Power Automate + AI Builder: Document Processing Guide

What Is Intelligent Document Processing?

Intelligent Document Processing (IDP) uses AI to automatically extract, classify, and process data from unstructured documents like invoices, receipts, contracts, and forms. Instead of manually typing data from a PDF into a spreadsheet, IDP reads the document for you.

In the Microsoft ecosystem, this is powered by AI Builder — a low-code AI capability built directly into the Power Platform. Combined with Power Automate, you can build end-to-end pipelines that:

• Receive a document (via email, SharePoint upload, or Teams message)

• Extract structured data using an AI model

• Validate and route the data to downstream systems

• Notify stakeholders and log an audit trail

All without writing a single line of traditional code.

Why This Matters in 2026

The push toward hyperautomation — automating entire business processes rather than individual tasks — has made document processing one of the highest-ROI automation targets.

• 80% of enterprise data is unstructured (PDFs, scanned images, emails)

• AI Builder models are pre-trained on millions of documents

• Custom AI models let you train on your own document formats with as few as 5 sample documents

• Power Automate integration means extracted data flows directly into SharePoint, Dataverse, Dynamics 365, or any of 1,000+ connectors

Prerequisites

• Power Automate Premium license (AI Builder requires Premium or per-user plan)

• AI Builder credits — included with certain Microsoft 365 and Dynamics 365 plans

• SharePoint Online — for document storage and triggers

• A Microsoft 365 environment — developer tenant works fine for testing

Step 1: Choose Your AI Model

AI Builder offers several prebuilt models that require zero training:

Model	What It Extracts	Best For

|-------|-----------------|----------|

Invoice Processing	Vendor, amounts, line items, dates, PO numbers	Accounts payable automation

Receipt Processing	Merchant, total, tax, date, payment method	Expense report automation

Identity Document Reader	Name, DOB, address, document number	KYC/onboarding workflows

Business Card Reader	Name, title, company, email, phone	CRM lead capture

Text Recognition (OCR)	Raw text from any image or PDF	General-purpose extraction

Custom Document Processing	Any fields you define	Custom forms, applications

Step 2: Build the Power Automate Flow

Trigger: When a File Is Created in SharePoint

• Go to make.powerautomate.com

• Click Create → Automated cloud flow

• Name it: Invoice Processing Pipeline

• Trigger: When a file is created (properties only) — SharePoint

Action: Extract Information from Invoices

• Click New step → search for AI Builder

• Select Extract information from invoices

• For the Invoice file parameter, use the dynamic content picker

AI Builder returns: Invoice ID, dates, vendor info, amounts, line items, and confidence scores.

Action: Create Item in SharePoint List

Map the AI Builder outputs to your SharePoint list columns.

Action: Handle Low-Confidence Results

Add a Condition to route uncertain results for human review with a threshold of 0.85.

Step 3: Train a Custom AI Model (Optional)

For custom forms, insurance claims, or government applications, custom models give you granular control:

• Go to make.powerapps.com → AI Builder → Explore

• Click Document processing → Create custom model

• Define fields you want to extract (e.g., "Claim Amount", "Claimant Name", "Date of Incident")

• Upload 5+ representative sample documents

• Manually tag each field in the samples so the model learns patterns

• Train the model (typically takes 5-10 minutes)

• Test with a new document to validate accuracy

• Publish to make available in Power Automate

The more varied your training samples, the better the model generalizes to new documents. A model trained only on scanned invoices will perform poorly on digital PDFs.

Step 4: Production-Ready Patterns

Error Handling

Wrap AI Builder actions in Try-Catch using Scope actions.

Batch Processing

Use Apply to each for high-volume document processing.

Approval Workflow Integration

Combine with Power Automate Approvals for procure-to-pay workflows with tiered approval thresholds.

Real-World Performance Metrics

Metric	Before IDP	After IDP

|--------|-----------|-----------|

Invoice processing time	15-20 min/invoice	30 seconds

Error rate	5-8% (human entry)	1-2% (AI + human review)

Monthly capacity	~500 invoices/person	~5,000 invoices/person

Licensing and Costs

AI Builder uses a credit-based consumption model. Prebuilt and custom models cost ~1 credit per page. Credits are included with Power Automate Premium and related licenses.

Tips and Best Practices

• Start with prebuilt models first

• Use PDF format when possible

• Set confidence thresholds per field (typically 0.8–0.95 depending on risk tolerance)

• Build a feedback loop for accuracy tracking

• Test with edge cases (multi-page, rotated, low-quality)

• Monitor AI Builder credit consumption in production (prebuilt: ~1 credit per page; custom: similar)

FAQ: Confidence Scores & Production Deployment

What's a confidence score and how do I use it?

AI Builder returns a confidence score (0–1) for each extracted field. A score of 0.95 means the model is 95% confident in the extraction. In your flow, check the confidence: if it's below your threshold (e.g., 0.85), flag the document for human review instead of auto-processing.

Example condition in Power Automate:
``
if(greater(outputs('Extract information from invoices')['body/vendorName/confidence'], 0.85),
'Process automatically',
'Route for manual review'
)
`

How many documents do I need to train a custom model?

Minimum: 5 samples. Recommended: 20–50 for good accuracy. The more varied (different layouts, quality levels, languages), the better the model generalizes.

What happens if AI Builder can't extract a field?

The field will be returned as null or empty. Always wrap downstream actions in conditions that check for null values using coalesce()` to provide sensible defaults.

Can I retrain a model after it's published?

Yes. Upload new samples, re-tag, and retrain. Publish a new version. Power Automate always uses the latest published version by default.

How do I monitor production accuracy?

Log every extraction result and the actual correct value to a SharePoint list. Periodically audit: compare AI-extracted values to ground truth. If accuracy drops below your threshold, trigger retraining.

What's Next for AI Builder in 2026

• Multi-modal document understanding

• Cross-document intelligence

• Continuous learning

• Natural language queries via Copilot

AI Builder is becoming the standard way to bridge physical documents and digital business processes in the Microsoft ecosystem.