Can Power Automate Extract Invoice Data?

Last updated August 2026.

Yes. Power Automate extracts invoice data using the AI Builder invoice processing prebuilt model, which reads a PDF or image invoice and returns named fields including the vendor, invoice number, dates, subtotal, tax, total, and individual line items. You call it as an action inside a cloud flow, so the extracted data can be written straight to Excel, SharePoint, Dataverse, or an accounting system. It works without training, on invoice layouts it has never seen.

That is the short answer. The longer answer matters if you are deciding whether the Microsoft stack you already pay for can replace a dedicated extraction tool, because AI Builder carries real limits that are easy to miss until a flow starts failing in production.

What fields does AI Builder extract from an invoice?

The prebuilt invoice processing model returns a defined field list rather than raw text. Microsoft's documentation groups them into header fields and repeating collections. The header fields cover both sides of the transaction:

Three collections repeat per invoice. Items holds the line items, each with Description, Quantity, Unit, UnitPrice, Amount, ProductCode, Date, Tax, and TaxRate. TaxDetails holds tax Amount and Rate. PaymentDetails holds banking references including IBAN, SWIFT, and BankAccountNumber. Every one of those fields also carries a confidence score, which is the hook most well-built flows use to decide what a person needs to look at.

The line-item support is the part people underestimate. Plenty of tools stop at the invoice total. If you need line detail for job costing or three-way matching, AI Builder does return it.

What are the file limits for Power Automate invoice processing?

These are the constraints that break flows in production, and they are worth reading before you design around them. Microsoft publishes them directly:

That throughput cap is shared across document processing models in the environment, so receipt processing and invoice processing draw from the same 360 calls per minute. Microsoft also advises providing one clear photo or scan per invoice, which means a batch-scanned stack of ten invoices in a single PDF is not the input the model expects.

How do you build the flow?

The common pattern is four actions and takes an afternoon rather than a project:

1. Trigger. Start the flow when a file is created in a SharePoint document library, or when an email with an attachment lands in a monitored mailbox.
2. Extract. Add the AI Builder action for the invoice processing prebuilt model and pass it the file content.
3. Branch on confidence. Add a condition on the confidence score of the fields you care about. Microsoft's own worked examples use a threshold of 0.65 on the invoice total, routing anything below it for review.
4. Write the result. Add rows to an Excel table, a SharePoint list, or Dataverse, then post an approval to Teams if you need one.

The confidence branch is what separates a flow you can trust from one that quietly writes wrong totals into your books. Do not skip it.

What if AI Builder misses a field you need?

Microsoft documents three routes, and they are worth knowing because each has a different cost in effort:

The first is a custom invoices model, which augments the prebuilt model with additional fields or with samples of documents it handles poorly. The second is reading the raw OCR output: every time the prebuilt model runs it also performs a full OCR pass, and the detected text is available in the output, so a search across that content sometimes gets you the value without any training. The third is a separate document processing model you train yourself on your own documents.

Microsoft's documentation also shows a fallback pattern worth copying: train a custom model on your top 20 suppliers, then call the prebuilt model for everything else. New and low-volume vendors get handled without anyone configuring a template for them.

When is Power Automate the wrong tool for this?

It is the right tool when your organization already runs on Microsoft 365, invoices arrive in a predictable place like a SharePoint library, and someone on the team is comfortable building and maintaining cloud flows. In that setting you are using licensing you already hold.

It is the wrong tool in three situations. First, when nobody owns the flow: AI Builder consumes credits and cloud flows need maintenance, so an unowned flow becomes a silent failure nobody notices for a month. Second, when your documents are receipts rather than invoices, because crumpled thermal paper and phone photos are a different extraction problem from a clean supplier PDF. Third, when you just need a spreadsheet: if the goal is to turn a folder of invoices into an Excel file once a month, building and licensing a flow is a lot of machinery for an outcome a browser tool delivers in a few minutes.

There is also a language consideration that gets missed. The supported language list is broad and includes English (United States), but the model is tuned for the layouts it was trained on. Unusual formats still deserve the confidence check.

How does this fit into a wider automation workflow?

Extraction is one stage of five. A complete workflow captures invoices, extracts their data, validates it against purchase orders and duplicate checks, routes it for approval, and posts it to your accounting system. Power Automate handles the routing and approval stages well, because those are rules. Our guide to invoice processing automation walks through all five stages and is honest about which ones automate cleanly and which still need a person.

Cost is the other half of the decision. AI Builder is credit-based, while the underlying document AI services are priced per page or per document, and the spread between vendors is wide enough to change which approach makes sense at your volume. We keep a verified breakdown on the OCR API pricing page, and if you want the extraction step without building a flow at all, the invoice OCR software route gives you the same structured output as an Excel or CSV file. Teams whose documents are receipts rather than supplier invoices should start with the OCR receipt scanner instead.

One practical note from running these flows: extraction solves getting data in, but it creates a second problem downstream. Once thousands of invoices and their extracted fields are spread across SharePoint libraries and Excel tables, locating one specific document a year later is genuinely hard, which is the gap enterprise search across your internal systems is built to close. Worth planning for before the archive gets big rather than after.

The short version

Power Automate does extract invoice data, and the AI Builder prebuilt model is more capable than its reputation suggests, particularly on line items and confidence scoring. Build the confidence branch, respect the 20 MB and 360-calls-per-minute limits, and decide honestly whether you have someone to own the flow. If the answer is no, or if you simply want invoices turned into a clean spreadsheet without a build, a browser-based extraction tool gets you the same data with nothing to maintain.