# What AI document processing for invoices and receipts costs per page and where documents go

Amazon Textract and Azure charge $0.01 a page for invoices and receipts; Google charges $0.10 for up to 10 pages, so a one-page receipt costs ten times as much. Read on each vendor's own pages, 2026-11-06.

If you want software to read invoices and receipts and pass the fields on to your books, three cloud services do this as a pay-per-page service: Amazon Textract, Microsoft Azure Document Intelligence and Google Document AI. This page compares what each vendor publishes about price, about what it does with the documents you send, and what the one independent study located says about accuracy. Prices and data terms are from the vendors' own pages, read on 2026-11-06, in US dollars. Nothing here was tested on a real set of invoices. What QuickBooks Online, Xero and Excel read from a PDF invoice themselves is on [PDF invoices into QuickBooks, Xero and Excel](/research/pdf-invoices-into-accounting-software).

## What they charge

- **Amazon Textract, Analyze Expense:** $0.01 a page in the US West (Oregon) region for the first million pages, $0.008 a page after that. Amazon's own worked example puts 100,000 invoices at $1,000.[1] New AWS customers get 100 Analyze Expense pages a month free, for three months.[1]
- **Azure Document Intelligence, prebuilt receipt and invoice models:** $10 per 1,000 pages, which is $0.01 a page, and a free tier of 0 to 500 pages a month.[2] Microsoft says "Prices are estimates only and are not intended as actual price quotes."[2]
- **Google Document AI, Invoice parser and Expense parser:** $0.10 per count, and Google's footnote reads "1 count equals up to 10 pages in a document."[3] Its examples: a document of 1 to 10 pages costs $0.10, and 11 to 20 pages costs $0.20.[3] Google lists a lower rate of $0.09 or $0.08 a count under one- and three-year savings plans.[3]

Arithmetic from those prices (this page's own, not a vendor figure): 1,000 one-page receipts cost about $10 on Amazon or Azure and $100 on Google, because each one-page receipt is billed as a full count. If a firm scanned several receipts into one multi-page PDF, Google's price per page would fall to about the others'. The same arithmetic with an assumed 500 documents a month gives $5 or $50. Google's page also says that if a document has "multiple eligible forms, only the first is processed and charged."[3] Whether that matters for a PDF holding many receipts is not established here.

## Where your documents go

- **Amazon:** the page on AI services opt-out policies says "AWS AI services may use and store customer content for service improvement, such as fixing operational issues, evaluating service performance, debugging, or model training", and that this content may be stored in a Region other than the one you use.[4] You can opt out with an AWS Organizations policy, and Textract is one of the services that policy covers.[4][5] Opting out deletes earlier stored copies that are not needed to provide the service.[4] Textract's own FAQ answer on this is not cited: its answers are hidden behind a click on the page.
- **Microsoft:** documents are processed in the region where you created the resource, and "temporarily encrypted and stored in Azure Storage."[6] The page says the service keeps the submitted input and the results for 24 hours after an analysis finishes and deletes both, and that you can delete them sooner with the Delete Analyze Result call.[6] That page does not say whether Microsoft uses the documents to train models, either way.
- **Google:** "At Google Cloud, we never use customer data to train our Document AI models."[7] For online requests the document is "processed in memory, encrypted in flight, and not persisted to disk", though Google logs some request metadata to "improve our service and combat abuse."[7] For batch requests the document is stored for a short period, "typically deleted immediately after the processing, with a failsafe Time to live (TTL) of one day."[7]

The same question for a chatbot, what OpenAI keeps and whether it trains on it, is on [OpenAI data use and chatbot liability](/research/ai-privacy-and-chatbot-liability).

