AI Builder vs Azure Document Intelligence

Last updated August 2026. Rates below were read from Microsoft's own documentation on learn.microsoft.com in August 2026.

Short answer: AI Builder and Azure AI Document Intelligence give you what is functionally the same prebuilt invoice model, returning the same fields. The difference is packaging and price. Azure AI Document Intelligence is the direct API. AI Builder is that same capability resold inside Power Platform with a workflow designer, Dataverse and admin governance wrapped around it, metered in credits. For a customer signing up in 2026, AI Builder works out to roughly $80 per 1,000 invoice pages against the $10 per 1,000 last published for the Azure prebuilt invoice model. You are paying about eight times more for the wrapper, not for better extraction.

The comparison in one table

Is AI Builder the same as Azure Document Intelligence?

Not the same product, but the invoice extraction underneath is the same work. The clearest evidence is the output schema. AI Builder's prebuilt invoice model returns 31 header fields plus nested collections for payment details, tax details, installments and line items. That list matches the Azure AI Document Intelligence prebuilt invoice schema field for field, including two fields nobody would independently arrive at: KVKNumber, a registration identifier specific to businesses in the Netherlands, and BPayBillerCode, an Australian bill payment identifier.

Two teams building invoice models separately do not both decide to ship a Dutch chamber of commerce number and an Australian BPay reference in the default field set. The practical takeaway for a buyer is simple: do not expect an accuracy difference between the two on the same invoice. Evaluate them on cost, on integration, and on what happens when a document fails.

Which is cheaper, AI Builder or Azure Document Intelligence?

Azure, by a wide margin, for the extraction itself. AI Builder prices document work at 8 Copilot Credits per page. At the pay-as-you-go rate of $0.01 per Copilot Credit that is $0.08 a page, or $80 per 1,000. A prepaid pack at $200 a month for 25,000 credits brings the effective rate to $0.064 a page, or $64 per 1,000.

There used to be a much cheaper path. AI Builder capacity add-ons priced invoice pages at 32 AI Builder credits, and the Tier 1 add-on gave you 1,000,000 credits for $500 a month, which works out to $0.016 a page or $16 per 1,000. Microsoft closed that add-on to new customers on November 1, 2025. If you did not already hold one, you cannot buy it. The full ladder is broken down on our AI Builder pricing page, including what each Power Platform license actually seeds and what happens on November 1, 2026.

One caveat on the Azure number, and we would rather flag it than paper over it. As of August 2026 the Azure AI Document Intelligence pricing page renders every rate as a dash instead of a figure and directs you to sign in to the pricing calculator. The $10 per 1,000 pages figure is the last one we were able to verify from a published source. Confirm it against your own agreement in the calculator before you build a business case on it. We track what we can verify on our Azure Document Intelligence pricing breakdown.

What you actually get for the difference

It would be unfair to present the gap as pure markup. AI Builder buys you things that have real value if you need them.

• No code between the document and the process. An invoice lands in a SharePoint library, a flow fires, fields come back, a row appears in Dataverse and an approval goes out. Nobody wrote a service to do that.
• Tenant governance. Data loss prevention policies, environment routing, audit logs and admin controls over which environments can run AI at all. In a regulated business this often decides the question before price is discussed.
• Confidence scores you can branch on. Every field comes back scored, and a flow can route anything under a threshold to a human without custom code.
• Raw OCR on the same call. Alongside the named fields you get the full detected text and every key-value pair, so a field the model does not name is still reachable without paying for a second pass.

The reason teams reach for AI Builder in the first place is rarely the extraction quality. It is that the invoice arrives inside a process that already knows how to route an exception to the right approver, and rebuilding that layer around a raw API is weeks of work nobody budgeted for.

Where AI Builder gets awkward

Three things catch teams out, and none of them appear in a licensing deck.

Capacity does not roll over, and running out blocks you. Credits reset on the first of the month and unused capacity is forfeited, so you have to size for your busiest month and pay for idle headroom the rest of the year. When you exceed capacity, Microsoft does not bill overage. It blocks the environment, and flows fail with EntitlementNotAvailable, NoCapacity or QuotaExceeded. Microsoft states the consumption figure that triggers the block is computed on a lag that can exceed five days, so the stop arrives with no warning. Azure just keeps processing and keeps charging, which is the failure mode most AP teams would prefer.

The credits are not the whole bill. AI Builder actions are premium connectors, so the premium Power Platform licensing underneath is required regardless. On pay-as-you-go, each flow run also bills through the cloud flow run meter, separately from and in addition to the credits the model consumes. Worth noting: Microsoft's own two licensing articles currently contradict each other on whether an AI Builder action makes a flow a premium flow. Both agree it makes an app a premium app. If that distinction moves your total materially, get the answer in writing.

The locale list is narrower than it looks. The prebuilt invoice model publishes 36 supported locales. It covers English for the US, Canada, the UK, Australia and India, and it covers Spanish (Spain), Portuguese (Portugal) and French (France). It does not list Spanish for Mexico or the United States, Portuguese for Brazil, or French for Canada. A US importer buying from Monterrey or Sao Paulo is outside the published support matrix, which is an odd gap for a product sold into North America.

November 1, 2026 forces the decision anyway

Microsoft announced the progressive end of AI Builder credits in October 2025. The features stay; the currency is being retired on a published calendar.

There is no automatic migration and no conversion ratio between the two currencies. Seeded credits are removed, not converted. Contracts that started before November 1, 2025 are honored for their full term, so an enterprise agreement may buy you time, but the direction is settled. If your invoice automation currently runs on seeded credits you already own, you are on a clock, and the replacement costs five times as much per page.

Can AI Builder extract invoice line items?

Yes, and this is a genuine strength worth stating plainly. Line items come back with amount, date, description, quantity, product code, tax, tax rate, unit and unit price, each with its own confidence score. That is real line-item extraction, not header-only capture, and it is better than what several accounting packages manage natively. QuickBooks Online's own receipt and bill capture, by contrast, reads header fields and does not return line items at all.

Do I need Azure to use AI Builder?

No. AI Builder is licensed and consumed entirely through Power Platform. You need an environment with Dataverse and the appropriate premium licensing, and capacity in either AI Builder credits or Copilot Credits. You never touch an Azure subscription, never provision a resource and never handle an API key. That convenience is a large part of what the price premium is buying.

How to choose

Pick Azure AI Document Intelligence if you have developers, your volume is meaningful, and you want the lowest per page cost with no capacity cliff. You will build the orchestration, the retry logic and the exception queue yourself, and at a few thousand pages a month the savings pay for that quickly.

Pick AI Builder if your volume genuinely fits inside credits you already hold, the value is in the approval workflow rather than the extraction, and keeping every document inside your Microsoft tenant is a compliance requirement. Just do the arithmetic on your busiest month first, and check what happens to your seeded credits after November 1, 2026.

Pick neither if what you actually need is invoices turned into a spreadsheet or an accounting import without building anything. That is a different shape of problem. If your volume is a few thousand pages a month and the destination is Excel, CSV or your ledger, a flat plan is both cheaper than either option above and considerably less work: our Plus plan works out to $7.40 per 1,000 pages on annual billing, with no credits to allocate, no premium connectors and nothing that blocks mid-month. You can drop a folder of invoices in at once with bulk invoice upload, or wire it into your own stack through the invoice data extraction API.

If you are already committed to Power Platform and want the flow-based approach done properly, our walkthrough on Power Automate invoice extraction covers building the flow itself, and the AWS Textract pricing breakdown covers the third major cloud option at $10 per 1,000 pages for Analyze Expense.