AI Builder Document Processing: Extraction and Review Steps

AI Builder document processing is a Microsoft Power Platform feature that extracts structured field values from documents—invoices, receipts, forms, contracts—using a model you train on your own examples. The model learns the layout of your documents so it can read new ones without manual configuration for each file.

Building the extraction step is straightforward; keeping the decisions about exceptions consistent across runs is harder. Tools like fenbs retain the extraction rules and the decisions made when a model fails or a value is disputed, so each run starts from the same instructions even when different people or assistants are involved.

How AI Builder document processing works

You create a document processing model in AI Builder by labeling the fields you want to extract across a set of sample documents. AI Builder uses those samples to learn where each field typically appears and what it looks like. Once trained, the model can extract the same fields from new documents without you labeling each one.

The model returns a confidence score alongside each extracted value. A high score means the model is fairly certain about that value; a low score flags it for review. Neither score confirms that the extracted value is correct—it describes the model's certainty, not the fact. Verification is still required for any value that will drive a financial, legal, or compliance decision.

How to prepare training documents

The quality of the trained model depends on the range of documents you use to build it. A model trained on invoices from a single vendor will struggle with invoices from others that use a different layout.

1. Gather at least five examples of each document layout you want the model to handle. More variety produces a more robust model.
2. Make sure the examples cover the range of formats you will actually see: different vendors, different paper sizes, printed versus scanned.
3. Label every field consistently across all examples. A field labeled "Invoice Total" in some documents and "Amount Due" in others will confuse the model.
4. Review the model's extraction on a test set of documents it has not seen before you use it in production.

How to handle low-confidence and missing values

In your Power Automate flow, route documents with low-confidence or missing values to a review step rather than letting them pass through unchecked. The review step should show the reviewer the original document alongside the extracted value, so the comparison takes seconds rather than minutes.

Decide in advance what threshold triggers a review and what to do with a document that cannot be processed at all—whether to reject it, queue it for manual entry, or escalate it. Record those decisions before you build the flow, not during an incident when a batch is stuck. A workflow that handles exceptions cleanly is far easier to maintain and audit than one that drops them silently.

How to build a review step in Power Automate

A practical review step in Power Automate sends an approval request to a named reviewer when the confidence score falls below a threshold you set. Include the extracted value, the confidence score, and a link or thumbnail of the source document in the approval message. The reviewer confirms or corrects the value, and the flow continues with the verified result.

• Set the confidence threshold in one place so it is easy to adjust as the model improves.
• Route the approval to the person best placed to judge the value—an accounts team member for invoice amounts, a procurement team member for purchase-order numbers.
• Log every manual correction so you can see whether a field is systematically uncertain and retrain the model with better examples if it is.

How to improve the model over time

Each manual correction is a signal that the model needs more examples of that field layout. Keep a record of which fields trigger the most corrections and retrain the model with those examples added to the training set. Three or four retraining cycles on a real document set usually produce a model that handles most of its cases automatically.

For workflow automation that acts on the extracted data, see AI workflow builder.

How to do this with fenbs

fenbs does not process documents or train models. It keeps the rules, exception decisions, and improvement history so each run of the workflow starts from the same agreed baseline.

1. Create a free account and open a project for the document processing workflow.
2. Write the field definitions as AI context notes: field name, description, expected format, and the confidence threshold for that field.
3. Record each exception-handling rule as a Decisions and rules entry: what triggers a review, who reviews it, and what to do if a document cannot be processed.
4. After each batch, create a task for every field that needed manual correction more than once. Use those tasks to track retraining work.
5. Before changing a threshold or retraining the model, record the change and the reason as a decision so the rationale is available to the next person who reviews the workflow.