PurchaseOrders is an Azure Document Intelligence alternative built only for purchase orders. Azure has no prebuilt purchase order model: Microsoft routes POs through the prebuilt invoice model, whose schema is shaped for accounts payable, or tells you to label and train a custom one. Here the PO fields are already defined, output is Excel or CSV as well as JSON, and there is a free tier. Try it on a real PO below.
Submit your purchase orders
Drop documents here, or click to file
Up to 50 files per batch
Uploading...
Azure AI Document Intelligence, renamed from Azure Form Recognizer in July 2023, is a strong and actively developed service with an unusually broad prebuilt model catalog. For purchase orders specifically, the gaps are concrete. Details below reflect Microsoft's published documentation and East US pricing as of July 2026.
Microsoft ships prebuilt models for invoices, receipts, IDs, contracts, bank statements, checks, pay stubs, W-2s, and mortgage forms. There is no purchase order model. Microsoft's guidance is to run POs through the invoice model, or to train a custom one.
The prebuilt invoice schema is accounts payable shaped: InvoiceId, DueDate, AmountDue, RemittanceAddress. It does have a PurchaseOrder field, but Microsoft defines it as a purchase order reference number printed on an invoice. There are no PO-native fields for requested delivery date, buyer, or ship-via.
Custom neural models need a minimum of five labeled samples, and Microsoft advises five per layout variation. A multi-vendor PO estate means labeling in Document Intelligence Studio for each supplier layout, storing the training set in Blob Storage, and retraining when a vendor changes their template. Custom extraction also costs $30 per 1,000 pages, three times the prebuilt rate.
The API returns JSON with confidence scores. No CSV or XLSX export is documented, so somebody flattens the nested line-item array into rows. Every call is asynchronous: you POST, get a 202 with an Operation-Location header, and poll. And the free F0 tier analyzes only the first two pages of any request, so it cannot evaluate a real PO.
PurchaseOrders is tuned for one document. The purchase order fields are already defined, the output is a spreadsheet if you want one, and there is nothing to label, train, or poll.
The PO number, supplier, ship-to and bill-to, order and requested delivery dates, payment terms, and the full line-item table come back named as what they are. You are not reading PO data out of a schema designed around InvoiceId and AmountDue.
No five samples per vendor layout, no Studio labeling sessions, no Blob Storage container for a training set, and no retraining when a supplier redesigns their template. A brand new vendor works on the first upload.
Download a spreadsheet with consistent headers and the line items already flattened into rows. JSON is available through the API, but you are not obliged to write the flattening code to get usable data.
No Azure subscription, no resource provisioning, no key management, and no polling loop with retry logic for throttled requests. Upload a purchase order in the browser and read the result on screen.
If the destination is a spreadsheet, the purchase order PDF to Excel converter gets you there in one step, and an importer-ready file comes from the purchase order PDF to CSV converter. Developers who want JSON use the purchase order API. The nested line-item table is exactly where these builds usually break, which is covered in purchase order line item extraction. Comparing the other cloud OCR services? Read the Amazon Textract alternative and Google Document AI alternative breakdowns. Microsoft shops pushing the result onward will also want purchase order data into Business Central, and the full field is lined up in the best purchase order OCR software.
An honest side by side for teams extracting purchase orders. Azure figures reflect Microsoft's published documentation and East US pay-as-you-go pricing in USD, verified July 2026.
