PurchaseOrderOCRbyKoncile

AWS Textract alternatives for purchase orders: 7 tools compared (2026)

Published 2026-11-13 · By the Koncile team

The main AWS Textract alternatives for purchase orders are Koncile, Azure Document Intelligence, Google Document AI, Mindee, Veryfi, Nanonets and Parseur. Teams usually leave Textract because it has no purchase order model, so PO fields, the totals check, matching and a review screen all become code to write and maintain. Below: what Textract does well, what one PO really costs on it, and seven alternatives with their own weaknesses.

Hands holding purchase orders, surrounded by laptops displaying data analysis software, a calculator, and a pen.

The short answer

Amazon Textract is a set of OCR building blocks on AWS: it reads text, tables, key value pairs and invoices, and leaves the rest to your code. It has no purchase order (PO) model. Teams look for an alternative when the code around Textract (PO fields, the totals check, matching with the catalog or the invoice, a screen for doubtful fields) costs more than the OCR itself. Pick by the reason you are leaving:

  • You want PO lines, recomputed totals and matching in one response, priced per page: Koncile. It is our product, so read its weaknesses below with the same care as the others'.
  • You want to stay with a hyperscaler but get a model that knows POs: Azure Document Intelligence. Its prebuilt invoice model also lists purchase orders among the documents it reads.
  • You are already on Google Cloud: Google Document AI, with a custom extractor trained on your POs.
  • You want a typed PO API with SDKs: Mindee, or Veryfi if receipts and POs go through the same API.
  • You want the workflow too (approval, export): Nanonets. You only need a mailbox parser: Parseur.

Every fact below comes from the vendor's own site or documentation, read on 13 November 2026 unless another date is given. Prices exclude tax and are for the US region when the vendor shows one. For the full ranking of ten tools, see our comparison of the best purchase order OCR software.

What Textract does well

Textract is cheap, scales without a contract and stays inside your AWS account, which matters when documents must not leave it. AnalyzeExpense, its invoice and receipt API, returns standard summary fields (vendor, receiver, order date, delivery date, PO number, subtotal, tax, total) and line items with description, product code, quantity, unit price and total price normalized, at $0.01 per page for the first million pages a month. AnalyzeDocument reads any table, form or answer to a question you ask (Queries), and you can train an adapter on your own layouts (Custom Queries).

It also reads handwriting, accepts PDFs up to 3,000 pages in asynchronous mode, and new AWS customers get a free tier for three months (100 pages a month on AnalyzeExpense). If your team already runs Lambda, S3 and Step Functions, Textract slots into a pipeline you know.

Why teams look for a Textract alternative

There is no purchase order model

A PO is not an invoice, but AnalyzeExpense is the closest model Textract has, and its taxonomy is an invoice one: the document number comes back as INVOICE_RECEIPT_ID, and there is no standard field for the unit of measure (UoM), the requested delivery date of a line or the buyer's item number, which stay in a raw EXPENSE_ROW string your code has to split. With AnalyzeDocument you get the table cell by cell instead, and you write the mapping from column headers to your fields for every supplier layout. Either way, Textract tells you what is on the page; deciding which number is the unit price is your job.

Here is the kind of line you get back from AnalyzeExpense for a printed PO row:

{
  "LineItemExpenseFields": [
    { "Type": { "Text": "PRODUCT_CODE" }, "ValueDetection": { "Text": "GL-NIT-M" } },
    { "Type": { "Text": "ITEM" },         "ValueDetection": { "Text": "Nitrile gloves, box of 100" } },
    { "Type": { "Text": "QUANTITY" },     "ValueDetection": { "Text": "40" } },
    { "Type": { "Text": "UNIT_PRICE" },   "ValueDetection": { "Text": "7.85" } },
    { "Type": { "Text": "PRICE" },        "ValueDetection": { "Text": "314.00" } },
    { "Type": { "Text": "EXPENSE_ROW" }, "ValueDetection": { "Text": "GL-NIT-M Nitrile gloves, box of 100 40 BX 7.85 314.00" } }
  ]
}

Illustration built from the field list in the AWS documentation; the values are a fictitious PO line. The unit (BX) is only in the raw row, and nothing tells you whether 40 × 7.85 really makes 314.00: that check is yours.

The price depends on the features you combine

Textract bills each page once per feature combination. For a purchase order, the realistic choices are these (US West, first million pages a month):

Amazon Textract price per page by feature combination, from aws.amazon.com/textract/pricing (read on 13 November 2026). Illustration for a PO, not a quote.
What you callWhat you get for a POPrice per page
AnalyzeExpenseInvoice style header and lines, UoM in a raw row$0.010
AnalyzeDocument, Tables + QueriesEvery table cell, plus answers to up to 15 questions per page (30 in async)$0.020
AnalyzeDocument, Custom QueriesAnswers from an adapter trained on your layouts$0.025
AnalyzeDocument, Tables + Forms + QueriesAll key value pairs, tables and answers$0.070

At 500 two page POs a month, that is $10 with AnalyzeExpense and $70 with Tables, Forms and Queries. The OCR line is small. The real cost is the code and the hosting around it: splitting rows, recomputing quantity times unit price, adding the lines up and comparing them with the subtotal, matching SKUs with your catalog, storing low confidence fields for a person to check, and keeping all of that working when a supplier changes its template.

