Power Platform AI Builder Document Processing
Offices drown in invoices, receipts, and forms. AI Builder reads these documents, extracts the key fields, and passes them to an app or flow. Staff stop retyping and start reviewing.
What AI Builder Offers
AI Builder is a set of ready-to-use AI skills inside Power Platform. Some models work without training, and others learn from your own samples.
| Model | Job |
|---|---|
| Invoice processing | Reads vendor, date, total, and line items from invoices |
| Receipt processing | Reads merchant, date, and amounts from receipts |
| Document processing | Learns fields from your own document layouts |
| Text recognition | Converts images of text into editable text |
| Category classification | Sorts text into labels, such as complaint or praise |
| Sentiment analysis | Rates text as positive, neutral, or negative |
| Prediction | Estimates an outcome from historical data |
Model names and availability change as Microsoft updates the product, so check the AI Builder page in your environment for the current list.
The Document Processing Pipeline
[Email arrives with PDF invoice]
|
v
[Flow saves attachment to SharePoint]
|
v
[AI Builder: Extract information from invoices]
|
v
Vendor = "Green Supplies" Total = 1,240.00 Due = 15 Nov
|
v
[Confidence above 80%?]
| yes | no
v v
[Create record in Dataverse] [Send to human review in Power Apps]
Using a Prebuilt Invoice Model
- Create a cloud flow that triggers on a new email with attachments.
- Add the action Extract information from invoices.
- Pass the attachment content into the action.
- Read the extracted fields from the dynamic content panel.
- Save the fields to a Dataverse table or a SharePoint list.
Training a Custom Document Model
Collect Samples
Gather at least five documents with the same layout. More samples with slight variations improve accuracy.
Define Fields
Create the fields you want, such as Policy Number, Customer Name, and Premium Amount.
Tag the Samples
Draw boxes around each value in every sample. The model learns where fields appear.
Train and Publish
Train the model, check the accuracy score, and publish it. Flows and apps can then call the model.
Confidence Scores
Every extracted field carries a confidence score between zero and one. A score of 0.95 signals a reliable read, and a score of 0.4 signals a doubtful one. Set a threshold in the flow. High scores go straight through, and low scores go to a person. This human-in-the-loop pattern protects accuracy while still saving time.
Layman Example: A Mailroom Clerk
A mailroom clerk opens each envelope, finds the sender and the amount, and writes them in a register. Dozens of letters take an hour. AI Builder acts as a fast clerk who reads every page in seconds and flags only the unclear ones for a colleague.
Credits and Licensing
AI Builder usage consumes AI credits, and the amount depends on the model and the plan. Review the licensing guide from Microsoft and monitor consumption in the admin center before launching a high-volume solution.
Good Practices
- Use clear scans with good lighting and straight alignment.
- Protect personal and financial data with proper security roles.
- Store the original document next to the extracted data for audits.
- Measure accuracy monthly and retrain the model when layouts change.
Handling Errors in Document Flows
Scans arrive blurry, rotated, or empty. Add a condition that checks whether the model returned any fields. Send failed documents to a review folder and notify the owner. Keep a log table with the document name, the processing date, and the outcome, so the team can spot patterns such as one vendor whose invoices always fail.
Key Takeaways
AI Builder converts paper-style data into structured rows. Combine extraction, confidence checks, and human review to build document workflows that stay fast and trustworthy.