AI Builder Credits End for New Licenses: Copilot Costs and Azure Document Intelligence Options

What the roundup covers​

The guide sorts document-extraction products into enterprise IDP suites, cloud APIs, AI-native parsing libraries, accounts-payable tools, RPA platforms and small-business tools. Its practical advice is sensible:

• Start with the documents you actually process.
• Decide where the extracted data has to end up.
• Define an acceptable error rate and who reviews low-confidence results.
• Count total cost of ownership, not only the subscription.
• Budget weeks, not days, for calibration.

It also tells readers to ignore vendor accuracy figures and run a proof of concept on their own documents. That is the most useful line in the piece.

Treat the rest with caution. The copy I received says it covers 32 tools, but the live page reportedly describes 46 options, so don't rely on either number. The vendor claims on accuracy, pricing, deployments and market position are the publisher's assertions. I haven't independently benchmarked or verified them.

The licensing claim that is now wrong​

The guide says AI Builder includes "5,000 AI Builder credits" per Microsoft 365 plan and quotes a Premium price. Neither matches Microsoft's current licensing documentation. The seeded credits attach to certain Power Platform and Dynamics 365 licenses, not Microsoft 365 plans generally, and they are ending.

Microsoft's licensing overview says that starting November 1, 2026, no new or renewed Power Platform or Dynamics 365 licenses will receive seeded AI Builder credits. Licenses bought or renewed before that date keep their seeded credits through the contract term. Microsoft also says AI Builder capabilities will continue to be supported through Copilot Credits. The features aren't being retired. The billing currency is changing.

The practical points for admins:

• No automatic conversion. Microsoft says there is no automatic conversion of AI Builder credits into Copilot Credits.
• New customers. Microsoft's licensing page says AI Builder capacity add-ons are only available as a renewal or true-up for existing customers. New customers must buy Copilot Credits to run AI Builder features.
• Fallback behavior. In apps and flows, usage tries AI Builder credits first and then Copilot Credits. If neither is available in the environment, model runs are blocked.
• Building is free. Building and testing models doesn't consume credits. Running them in apps, flows or agents does.
• Premium flows. A Power Automate flow that uses any AI Builder action is a premium flow. It needs premium Power Automate licensing, and pay-as-you-go runs are billed separately from credit consumption.
• No rollover. Consumption resets monthly and unused credits don't roll over. Size capacity to your peak month.

Microsoft's rate table lists receipt, invoice and identity-document analysis at 32 AI Builder credits per page, or 8 Copilot Credits per page. Custom document processing is 100 AI Builder credits per page, or 8 Copilot Credits. Microsoft's table puts one Copilot Credit at $0.01 under pay-as-you-go, which is about 8 cents per page for those actions. One consultancy notes that this can cost considerably more than before, depending on usage. Model that against your real volumes before renewal.

I'm not repeating the guide's quoted monthly price, because I couldn't verify it against plan-specific Microsoft pricing.

Azure AI Document Intelligence: the developer route​

Azure AI Document Intelligence is the service formerly called Form Recognizer. Microsoft documents three groups of models:

• Read and Layout. Read handles printed and handwritten text. Layout extracts text, tables, selection marks and structural elements such as titles, headings, headers and footers.
• Prebuilt models. These cover invoices, receipts, identity documents, tax and mortgage documents, contracts, bank statements and pay stubs.
• Custom models. Microsoft says a custom extraction model needs only five examples of the same form type to get started. Custom classification needs at least two classes and five samples per class. A composed model can bundle up to 200 trained custom models behind one model ID.

Limits from Microsoft's current model overview:

• PDFs and TIFFs: up to 2,000 pages.
• File size: 500 MB on the paid S0 tier and 4 MB on the free F0 tier.
• Free tier: only the first two pages of a document are processed.
• Text height: a minimum of 12 pixels on a 1024 x 768 image, roughly 8-point text at 150 dpi.