A few limits that hit purchase orders

  • Handwriting is supported in English only, and key value extraction (Forms) in six languages: English, Spanish, Italian, Portuguese, French and German. Other APIs, Queries included, are English only.
  • Synchronous calls take one page per PDF. A PO whose lines run onto a second page goes through the asynchronous API, and stitching the table back together is up to you.
  • Text under 15 pixels high is not detected, about 8 point type at 150 DPI: small print on a faxed PO is a common miss.
  • There is no review screen. Low confidence fields have to be routed to a person by a tool you add.

We did not quote user reviews: the review sites we tried blocked our reader. Every limit above is from the AWS documentation.

Textract alternatives at a glance

Facts from each vendor's site, read on 13 November 2026 unless stated otherwise (Koncile and Nanonets on 12 November, Mindee on 4 November, Veryfi and Parseur on 29 October). Prices exclude tax.
ToolBest forPO modelChecks built inFree startEntry price
Amazon TextractAWS teams building their ownNoNo100 pages a month, 3 months$0.01/page (AnalyzeExpense)
KoncilePO lines, totals and matching in one callYesLine totals, catalog, PO / delivery note / invoice20 credits once€199/month for 1,000 pages
Azure Document IntelligenceTeams on AzureInvoice model that lists POsNo500 pages a month$10 per 1,000 pages (prebuilt)
Google Document AITeams on Google CloudNo, custom extractorNoGoogle Cloud creditsPer page, see pricing page
MindeeDevelopers, customer POsYesYour code14 day trial$53/month billed annually
VeryfiPOs plus receipts in one APIYesTwo and three way announced100 docs a month$500/month minimum
NanonetsWorkflows built in one platformYesThree way matching announced$50 of credits$100/month, billed per block
ParseurA PO mailbox, small volumeYesNot built in20 pages every monthPriced by volume

1. Koncile: the checks Textract leaves to your code

Koncile is the company behind this site. Its purchase order template returns the header (PO number, dates, buyer, vendor, bill to, ship to, terms, subtotal, tax, total) and one row per line with SKU, description, quantity, UoM, unit price and line total, already split. The work you would write around Textract is in the response: quantity times unit price recomputed on each line, lines added up and compared with the subtotal, and each line matched to your item catalog or price list. Delivery notes and invoices use the same template library, so a PO can be compared line by line with what was delivered and billed. Handwritten order forms and faxes are read too.

For developers, the purchase order OCR API page shows a request and the JSON a PO returns. Pricing is public: 20 free credits at signup, then Advanced at €199 per month (about $229 on the US price list) for 1,000 credits, one credit per extracted page (pricing, read on 12 November 2026). Some operations, matching among them, have their own price on the full price list.

Where Koncile is weaker. Per page, it costs far more than Textract's $0.01: it only makes sense if it replaces code you would otherwise build and maintain. It does not run inside your AWS account. It is not a procurement or AP suite: it does not create or approve POs and does not pay invoices, and there is no native app inside NetSuite, Dynamics or Odoo (data goes out through the API, webhooks, Make, n8n, Zapier or files). The 20 credits are given once. Cramped handwriting and low resolution scans remain its first sources of error.

2. Azure Document Intelligence: the closest hyperscaler swap

Microsoft's prebuilt invoice model reads "sales invoices, utility bills, and purchase orders" according to its documentation, with header fields and line items, from phone photos, scans and digital PDFs. All prebuilt models cost $10 per 1,000 pages, custom extraction $30 per 1,000, and the free tier covers 500 pages a month. For a Textract team, the switch is mostly an SDK change. The weakness is the same as Textract's: there is no totals check, no catalog matching and no review screen, so the code around the OCR stays yours.

3. Google Document AI: custom extraction on Google Cloud

Google's processor list has an invoice parser and an expense parser but no purchase order processor. For POs you train a custom extractor on your own samples, billed per page on the Document AI pricing page, with savings plans for committed volume. It suits teams whose data already lives in Google Cloud. Training, evaluating and maintaining the extractor is work Textract users know well, and matching remains your code.

4. Mindee: a typed PO API with SDKs

Mindee has a dedicated Purchase Order OCR API aimed at sales order entry, with line items, SDKs in six languages and published accuracy figures on its own test set (product page read on 29 October 2026). Starter costs $53 per month billed annually for 500 credits a month, or $59 billed monthly (read on 4 November 2026), with a 14 day trial. Coming from Textract, you trade a cheaper raw page for a typed PO schema; matching against a price list is still code you write, as our page on Mindee alternatives details.

