Most organizations that want to add AI to their workflows face the same problem: the tools built for data scientists are too complex, but the marketing promises of “AI in one click” don’t deliver anything useful in the real world. AI Builder sits in a narrow but practical middle ground — and understanding what it actually does (and doesn’t do) is worth your time if you work inside the Microsoft ecosystem.

What AI Builder is Link to heading

AI Builder is Microsoft’s low-code AI layer built into Power Platform. It gives Power Apps and Power Automate access to pre-built and custom-trained AI models without requiring you to write code, manage infrastructure, or understand machine learning pipelines.

The key word is “integrated.” AI Builder isn’t a standalone AI tool — it’s a set of models you call from within your existing Power Platform workflows. An invoice arrives by email → Power Automate triggers → AI Builder reads the PDF and extracts fields → data lands in SharePoint. That’s the pattern.

It’s available as part of Power Platform Premium licensing, with additional AI Builder capacity measured in “service credits” — more on that later.

What it can actually do Link to heading

AI Builder covers two broad categories: pre-built models that work immediately and custom models you train on your own data.

Pre-built models Link to heading

These require zero training and are ready to use:

  • Invoice processing — extracts vendor name, invoice number, date, line items, VAT, and total from PDFs. Works across many invoice formats out of the box. Not perfect on highly unusual layouts, but handles the 80–90% of standard supplier invoices reliably.
  • Receipt processing — similar extraction for retail receipts.
  • Document classification — takes a document and tells you which category it belongs to (requires a small labeled training set, but significantly less than building from scratch).
  • Business card reader — extracts contact details.
  • Text recognition (OCR) — general-purpose text extraction from images.
  • Sentiment analysis — classifies text as positive, negative, or neutral.
  • Key phrase extraction — pulls out the main topics from a block of text.
  • Language detection — identifies the language of a text input.
  • Entity extraction — identifies named entities (people, places, organizations) in text.

The document-oriented models — invoice processing and OCR in particular — are where AI Builder earns its place. They’re accurate enough for real production use and drop directly into Power Automate flows without any setup.

Custom models Link to heading

For cases where pre-built doesn’t fit:

  • Custom document processing — you label your own document type (a specific contract format, a custom form, an internal report template) and the model learns to extract the fields you define. Takes 5 samples minimum to train, but more training data gives better results.
  • Custom prediction — trains a binary or multi-class classifier on your own tabular data from Dataverse. Useful for scenarios like “is this sales opportunity likely to close?” or “does this support ticket need escalation?” — provided you have enough historical labeled data.
  • Object detection — identifies and locates specific objects in images. Less commonly needed in business automation.
  • Category classification — text classification trained on your own categories and examples.

The custom document processing model is genuinely useful and underused. If your organization has a proprietary form — a purchase order format, an internal approval document, an industry-specific report — you can train a model to extract fields from it in an afternoon, without any ML background.

Where it works well Link to heading

High-volume, repetitive document intake Link to heading

The clearest win for AI Builder is any process where the same type of document arrives repeatedly and someone is manually copying data out of it. Invoice processing is the canonical example, but it applies equally to:

  • Expense receipts submitted by field teams
  • Delivery notes from suppliers
  • Application forms scanned from paper
  • Inspection reports from technicians

If you’re processing more than 20–30 documents a week manually, the automation almost always pays for itself quickly.

Enriching existing Power Platform workflows Link to heading

Because AI Builder plugs directly into Power Automate, it adds AI capability to flows you already have. A flow that saves email attachments to SharePoint can be extended — with a single additional step — to also classify the document type, extract key fields, and route the file to the right folder. The AI capability slots in without rebuilding anything.

Power Apps with embedded AI Link to heading

In canvas apps, AI Builder models are available as components. A field technician app can include a camera control that feeds directly into a custom document processing model, extracts values from a photographed form, and pre-fills the submission fields automatically. This works without any code — just connecting controls in the Power Apps studio.

Where it struggles Link to heading

Credit consumption and cost Link to heading

AI Builder runs on a credit system. Every model invocation consumes credits, and credits are finite based on your license. Pre-built model calls cost 1 credit each; custom model runs can cost more. For low-volume use, the included credits in Power Apps/Power Automate per-user plans cover most needs. For high-volume automation — thousands of documents per month — you need to purchase additional capacity, and costs scale meaningfully.

