Azure AI Vision vs Document Intelligence for OCR: Which Azure Service Should You Use?

Short answer: use Azure AI Document Intelligence for scanned and digital documents, and Azure AI Vision only for general images such as photos, signs, posters and product labels. That is not our opinion, it is Microsoft's. Its OCR overview states plainly: "If you're extracting text from scanned and digital documents, use Document Intelligence Read OCR." Both services cost about $1.50 per 1,000 at the standard rate, so the choice is about the response shape, not the price.

The reason this question keeps coming up is that there is no product called "Azure OCR". There are two services that read text, they share the same underlying engine, and Microsoft has renamed both of them at least once. Here is how they actually differ, checked against Microsoft's own documentation and price feed on 17 August 2026.

Is Azure AI Vision the same as Azure Document Intelligence?

No. They are separate services on separate meters that happen to share the same Read OCR engine. Microsoft describes Document Intelligence as including "a document-optimized version of Read as its OCR engine while delegating to other models for higher-end insights". Vision is an image API billed per transaction. Document Intelligence is a document API billed per page.

The practical consequences follow from that framing rather than from accuracy. Vision gives you a synchronous call that returns text and bounding boxes, which is easy to put behind a screen where somebody is waiting. Document Intelligence gives you an operation you submit and poll, which is what you want for a 200-page loan file, and its response carries paragraphs, tables, selection marks and, on the prebuilt and custom models, named fields.

The third option Microsoft tells you not to use

There is a legacy path, and it still appears in tutorials, Stack Overflow answers and running production code. Microsoft's OCR overview opens with a warning: "We don't recommend using this service, including the Azure Vision in Foundry Tools legacy OCR API v3.2 and RecognizeText API v2.1." The same page states that "All future Read OCR enhancements are part of the two services listed previously. There are no further updates to Azure Vision v3.2."

Here is where it gets awkward, and it is the single most useful thing in this article. On the Azure AI Vision pricing page, Image Analysis features are priced in groups. The feature literally named OCR sits in Group 1. The feature named Read sits in Group 2, alongside Describe, Caption and Dense Captions. Group 1 costs $1.00 per 1,000 transactions. Group 2 costs $1.50.

So the cheaper Azure Vision text meter is the deprecated one. A team optimizing on price alone, reading only the pricing page, will pick the $1.00 feature and build on a frozen API. On a million pages a year that choice saves $500 and buys a migration later.

Per 1,000 transactions, East US, from the public Azure Retail Prices API on 17 August 2026. Group 2 has a single break at one million, so its rate is flat above that.

Which is cheaper, Azure AI Vision or Document Intelligence?

For plain text, neither. Both land on $1.50 per 1,000 at the standard rate and both fall to $0.60 above a million a month. That symmetry is deliberate: it is the same engine, priced the same way, wrapped in two different products.

The price only diverges when you ask for structure, and only Document Intelligence offers it. Tables, selection marks and paragraph roles mean Layout, and Layout is billed as a prebuilt model at $10.00 per 1,000 pages, the same as invoice extraction. Custom extraction is $30.00. So the real cost question on Azure is never Vision against Document Intelligence, it is Read against everything else, and that decision is worth 6.7 times.

When Azure AI Vision is genuinely the right pick

Vision wins on interaction, not on documents. If a user is holding a phone at a shelf label, a serial plate, a parking sign or a whiteboard and expects a result in under a second, the synchronous call is the right shape and the async poll is not. Microsoft's own framing for the version 4.0 image edition is that it is "optimized for general, non-document images with a performance-enhanced synchronous API that makes it easier to embed OCR in your user experience scenarios."

Vision also makes sense when text is one of several things you want from the same image. If you are already paying a Group 1 transaction for object detection or tagging, adding text extraction to that pipeline is cheaper and simpler than standing up a second service.

When Document Intelligence is the only sensible answer

Anything that arrives as a file rather than a camera frame. Invoices, statements, claims, contracts, tax forms, shipping paperwork. Three things push you here and none of them are accuracy:

• Page counts. Documents come in tens and hundreds of pages, which is what the asynchronous operation model exists for.
• Structure. A total sitting inside a table cell is a different problem from a total sitting somewhere in a wall of text, and only Document Intelligence models the table.
• Named fields. The prebuilt models return an invoice number as an invoice number. Vision returns it as a string with coordinates and leaves you to work out which string it was.

That last point is where most Vision-for-documents projects quietly go wrong. The OCR is fine. The 400 lines of regex and positional heuristics written on top of it are what breaks when a vendor changes their template.

Do both services bill the same way for a PDF?

No, and this trips people up on the invoice. Document Intelligence bills by pages analyzed, and Microsoft's own billing note is explicit: "When a document is submitted for analysis all pages are analyzed unless a page range is specified with the pages parameter in the request." So a 40-page PDF is 40 pages even if you wanted page 3, and passing a page range is the cheapest optimization available on Azure.

Azure AI Vision bills per transaction, and a transaction is a call for a feature. If you request two features on one image, that is two transactions. Vision's image edition is aimed at single images rather than long files, which is another way of saying it was never designed for the 40-page PDF in the first place.

What about Azure Content Understanding?

It is a third, newer service, and it is not a replacement for either of these. Content Understanding takes documents, images, audio and video and returns fields you describe rather than fields you train. If you are weighing it against the incumbent, we have a dedicated breakdown of Azure Content Understanding vs Document Intelligence with the rates side by side.

The useful mental model for the whole Azure lineup: Vision reads pictures, Document Intelligence reads documents, Content Understanding reads mixed media and describes it. All three bill per unit of input and none of them ship the pipeline around the call.

How to decide in one pass

Ask what arrives. If it is a camera frame and a person is waiting, use Azure AI Vision and its Read feature, not the Group 1 OCR feature. If it is a file, use Document Intelligence, start with prebuilt-read, and only move to Layout when you have proven you actually need the table structure, because that move costs 6.7 times as much per page.

Then price it properly. Committed spend and per-model meters make an Azure document bill harder to predict than the headline rate suggests, and teams running several clouds usually end up wanting a read-only view of what each service is actually costing them rather than a spreadsheet somebody updates when they remember. The full meter-by-meter breakdown for both services, with the commitment tier break-evens, is on our Azure OCR pricing page.

One last piece of honesty about all of it. Choosing correctly between these two services gets you a good OCR response. It does not get you extracted data you can post into a system of record. Between the API call and that outcome sit the queue, the retries, the confidence threshold, the review screen, and the export. On most projects that surrounding work costs more than the meter does, whichever Azure service you land on. If you would rather test that end state than build it, our OCR API pricing comparison puts Azure next to the alternatives on the same per-page basis.