Power Apps AI Builder: Complete Guide for Business

Your team is drowning in manual data entry—copying information from invoices into spreadsheets, sorting through email attachments to extract key details, or manually reviewing forms for completeness. AI Builder brings pre-built AI models directly into Power Apps, letting you automate document processing, predict outcomes, and recognize text without writing code or hiring data scientists. This guide walks you through building your first AI-powered app in under two hours.

What You'll Learn

How to enable AI Builder in your Power Platform environment and understand licensing requirements
Build a canvas app that extracts data from invoices and receipts using pre-built document processing models
Train a custom form processing model to read your specific business documents like purchase orders or contracts
Add prediction models to forecast sales outcomes or prioritize customer requests automatically
Connect AI Builder output directly to Dataverse or SharePoint to eliminate manual data entry
Set up approval workflows in Power Automate triggered by AI confidence scores

Prerequisites

Power Apps Premium license ($20/user/mo) or Power Apps per app ($10/app/user/mo) — AI Builder requires premium licensing
Administrator or maker permissions in a Power Platform environment (not the default environment for production use)
Sample documents ready: 5-10 examples of invoices, receipts, or forms you want to process
Basic familiarity with Power Apps canvas apps (knowing how to add controls and navigate screens)

Step 1

Verify AI Builder Credits and Environment Setup

Navigate to the Power Platform Admin Center (admin.powerplatform.microsoft.com) and select your target environment. Click Settings > Features and confirm AI Builder is enabled. Check your AI Builder credit allocation under Resources > Capacity > Add-ons—Premium licenses include 500 credits/user/month, and each AI model prediction consumes 1-5 credits depending on complexity. Invoice processing uses approximately 5 credits per document. If you're running low, you can purchase additional credit packs for $500 per 1 million credits.

💡 Tip: Start in a dedicated development or sandbox environment, not your default environment. This prevents accidental consumption of production credits during testing and allows you to control governance policies separately.

Step 2

Create a Canvas App with the AI Builder Component

Open Power Apps (make.powerapps.com), select your environment, and click Create > Canvas app from blank. Name it "Invoice Processing App" and choose Tablet format for better document viewing. In the Insert menu, expand AI Builder and select the Form Processor component—this drag-and-drop control handles document upload, AI processing, and result display without custom code. Position it prominently on your first screen. The component automatically provides upload buttons and displays extracted fields once processing completes.

💡 Tip: Use Tablet format rather than Phone for document-heavy apps. Users need screen real estate to review extracted data alongside the original document image.

Step 3

Configure the Pre-Built Invoice Processing Model

With the Form Processor component selected, open the Properties pane and click AI model > Select model. Choose the pre-built "Invoice processing" model—this Microsoft-trained model extracts vendor name, invoice date, total amount, line items, and tax without any training on your part. It works immediately with invoices in English, Spanish, French, German, Italian, and Portuguese. Set the ModelType property to "Prebuilt" and OnResults action to store extracted data. The model handles varied invoice layouts automatically, from simple PDFs to scanned images.

⚠ Watch out: Pre-built models work for 80% of standard invoices, but custom forms like internal purchase orders require training a custom model in Step 5. Test with your actual documents before committing to the pre-built approach.

Step 4

Capture AI Results in a Collection or Dataverse Table

In the Form Processor's OnResults property, write a formula to capture extracted data: ClearCollect(colInvoiceData, FormProcessor1.Results). This stores vendor name, invoice number, date, amount, and line items in a collection you can display in a gallery or send to approval workflows. For persistent storage, create a Dataverse table named "Processed Invoices" with columns matching the extracted fields (Vendor as text, InvoiceDate as date, TotalAmount as currency). Use Patch(ProcessedInvoices, Defaults(ProcessedInvoices), {Vendor: FormProcessor1.Results.VendorName, Amount: FormProcessor1.Results.InvoiceTotal}) to write directly to Dataverse. Display the collection in an editable gallery so users can correct any extraction errors before final submission.

💡 Tip: Always include a confidence score column in your storage. FormProcessor1.Results.Confidence returns 0-100 for each field—flag anything below 80 for manual review to maintain data quality.