It’s worth modeling your expected monthly volume before committing. A flow processing 500 invoices a month uses credits at a rate that can surprise teams who didn’t plan for it.

Model accuracy on unusual documents Link to heading

Pre-built models are trained on broadly representative document sets. They perform well on standard European and US invoice formats. They start to struggle on:

  • Handwritten documents
  • Low-quality scans or photos taken in poor lighting
  • Documents in non-Western layouts or scripts
  • Highly customized formats with non-standard field positions

When accuracy matters and documents are unusual, the custom document processing model (trained on your specific format) consistently outperforms the pre-built equivalent.

The Dataverse dependency for custom prediction Link to heading

The custom prediction model pulls training data from Dataverse — not from Excel, SharePoint lists, or external databases. If your historical data lives elsewhere, you need to import it into Dataverse first, which adds friction and sometimes requires data modeling work that wasn’t planned.

Not a replacement for Azure AI or custom ML Link to heading

For complex, high-stakes predictions — fraud detection, demand forecasting, churn modeling across millions of records — AI Builder isn’t the right tool. Its custom prediction model is accessible but limited in the sophistication of what it can learn. Teams with real ML requirements, large datasets, or the need for model explainability beyond what Power Platform provides should look at Azure Machine Learning instead. AI Builder’s value proposition is specifically low-code AI for business process automation, not general-purpose ML.

A practical example: document classification + invoice OCR Link to heading

Here’s how a real automation combining AI Builder’s document classification and invoice processing models works:

  1. Trigger: email arrives in a shared inbox with a PDF attachment
  2. Document classification: the PDF is passed to an AI Builder classification model trained on the organization’s document types (invoices, delivery notes, contracts, correspondence). If classified as “not invoice,” the flow routes it to a holding folder and stops.
  3. Invoice processing: if classified as an invoice, it goes through the pre-built invoice processing model, which returns structured fields: vendor, number, date, line items, VAT, total.
  4. Data write: the extracted fields are written to a SharePoint list row with the original PDF attached.
  5. Notification: Power Automate sends a Teams message to the finance team with a link to the new record.

Total development time for an experienced Power Platform developer: a few hours. The flow handles a document type that previously required manual data entry for every instance.

The accuracy depends on invoice variety. For a supplier base with consistent formats, expect 85–95% accuracy with minimal manual corrections. For a very diverse supplier base with unusual layouts, you’ll catch more errors and may need a human review step before data is committed.

Comparing to alternatives Link to heading

AI Builder Azure AI Document Intelligence Custom ML (Azure ML)
Technical barrier Low — no code Medium — API calls High — ML expertise needed
Integration with Power Platform Native Requires connector/custom code Requires custom connector
Accuracy on standard docs Good Better Depends on model
Customization Limited High Full control
Cost model Credit-based, included with license Per-page pricing Infrastructure + compute
Best for Business users, Power Platform teams Dev teams needing accuracy Data scientists, complex requirements

If you’re already running Power Platform and need document AI for standard business documents, AI Builder is the pragmatic starting point. If you need higher accuracy on unusual formats, or are working with a development team comfortable with APIs, Azure AI Document Intelligence offers more control. Custom ML is for when neither of these fits.

The bottom line Link to heading

AI Builder doesn’t try to be everything. It’s a set of purpose-built AI models integrated directly into the tools that business users and low-code developers already work in. The pre-built document models — invoice processing especially — are accurate enough for real production workflows. The custom model training is accessible to people without ML backgrounds, given enough representative training data.

Its limitations are real: the credit model adds cost at volume, the custom prediction model requires Dataverse, and it’s not the right tool when accuracy requirements are very high or the problem is genuinely complex. But for the category of problems it targets — automating repetitive document handling inside the Microsoft ecosystem — it works, and it works without needing a data scientist in the room.

That’s a harder combination to find than it sounds.

Sources: AI Builder documentation — Microsoft Learn · AI Builder licensing · Azure AI Document Intelligence