| PurchaseOrders | Azure Document Intelligence | |
|---|---|---|
| Focus | Purchase orders only, pre-tuned | Broad document AI service |
| Prebuilt PO model | Yes, that is the product | None, POs go through the invoice model |
| Field schema on a PO | PO number, delivery date, buyer, ship-to | InvoiceId, DueDate, AmountDue |
| Custom model effort | None | Five labeled samples per layout, in Studio |
| Listed page rate | Per document, free tier | $10 per 1,000 prebuilt, $30 custom |
| Output | Excel, CSV, JSON, Sheets, API | JSON only, no documented CSV export |
| Call pattern | Upload and read the result | Async: POST, 202, poll for the result |
| Free tier | Reads a whole PO | F0 analyzes only the first 2 pages |
| Table extraction | PO line items as rows | Strong, with cell-level confidence |
| Best for | Teams that just need PO data | Azure engineering teams, many doc types |
Azure pricing and model details reflect Microsoft's published documentation and East US pay-as-you-go pricing in USD as of July 2026 and may change; confirm current rates at azure.microsoft.com. Azure AI Document Intelligence was renamed from Azure Form Recognizer in July 2023 and is not deprecated. Microsoft's invoice model documentation does list purchase orders among its supported document types; our point is that the extracted schema is invoice shaped, not that POs are rejected. Azure genuinely wins on breadth of prebuilt models, table extraction, Power Automate and AI Builder integration, US Gov regions, and disconnected containers. PurchaseOrders captures purchase order data and returns Excel, CSV, JSON, or an API response. It does not create POs, route approvals, perform three-way matching, or post to an ERP. PurchaseOrders is not affiliated with, endorsed by, or sponsored by Microsoft Corporation. Microsoft, Azure, Azure AI Document Intelligence, and Azure Form Recognizer are trademarks of Microsoft Corporation.
No Azure subscription, no Studio labeling, no polling loop. Upload a real purchase order and see the extracted data in seconds.
Drag in a PDF, scan, or photo of a PO from any supplier. There is no resource to provision, no key to manage, and no training set to label first.
Tip: Upload a PO from a supplier you have never processed before. There is no per-layout training.
The AI reads PO number, supplier, ship-to and bill-to, order and requested delivery dates, line items, SKUs, quantities, units of measure, unit prices, terms, and totals.
Export clean Excel, CSV, JSON, or Google Sheets, or send the data through the API into your ERP or accounting system.
No. There is no prebuilt purchase order model. Microsoft's invoice model documentation lists purchase orders among the document types it supports, so POs are accepted, but the schema you get back is invoice shaped: InvoiceId, DueDate, AmountDue. For PO-native fields, Microsoft's own guidance is to train a custom model.
Yes. Microsoft renamed Azure Form Recognizer to Azure AI Document Intelligence in July 2023, with no pricing change and no breaking API changes. It is now branded within Microsoft Foundry tools. The service is not deprecated: the current generally available REST API is v4.0.
On East US pay-as-you-go pricing in July 2026, Read OCR is $1.50 per 1,000 pages, Layout and prebuilt models including Invoice are $10 per 1,000 pages, custom extraction is $30 per 1,000 pages, and custom classification is $3 per 1,000 pages. Custom neural training is free up to a monthly allowance, then billed hourly.
Microsoft states you need only five examples of the same form or document type to get started with a custom neural model. In practice its own best-practice guidance is at least five labeled samples per layout variation, so a purchase order estate with a dozen supplier formats means a labeling pass for each one.
No. The documented outputs are JSON for all models, Markdown for layout, and searchable PDF for read. There is no CSV or XLSX export in the API. Note the common trap: Excel and Word files are supported input formats, not output formats, so turning PO line items into spreadsheet rows is code you write.
The free F0 tier allows 500 pages a month, but it analyzes only the first two pages of any request and caps files at 4 MB with one transaction per second. A three page purchase order returns two pages of data, so F0 is fine for a smoke test and not usable for evaluating real purchase orders.
Neither has a purchase order model, so both mean adapting an invoice or expense model or building a custom one. Azure has broader prebuilt coverage and stronger Power Platform integration; Textract is cheaper per page for expense parsing. If nobody on your team wants to own that pipeline, a PO-specific tool avoids the choice.
Compare with AWS Textract for POs.
Compare with Google Cloud Document AI.
Get PO data into Business Central.
See all the PO OCR tools compared.
How the line-item table is captured.
Z tej samej rodziny narzędzi
