AI Builder explained

AI Builder is the no-code machine-learning capability inside the Power Platform. It exists to put trained ML models in the hands of makers — citizen developers, business analysts, power users — without needing data scientists, Python notebooks, or Azure ML knowledge. For Dynamics 365 customers, AI Builder is the cheap, fast way to add AI to a Power App or Power Automate flow without an enterprise data-science programme.

Pre-built models

Microsoft ships ready-to-use models for common business tasks:

• Form processing — extract structured data from invoices, receipts, IDs, and custom forms with a UI for training on samples.
• Object detection — locate and count objects in images (boxes, pallets, equipment).
• Sentiment analysis — score text for positive/negative/neutral tone across many languages.
• Key phrase extraction — pull out salient terms from text.
• Language detection — identify the language of a document.
• Text translation — translate between languages.
• Business card reader — extract name, title, email, phone from a card image.
• Receipt reader — extract structured line items from receipt photos.
• Text recognition (OCR) — extract printed and handwritten text from images.
• Category classification — classify text into a custom taxonomy.

Custom models

Beyond the pre-built set, AI Builder supports custom-trained models for prediction, category classification, entity extraction, and form processing. Training is point-and-click: provide labelled data from Dataverse or a file, the system trains, and reports accuracy. Re-training on a schedule keeps models current.

Where models run

Trained models live as AI Builder assets in an environment. They're invoked from:

• Power Automate flows — call the model as an action; pass inputs, receive outputs.
• Power Apps — invoke the model from a button or screen.
• Dataverse plug-ins — call from server-side logic.
• Direct APIs — for programmatic use.

Integration with Dynamics 365

Common use cases:

• Invoice automation — receipts and invoices arrive in a shared mailbox, AI Builder extracts data, a flow posts a draft vendor invoice in Business Central or F&O.
• Lead enrichment — AI Builder predictions enrich incoming leads with a quality score.
• Customer feedback — sentiment analysis on case descriptions, with auto-prioritisation.
• Inspection — object detection during field service inspections flags missing or defective components.

A worked example: vendor invoice OCR

The canonical AI Builder story is invoice OCR. The end-to-end shape:

1. Vendors send PDFs to invoices@company.com. A shared mailbox is a monitored source in Power Automate.
2. A cloud flow triggers on new email, filters on attachments with a PDF content type, and drops them into a SharePoint library keyed by month.
3. The flow calls an AI Builder custom form-processing model trained on samples from your biggest 20-30 vendors. It returns invoice number, invoice date, currency, subtotal, VAT, total, purchase-order reference (when present), and a line-item collection.
4. A Dataverse row is created in a staging table with confidence scores per field. Any field under a threshold (default 0.75) routes the row to a review queue as a model-driven app; anything above threshold posts straight through.
5. The flow then calls the Business Central or Finance and Operations API to create a draft vendor invoice with the extracted lines, and stamps the source PDF onto the record so an approver can eyeball the original.

The two things that make this work are the confidence threshold and the reviewer app. Without a threshold, one bad extraction posts a wrong number to the ledger. Without a reviewer UI, the "human in the loop" becomes an email chain no one owns.

Custom models

Beyond the pre-built set, AI Builder supports custom-trained models for prediction, category classification, entity extraction, and form processing. Training is point-and-click: provide labelled data from Dataverse or a file, the system trains, and reports accuracy. Re-training on a schedule keeps models current.

The three custom-model tips that trip most projects:

• Minimum training set is small (5 samples per form template for form processing; the docs give exact numbers per model) but that is a floor, not an aspiration. Aim for 30-50 varied samples per template — mixed layouts, different vendors, degraded scans, and at least one non-English document if the process handles them.
• Drift is real. A vendor rebrands their invoice, a new tax field appears, and accuracy quietly falls. Track per-field confidence in Application Insights and alert when the 7-day average drops.
• Environment strategy matters. A model trained in a developer environment does not automatically move to production; it moves via a solution export/import. Bake this into the ALM pipeline from day one, not after the first "why is my model not there" ticket.

Limits and when to reach for Azure ML or Azure OpenAI

AI Builder is not a substitute for Azure ML or Azure OpenAI for sophisticated requirements. Models are limited in size, training data, and customisation. Reach for the Azure services when:

• Model complexity is beyond the templates. Time-series forecasting on retail SKUs, computer-vision defect detection at pixel-level accuracy, or graph-based fraud detection — none of those fit the AI Builder templates. Azure ML gives you the notebook and the compute.
• The right answer is a language model, not a classifier. Summarising a case, drafting a response, extracting free-text fields from unstructured documents, or building an internal Q&A assistant — those are Azure OpenAI (or, from inside Copilot Studio, the built-in generative actions). AI Builder's text models are classifiers, not generators.
• Data cannot leave a specific region or subscription. AI Builder inherits the Power Platform environment's region; if compliance says the training data lives in a specific Azure subscription in a specific region, run it there.
• Volume pushes credit cost above the alternative. At a few thousand extractions per day AI Builder is cheap and effortless. At tens of thousands per day the credit maths often favours a Document Intelligence resource in Azure billed under consumption.

Licensing and credit consumption

AI Builder is metered in service credits. Every model call consumes a number of credits — form processing is a few credits per page, prediction is a fraction of a credit per row, GPT-style generative actions consume more heavily. Credits come from three places:

• A standalone AI Builder capacity add-on in the Power Platform admin centre, purchased in packs (the exact pack size varies over time).
• Entitlements bundled with certain Power Platform premium licences — the specific allocation is in the Licensing Guide, changes with waves, and is worth re-reading each release.
• Copilot Studio and per-user Copilot licences for the generative actions those products invoke on your behalf.

Two rules from the field:

• Reserve capacity per environment. Without a reservation, one runaway flow in dev burns the tenant's shared pool. Reservations let you cap risk and let admins see who used what.
• Instrument every flow that calls a model. Log the model, the input size, and the returned confidence to Application Insights. The alternative is a monthly credit surprise no one can trace.

Governance and the Copilot overlap

AI Builder used to be the only Microsoft no-code AI on the Power Platform. It now sits alongside Copilot Studio agents and the generative AI actions available inside cloud flows and canvas apps. The line is:

• AI Builder for a specific model call embedded in a flow or an app — extract fields from this PDF, score this lead, classify this text.
• Copilot Studio for a conversational agent that reasons across topics, tools, and knowledge sources.
• Generative AI actions in Power Automate for a one-off "call GPT with this prompt" step inside an otherwise deterministic flow.

They share credit pools, and the Power Platform admin centre reports usage across all of them together — worth remembering when the monthly credit chart spikes and the invoice-OCR flow is not the culprit.

FAQ recap

The full FAQ is at the top of the page; the two most-asked questions in the field are (1) "will AI Builder be replaced by Copilot Studio?" — no, they solve different problems, and Microsoft has kept improving both — and (2) "how do I stop credit runaway?" — reserved capacity plus per-environment DLP that blocks the AI Builder connector from experimental environments.