AI Builder Basics

Analogy categoryAnalogies

What AI Builder Adds to Power Automate

AI Builder is Microsoft's low-code AI capability built into the Power Platform, exposing both prebuilt models (already trained by Microsoft on general data) and custom models (trained on your own sample data) as actions you can drop directly into a Power Automate flow, a Power App, or a Copilot Studio agent. Prebuilt models cover common needs out of the box: 'Extract information from documents' handles invoices and receipts, 'Business card reader' pulls contact fields, 'Detect and count objects in images' locates items in photos, and 'Sentiment analysis' and 'Text recognition (OCR)' handle unstructured text. Every AI Builder action returns structured JSON output, confidence scores per extracted field, and often a bounding box or position, so a flow can branch on confidence (for example, routing anything below 80% confidence to a human reviewer) rather than blindly trusting every prediction.

🏏

Cricket analogy: Like a TV broadcast's ball-tracking overlay instantly estimating trajectory without a human plotting each frame, AI Builder's prebuilt models instantly extract structured data from a document without anyone hand-keying the fields.

Training a Custom Document Processing Model

When the prebuilt invoice or receipt model doesn't match your organization's document layout closely enough, you build a custom 'Document processing' model in the AI Builder studio by uploading a set of sample documents (Microsoft recommends at least five per distinct layout for reliable extraction) and manually tagging the fields you want extracted, such as PO Number, Vendor Name, or Line Item Total. The training process runs in the background and produces an accuracy score per field once complete; fields below the accuracy you need should get more tagged samples covering layout variation (different vendors, different page counts) rather than more samples of the same single layout, since the model generalizes from variety, not volume alone. Once published, the custom model appears as a new action in Power Automate ('Predict' or a model-specific action name) and returns the same structured JSON-with-confidence-scores pattern as the prebuilt models, so downstream flow logic doesn't need to change based on whether the model is prebuilt or custom.

🏏

Cricket analogy: Like a franchise scouting five different bowling actions to build a robust batting simulation rather than facing the same bowler five times, tagging varied document layouts trains a more generalizable custom model than five copies of one layout.

Calling AI Builder Actions from a Flow

In Power Automate, AI Builder actions plug straight into the standard designer flow: a trigger like 'When a file is created' feeds the file content into the 'Predict' or model-specific action, and the returned output can be parsed with 'Parse JSON' (using the schema AI Builder provides) to expose individual fields like invoice total or vendor name as dynamic content for later steps. Because AI Builder consumes 'AI Builder credits' (a capacity-based cost separate from standard Power Automate licensing, drawn from a per-tenant or per-environment credit pool), high-volume flows should batch or filter documents before invoking the model rather than calling it on every file unconditionally, and a Condition action checking the confidence score lets you send only low-confidence extractions to a human-in-the-loop review step, such as the AI Builder-provided 'Process and review' component in a model-driven Power App.

🏏

Cricket analogy: Like a stadium only running its expensive Hawk-Eye ball-tracking system for close-call reviews rather than every single delivery, a flow should filter which files actually need an AI Builder credit-consuming prediction call.

text

// Simplified flow logic: predict, then branch on confidence

Trigger: When a file is created in 'Incoming Invoices' folder
Action: AI Builder - Predict (Model: 'Custom Invoice Extractor')
         Input: File content
         Output: PredictionResult (JSON)

Action: Parse JSON  Content: PredictionResult
         Schema: { "invoiceTotal": {}, "vendorName": {}, "confidence": {} }

Condition: PredictionResult['confidence'] < 0.8
  If yes -> Create item in 'Human Review Queue' SharePoint list
  If no  -> Create item in 'Approved Invoices' table, Dataverse

AI Builder models can be built and tested for free in a trial or development environment, but production usage consumes AI Builder credits from a capacity add-on purchased separately from standard Power Automate per-user or per-flow licensing; check consumption in the Power Platform admin center before scaling a flow to run against a large document backlog.

A custom document processing model trained only on samples from one vendor's invoice layout will often extract confidently but incorrectly when it later encounters a genuinely different vendor's layout; always validate accuracy against a held-out test set covering every layout variation the flow will realistically encounter in production before trusting it to bypass human review.

AI Builder exposes prebuilt models (invoice extraction, object detection, sentiment analysis, OCR) that work with zero training as Power Automate actions.
Custom document processing models are trained by uploading and tagging sample documents, ideally covering multiple real layout variations.
Every AI Builder action returns structured JSON output with per-field confidence scores you can branch on.
AI Builder actions consume AI Builder credits, a separate capacity-based cost from standard Power Automate licensing.
Filter or batch documents before invoking a model to conserve the finite AI Builder credit pool.
Route low-confidence predictions to a human-in-the-loop review step rather than trusting every extraction automatically.
Validate a custom model's accuracy against varied test samples before relying on it to skip human review in production.