automate environment with invoice processing, the Power Automate consulting from NordFlux provides support with license selection, flow setup and a realistic cost estimate before invoices are processed in production. If you want to clarify in general whether AI Builder or another AI approach fits your own invoice processing, you can work this out in advance as part of AI consulting.

Frequently asked questions about automatically extracting invoices with AI Builder

How accurately does AI Builder recognize invoice data?

Accuracy depends on the field and is output per field as a confidence value between 0 and 1; Microsoft does not publish a blanket percentage for the overall model. For standardized invoices with a clear structure, the values are usually high; for unusual layouts or poor scans, you should specifically intercept low confidence values in your flow, for example through a manual review.

Do I always need a Premium license for AI Builder in Power Automate?

Yes. As soon as an AI Builder action is included in a flow, the entire flow counts as a Premium flow and requires a corresponding Power Automate or Power Apps license. A plain basic plan without Premium connectors is not sufficient for this.

What happens if my AI Builder credits run out during the month?

Processing first tries to switch to available Copilot credits. If these are also used up, model execution is blocked until additional capacity is allocated or consumption resets at the start of the month. According to Microsoft, there is no automatic extra billing beyond your quota.

Can I also use the standard model to extract invoices from atypical vendors?

Yes, the prebuilt model is designed for common invoice formats from many supported languages. If certain fields are persistently recognized poorly for a particular vendor, you can build a supplementary, individually trained document processing model and combine it with the standard model.