5. Veryfi: POs and receipts in one API

Veryfi has a pretrained PO model with SKU, quantity, UoM, unit price and total on each line, and announces two and three way matching on its PO page (read on 29 October 2026). The free plan covers 100 documents a month; the first paid plan is a $500 monthly minimum commitment, and the PO price is not public. Ask for it before you compare with Textract's per page price.

6. Nanonets: when you also want the workflow

Nanonets reads POs and lets you chain extraction, validation, approval and export into workflows. It bills each block a workflow runs, from $0.02 to $0.30, with $50 of free credits; ERP connectors sit in plans sold on quote (pricing read on 12 November 2026). It replaces the Lambda and Step Functions glue with a hosted workflow, at a cost per PO that depends on the blocks you chain. Our page on Nanonets alternatives works out one example.

7. Parseur: a parser for a PO mailbox

Parseur starts from the inbox: you forward POs to a dedicated address, it extracts header fields and table rows and exports to Excel, CSV, JSON, Google Sheets, Zapier, Make or Power Automate (read on 29 October 2026). Its free plan includes 20 pages every month. It is the shortest path from email to spreadsheet for a small volume, with no matching and no API first design.

How to choose, in one minute

  • Your bottleneck is the code, not the OCR bill: choose a tool whose response already holds what you were writing (split lines, totals check, matching), and compare its price with a few days of engineering a month.
  • Data must stay on a hyperscaler: Azure Document Intelligence or Google Document AI, and plan the checks yourself.
  • Your POs are handwritten, faxed or not in English: test that first, since Textract reads handwriting in English only.
  • Your volume is in the millions of pages: Textract and the hyperscalers stay cheapest per page; a specialized tool earns its price on the checks, not on the OCR.

How to test an alternative on your own POs

Vendor accuracy figures are measured on vendor test sets. A fair comparison takes an afternoon:

  1. Pick 20 real POs: clean PDFs, scans, a fax, a handwritten order, a PO whose lines run onto a second page, and two from your most difficult supplier or customer.
  2. Run them through your current Textract pipeline and through one or two alternatives, without retraining anything.
  3. Count, per PO, the lines whose SKU, quantity, unit price and line total are all right, not only the header fields.
  4. Check that totals which do not add up are flagged, and note how many fields a person still has to look at.
  5. Write down the code each option still needs before the data can enter your ERP, and put a price on it.
A laptop showing source code, the kind of glue code a Textract pipeline needs around the OCR
Count the code you still maintain, not only the price per page. Photo: Arnold Francisca on Unsplash.

The tool that wins is the one with the fewest corrections and the least code left to maintain, at your volume.

Frequently asked questions

01What is the best AWS Textract alternative for purchase orders?

It depends on what you were building around Textract. If the effort goes into splitting lines, checking totals and matching with the catalog or invoice, Koncile returns all of that in one response. If you must stay on a hyperscaler, Azure Document Intelligence has an invoice model that also reads purchase orders. For a typed PO API, Mindee.

02Does Amazon Textract have a purchase order model?

No. Its specialized APIs cover invoices and receipts (AnalyzeExpense), US identity documents (AnalyzeID) and mortgage documents (AnalyzeLending). For a PO you use AnalyzeExpense or AnalyzeDocument with tables and queries, then map the fields yourself.

03How much does Textract cost per purchase order?

From $0.01 per page with AnalyzeExpense to $0.07 per page with tables, forms and queries combined, for the first million pages a month in US West (pricing read on 13 November 2026). A two page PO therefore costs between $0.02 and $0.14, before the cost of the code and hosting around it.

04Is there a free AWS Textract alternative?

Open source OCR such as Tesseract is free but returns text, not PO fields. Among hosted tools, Parseur gives 20 pages every month, Azure Document Intelligence 500 pages a month, Veryfi 100 documents a month and Koncile 20 credits once at signup.

05Can Textract read handwritten purchase orders?

Yes, but AWS documents handwriting recognition for English only. A handwritten order in French, German or Spanish needs another tool, and cramped handwriting stays the main source of errors for every tool.

06Can Textract match a PO with an invoice?

No. Textract extracts; comparing a PO with a delivery note and an invoice (three way matching) is logic you write, or a feature of tools such as Koncile, Veryfi or Nanonets.

07Is Azure Document Intelligence better than Textract for POs?

For POs it has one advantage: its prebuilt invoice model is documented as reading purchase orders too, with line items, at $10 per 1,000 pages. It has the same gap as Textract on checks and matching.

08Should I replace Textract or build on top of it?

Build on it if you process millions of pages, have engineers who own the pipeline and few supplier layouts. Replace it when layouts change often, POs are handwritten or multilingual, or the team spends more time maintaining the parsing code than using the data.

Sources

Test a Textract alternative on your own POs

Upload a purchase order and get the header and every line back as fields, with the totals checked. 20 free credits, no card, no call.

Create a Koncile account, 20 free creditsTalk to an expert

20 free credits, no credit card. For professional use only.