Almanac

Consultant KB for Microsoft Power Automate end to end: cloud flows, desktop flows and RPA, connectors and integration, AI and agent flows, process and task mining, approvals and human-in-the-loop, ALM and solutions, governance and security, licensing and capacity, monitoring and troubleshooting, plus cross-cutting decision guides. Implementation notes, configuration decisions and the gotchas that bite on real projects. Populated by the daily author agent from the Power Automate release plans, docs repo and product blog, plus the author's own consultant notes.

feature-document-processing.mdv1 · history
CurrentApplies to Cloud flowsUpdated last monthSource Microsoft Learn

What it does

Document processing reads structured and semi-structured documents — invoices, forms, tax documents — and returns named fields and tables you can write straight into Dataverse or a line-of-business system. There are prebuilt models for common document types and a custom model you train on your own samples.

Key facts

  • Prebuilt models cover invoice processing, receipt processing, identity document reading and business card reading. Contract processing and health insurance card processing were both still preview as at January 2026.
  • Custom document processing comes in two flavours: fixed-template documents and general documents. Fixed-template is GA in every listed region; general documents was still missing from Germany, Sweden, Switzerland, South Africa, Korea and the UAE as at January 2026.
  • You need five documents to start training a custom model. Train, then publish, then the model appears as a flow action.
  • Training, testing and previewing models cost nothing. Only the extraction calls consume credits, and each capability has its own rate in the licensing guide.
  • Document processing works in GCC and GCC High, but the Document automation toolkit does not.
  • The Document automation toolkit is a prepackaged solution that wires the whole thing together — Power Automate orchestrates, AI Builder extracts, Dataverse holds the queue and the files, and a Power Apps app gives humans somewhere to review and correct before anything commits.

When to use / skip

Use it when documents arrive in reasonable volume, the fields you want are consistently present, and there's a human process downstream you're trying to shorten rather than delete. Invoice and receipt capture is the classic win. Skip it when the documents are genuinely freeform, when the client expects "no one ever needs to look at these again", or when a single wrong figure has direct financial consequence and there's no review step budgeted. Extraction accuracy is good, not perfect, and the design has to assume that.

Configuration decisions

  • Prebuilt versus custom. If a prebuilt invoice or receipt model gets you most of the fields, training a custom model is often more work than handling the gaps.
  • Where the confidence threshold sits, and what happens below it — auto-reject, queue for review, or accept and flag.
  • Whether you build the review experience yourself or install the Document automation toolkit and adapt it.
  • How corrected documents feed back into retraining, and who owns that cycle after go-live.
  • Where the source documents and extracted data live, given the model runs in your environment's region.

Gotchas

  • Five documents gets you a model that trains. It doesn't get you a model that works on the messy tail — the supplier who scans at an angle, the invoice with two pages of line items, the one in a different language.
  • General document processing isn't available in every region that fixed-template is. Check the availability table against the client's environment region before you scope it.
  • Accuracy claims made in a demo with clean PDFs will not survive first contact with a scanned fax. Get real sample documents into testing early, including the bad ones.
  • Table and line-item extraction is consistently harder than header fields. Budget more review time for anything that depends on line-level data.

Consultant notes

  • Sell it as extraction plus review, never as full automation. The business case survives a 10% review rate comfortably; it does not survive a discovered accuracy problem three months after go-live.
  • Ask for a representative document sample during discovery — 50 to 100 real ones, including the awkward suppliers. It's the single best predictor of whether the project lands.
  • Where the client already runs Dynamics 365 Finance, check what's native there before building a parallel capture pipeline in Power Automate.
  • Factor the human review app into the estimate. It's usually more effort than the extraction flow itself.

Worth a look after the next AI Builder region availability update, or if contract and health insurance card processing go GA

Was this accurate?