Five examples is a starting point, not a promise of production accuracy. The service is also multitenant and asynchronous. Microsoft says latency can vary at runtime, which matters if you promise users near-instant results.

For no-code use, Microsoft publishes a Power Automate and Logic Apps connector. You need a Document Intelligence resource in the Azure portal, plus its endpoint and key. Entra ID and managed identity authentication are also supported. The connector offers v3.x and v4.x document-analysis operations and lets you select pages, for example "2, 5-7", so you don't process an entire PDF. Several older per-model operations are marked deprecated, so new flows should use the newer analyze-document operations. Microsoft's own advice for a first test is to pick a prebuilt model such as prebuilt-invoice or prebuilt-layout, supply a PDF or image, and inspect the detected fields.

AI Builder and Power Automate: the low-code route​

AI Builder's prebuilt invoice model returns fields including invoice ID, vendor and customer details, dates, totals, tax and line items. Microsoft recommends one clear photo or high-quality scan per document. In Power Apps you can call it from Power Fx with the Predict function. A basic canvas-app setup looks like this:

1. Create a canvas app and add the "Invoice processing" data source.
2. Insert a picture control for the upload.
3. Insert a text label and point it at the predicted field you want, such as the invoice ID.
4. Save and run the app to test.

For custom models, Microsoft describes three types:

• Fixed template documents. For layouts where fields sit in similar places. Training is quick.
• General documents. For complex or unstructured formats. These are more powerful but take longer to train.
• Invoices. For extending the prebuilt invoice model with extra fields.

If you train on your own documents, you need at least five samples per layout, and each layout is a "collection". A model can hold up to 200 collections. Microsoft says models now run on Document Intelligence v4.0 or v3.1, depending on when you last edited them. v4.0 adds overlapping fields, signature detection and confidence scores for tables. You can move a v3.1 model to v4.0 by editing, retraining and republishing, with no re-tagging required.

My earlier notes from Microsoft's documentation include some further constraints: a 20 MB input cap, image dimensions between 50 x 50 and 10,000 x 10,000 pixels, a 300-field tagging limit in a cloud flow, and unlocking password-protected PDFs first. They also say fields split across page boundaries and table rows that break across pages are not currently supported. I haven't re-checked these particular numbers here, so confirm them in current documentation before you design around them. If your documents contain long multi-page tables, test early. That is where a clean demo tends to fall apart.

A third option: the document automation toolkit​

Microsoft also offers a document automation toolkit that combines four pieces:

• Power Automate orchestrates the process.
• AI Builder extracts the information.
• Power Apps provides manual review and approval screens.
• Dataverse holds the queue, files, data and configuration.

This matters because extraction alone is rarely the whole job. An invoice with a low-confidence total needs a human, a queue and an audit trail. The toolkit gives you that inside the Microsoft stack. It is one architecture, not the only one. Note also that AI Builder licensing requires a Power Platform environment with Dataverse.

Choosing between the Microsoft options​

This table is my own synthesis of Microsoft's documented capabilities. It is not a claim that either Microsoft product beats its competitors.

What to do before you commit​

• Check your contract dates. Know whether your Power Platform and Dynamics 365 licenses were bought or renewed before or after November 1, 2026, and what seeded credits you still hold.
• Estimate pages, not users. Credit consumption is per page or per token, so volumes drive cost.
• Build a hard-case test set. Include handwriting, bad scans, multi-page tables and weird vendor layouts.
• Define review rules. Decide the confidence threshold below which a person looks at the result, and who that person is.
• Verify connector availability. Confirm the connector and your chosen model work in your tenant and region.

Bottom line​

The Dynamic Business guide is a decent way to see how wide the IDP field is, and its advice to test on your own documents is sound. For Microsoft shops, the more important point is the one the guide gets wrong. Seeded AI Builder credits are going away for new and renewed licenses, and Copilot Credits are the replacement. Document extraction from Azure and Power Platform still works. The cost model is what has changed, so recheck it before your next renewal.