Step 5

Train a Custom Form Processing Model for Your Documents

For non-standard forms like your company's purchase orders, navigate to AI Builder in the left menu and select Build > Document processing > Form processing. Click Create and upload 5 sample filled purchase orders—AI Builder needs at least 5 examples but 15-20 dramatically improves accuracy. Draw bounding boxes around each field you want to extract (PO Number, Vendor, Date, Line Items, Approval Signature) and label them consistently across all samples. Click Train and wait 5-15 minutes. Once trained, the model shows accuracy metrics per field—aim for 85%+ before deploying. Return to your canvas app, change the Form Processor component's AI model to your custom "PO Processing" model, and test with new documents.

💡 Tip: Use documents with consistent layouts initially. Include examples with handwriting if applicable, and varied vendors to build robustness. You can add more training documents later to improve accuracy without starting over.

Step 6

Add Object Detection for Image-Based Workflows

If your process involves photos (damaged goods claims, inventory checks, safety inspections), add the Object Detector component from AI Builder. You can use pre-built models for common objects or train a custom model to recognize specific products, equipment, or defects. Create a new screen in your app, insert the Object Detector component, and configure it to capture from camera or file upload. For custom detection, go to AI Builder > Build > Object detection, upload 15+ tagged images showing the items you want to detect, and train the model. When the model runs in your app, it returns detected objects with confidence scores and bounding box coordinates—perfect for automated damage assessment or compliance checks.

⚠ Watch out: Object detection models require more training images than document processing (minimum 15 per object type, preferably 50+). Start with clearly defined objects before attempting complex detection scenarios.

Step 7

Build Prediction Models for Business Forecasting

Navigate to AI Builder > Build > Prediction and choose a scenario: predict a category (Will this lead close? Will this customer churn?) or predict a number (What will this deal's revenue be?). Connect historical data from Dataverse or your Excel file in OneDrive—you need at least 50 rows but 1,000+ produces reliable models. Select the field to predict (e.g., "OpportunityStatus") and let AI Builder choose relevant influence factors from your data columns. Train the model, review performance metrics (accuracy, precision, recall), and publish. Back in your app, use the Predict function: Set(varChurnRisk, 'Customer Churn Model'.Predict({LastPurchaseDate: DateValue(InputDate.Text), SupportTickets: Value(InputTickets.Text)})). Display the prediction and confidence score to guide user decisions.

💡 Tip: Prediction models shine when you have 6+ months of clean historical data. Start with simple binary predictions (yes/no outcomes) before attempting multi-class or numeric predictions.

Step 8

Connect AI Builder to Power Automate for End-to-End Automation

Create a cloud flow in Power Automate triggered when a file is added to a SharePoint document library (your invoice drop folder). Add the AI Builder action "Extract information from forms" and select your trained invoice model. The flow automatically processes new uploads without user intervention. Add a Condition action to check if AI Builder's confidence score is above 80%—if yes, automatically create the record in Dataverse and send an approval request to the finance manager with extracted details pre-filled. If confidence is low, route to a human for review. This eliminates the app interface entirely for routine documents while catching edge cases.

💡 Tip: Power Automate Desktop can capture documents from desktop apps or email attachments, save them to SharePoint, and trigger your cloud flow—perfect for businesses still receiving invoices via email rather than web upload.

Step 9

Implement Receipt Scanning for Expense Reports

Add a new screen to your app for mobile expense submission. Insert the Receipt Processor component from AI Builder—this pre-built model extracts merchant name, date, total, tax, and payment method from photos of receipts. Set the component to Camera mode so field employees can snap and submit receipts on the spot. In the OnResults property, write Patch(ExpenseReports, Defaults(ExpenseReports), {Merchant: ReceiptProcessor1.MerchantName, Amount: ReceiptProcessor1.Total, ExpenseDate: ReceiptProcessor1.TransactionDate, SubmittedBy: User().Email, ReceiptImage: ReceiptProcessor1.OriginalImage}). This creates an expense record with the receipt image attached, cutting expense report time from 20 minutes to under 2 minutes per submission.

⚠ Watch out: Receipt processing accuracy drops significantly with crumpled, faded, or poorly lit receipts. Add a preview screen so users can retake photos before processing—saves credits and improves data quality.