## Canada

Amazon lists a Canada (Central) endpoint for Textract, `textract.ca-central-1.amazonaws.com`.[8] Google lists Montréal (`northamerica-northeast1`) among the single regions where Document AI has "limited support".[9] Which of Google's processors run there, and which Azure regions offer the invoice model, is not established here. Amazon's quota table shows 1 request a second for synchronous Analyze Expense in regions other than the named ones; Canada (Central) is not named in the table, so it falls under "Other Regions" only on a reasonable reading.[8]

## What one study found about accuracy

Researchers at RMIT University in Melbourne tested Amazon Textract's receipt feature on 118 receipts from 59 Australian vendors: 77 photographed with three phones, 41 screenshots of electronic receipts, with no control over lighting or angle.[10] It is a preprint posted in December 2025, and the authors call their results qualitative. They report:

- "the receipts' totals were consistently detected."[10]
- Vendor names matched exactly on 68.7 percent of receipts where the name was printed as text (46 of 67) and 66.7 percent where it was a logo (26 of 39); 12 receipts with no readable name were left out of that count.[10]
- Dates were found but came in mixed formats, such as 16 Apr 2025, 07/04/25 and 2025-09-12.[10]
- Receipts laid out like invoices came back without line items: "Textract still extracted summary fields (vendor, total, date) but left item arrays empty".[10]
- Receipts with non-English product names were partly read; blurred, angled or shadowed photos "often resulted in missing or partially extracted fields", though the authors saw only a few examples.[10]

The study covers one service, one country's receipts and 118 photographs. It did not compare Amazon with Microsoft or Google. Several vendor blogs publish head-to-head benchmarks of the three services. They are not used here.

How chat apps and an academic test did at pulling numbers out of documents, set against human typing error rates, is on [Can AI do your data entry? Tests and costs](/research/ai-data-entry).

## Not yet proven

- **Which service is more accurate.** No independent comparison of all three on invoices is cited.
- **Products built for bookkeepers.** Receipt-capture apps sold to bookkeeping firms are not covered; this page is about the three cloud services they may be built on.
- **PDF invoices.** The RMIT study used images only, "no PDF-based receipts were included."[10]
- **Taxes and compliance.** Nothing here covers whether extracted figures are acceptable to a tax authority.
- **Current prices elsewhere.** Prices differ by region and currency; the figures here are the US ones.

If you run a bookkeeping practice and have measured an extraction service on your own invoices, I would like to compare notes: mackenzie@mackenziebowes.com.

## Sources

1. Amazon Web Services, Amazon Textract pricing, read 2026-11-06, undated. <https://aws.amazon.com/textract/pricing/>
2. Microsoft, Azure Document Intelligence pricing (US region, USD), read 2026-11-06, undated. <https://azure.microsoft.com/en-us/pricing/details/document-intelligence/>
3. Google Cloud, Document AI pricing (USD), read 2026-11-06, undated. <https://cloud.google.com/products/document-ai/pricing>
4. Amazon Web Services, AI services opt-out policies, AWS Organizations user guide, read 2026-11-06, undated. <https://docs.aws.amazon.com/organizations/latest/userguide/orgs_manage_policies_ai-opt-out.html>
5. Amazon Web Services, AI services opt-out policy syntax and examples (lists textract among the services), read 2026-11-06, undated. <https://docs.aws.amazon.com/organizations/latest/userguide/orgs_manage_policies_ai-opt-out_syntax.html>
6. Microsoft, Data, privacy, and security for Document Intelligence, read 2026-11-06, undated. <https://learn.microsoft.com/en-us/azure/foundry/responsible-ai/document-intelligence/data-privacy-security>
7. Google Cloud, Document AI security and compliance, read 2026-11-06, undated. <https://docs.cloud.google.com/document-ai/docs/security>
8. Amazon Web Services, Amazon Textract endpoints and quotas, AWS General Reference, read 2026-11-06, undated. <https://docs.aws.amazon.com/general/latest/gr/textract.html>
9. Google Cloud, Document AI regional and multi-regional support, last updated 2026-10-26. <https://docs.cloud.google.com/document-ai/docs/regions>
10. Oommen, Sanchez, Britto, Wang, Chiou and Spichkova (RMIT University), Towards Analysing Invoices and Receipts with Amazon Textract, arXiv 2512.19958, submitted 2025-12-23 (HTML version read). <https://arxiv.org/abs/2512.19958>

---

Published 2026-11-06. Web page: https://www.mackenziebowes.com/research/ai-document-processing-invoices-receipts