Step 10

Add Text Recognition for Handwritten Notes or Labels

For scenarios involving handwritten forms, field notes, or product labels, use the Text Recognizer component. This OCR functionality extracts printed and handwritten text from images. Insert the component, configure it for camera or upload, and capture the extracted text in a variable: Set(varExtractedText, TextRecognizer1.Results). Unlike form processing, text recognition returns unstructured plain text—you'll need to parse it with formulas (Split, Find, Mid functions) to extract specific values. This works well for serial numbers, handwritten inspection notes, or whiteboard captures during meetings where structure varies significantly.

💡 Tip: Text recognition uses fewer credits (1-2 per image) than form processing, making it cost-effective when you only need raw text rather than structured field extraction.

Step 11

Monitor AI Builder Usage and Cost in the Admin Center

Return to the Power Platform Admin Center > Resources > Capacity > Add-ons and review your AI Builder credit consumption. The dashboard shows credits used per environment, per model, and per day. Identify high-consumption models and optimize by increasing confidence thresholds (fewer retries), reducing image resolution for object detection, or batching predictions. Set up automated alerts when credits drop below 20% remaining so you can purchase additional packs or throttle usage before running out mid-month. Each invoice processed costs approximately 5 credits, receipts cost 2-3, and predictions cost 1, so a 500-credit allocation handles roughly 100 invoice submissions per user per month.

💡 Tip: Credits reset monthly with your subscription—they don't roll over. Schedule high-volume processing tasks (like end-of-month invoice batches) early in the billing cycle to avoid mid-month credit exhaustion.

Step 12

Deploy with User Training and Feedback Loops

Share your app with end users through the Power Apps portal, assign them Power Apps Premium licenses, and provide a 10-minute walkthrough focusing on document quality (lighting, focus, orientation) and when to override AI suggestions. Create a feedback mechanism—add a "Report Incorrect Extraction" button that logs the document ID, extracted values, and user corrections to a Dataverse feedback table. Review this data monthly to identify patterns (specific vendors the model struggles with, common field errors) and retrain your custom models with corrected examples. Plan to retrain quarterly as document formats evolve or new vendors are added. Document typical time savings (invoice entry drops from 5 minutes to 30 seconds) and report monthly hours saved to justify continued investment.

⚠ Watch out: AI models degrade over time as document formats change. Schedule quarterly retraining sessions with 10-15 new examples to maintain accuracy above 85%.

Summary

You've now built an AI-powered Power Apps solution that eliminates manual data entry from invoices, receipts, and custom forms. Your app uses AI Builder's pre-built and custom models to extract structured data, stores results in Dataverse for reporting, and can trigger automated approval workflows—all without writing traditional code or managing infrastructure. Typical implementations save 15-20 hours per employee monthly by eliminating manual document processing.

Next Steps

Expand to Power Automate Desktop for RPA: Use AI Builder models inside desktop flows to process documents from legacy desktop apps or email clients that can't integrate directly with Power Apps
Enroll in PL-7001: Create and Manage Canvas Apps to master advanced Power Apps patterns including offline data sync, complex formulas, and component libraries for scaling your solution across departments
Schedule a 30-minute assessment with Scott Hay to identify your highest-ROI automation opportunities and build a 90-day Power Platform implementation roadmap tailored to your business processes
Implement governance policies using DLP (Data Loss Prevention) to control which connectors citizen developers can use with AI Builder, preventing accidental data exfiltration while enabling innovation

Need Power Platform Built Right the First Time?

Power Platform is flexible—which means it's easy to build poorly. I help SMBs architect Power Apps, Power Automate, and Power BI solutions that scale. Governance included. 90-day delivery, you own the IP.

Book Power Platform Consultation

Scott Hay Microsoft Certified Trainer & AI Solutions Architect Microsoft Certified Trainer (MCT) • Delivers 12 Microsoft Copilot courses (MS-4002 through MS-4023) plus Azure AI, Power BI • Azure AI Agents, Semantic Kernel, Power BI (PL-300), Power Platform certified • Former Microsoft and Amazon — 30+ years building production systems • Builds custom AI solutions for SMBs with 90-day delivery