You are deciding whether to run your invoice and contract processing on a cloud API you integrate yourself, or on a packaged platform that arrives with a workflow. The per-page prices are public. The decision does not turn on them.
This page lists the 2026 published rates for Azure AI Document Intelligence, Google Document AI, AWS Textract and Rossum, works through what a real Hong Kong volume costs on each, and names what the per-page price does not include.
Prices below are the vendors' published list rates in US dollars. Verify against the vendor pricing page before you commit, because these change.
What does AI document processing actually cost per page in 2026?
Published list rates in 2026 range from US$0.60 to US$50 per 1,000 pages depending on which API call you make, not which vendor you pick. Raw text extraction sits near US$1.50 per 1,000 pages across all three hyperscalers. Structured field extraction costs ten to thirty times more.
That spread is the single most important fact on this page.
The same PDF invoice can cost US$1.50 or US$50 per thousand pages depending on whether you call plain text detection or full forms-and-tables analysis. Teams routinely default to the most expensive call because it returns the richest output.
Baseline per-1,000-page list rates, 2026
--- Plain text extraction (OCR only): about US$1.50 per 1,000 pages on Azure, Google and AWS alike.
--- Layout and structure parsing: about US$10 per 1,000 pages on all three.
--- Prebuilt document models, for example invoices and receipts: US$8 to US$10 per 1,000 pages.
--- Custom trained extraction: US$20 to US$30 per 1,000 pages.
--- AWS Textract forms analysis: US$50 per 1,000 pages, the highest single-call rate in this comparison.
How much does Azure AI Document Intelligence cost?
Azure AI Document Intelligence publishes 2026 rates of US$1.50 per 1,000 pages for Read (OCR), US$10 per 1,000 for Layout, US$10 per 1,000 for prebuilt models, and US$30 per 1,000 for custom extraction. Add-on features such as high-resolution, formula and barcode processing add US$6 per 1,000 pages.
The free tier is real but narrow.
Azure's free tier allows 500 pages per month. It processes only the first two pages of any document submitted, and caps file size at 4 MB. That is enough to test extraction quality on single-page invoices, and not enough to test a twelve-page contract.
The practical implication: budget for a paid tier during evaluation, not after it. Teams that pilot on the free tier often discover their real documents were never fully processed.
How much does Google Document AI cost?
Google Document AI charges US$1.50 per 1,000 pages for Enterprise Document OCR on the first 5 million pages each month, dropping to US$0.60 per 1,000 pages above that. Layout Parser is US$10 per 1,000 pages. Custom extractors and Form Parser cost US$30 per 1,000 pages, discounted to US$20 at higher volume.
Google is the only vendor in this comparison with a published volume discount on OCR.
At 5 million pages a month the rate falls 60%, from US$1.50 to US$0.60 per 1,000. For a high-volume scanning operation, that is a genuine structural advantage.
Below roughly 500,000 pages a month, the discount is irrelevant and Google prices the same as Azure and AWS. Do not let a volume tier you will never reach drive a platform decision.
How much does AWS Textract cost?
AWS Textract prices per API call: US$1.50 per 1,000 pages for DetectDocumentText, about US$15 per 1,000 for table extraction, US$50 per 1,000 for forms analysis, and roughly US$8 to US$10 per 1,000 for AnalyzeExpense, the purpose-built invoice and receipt API.
Textract has the widest internal price spread of any option here.
Calling DetectDocumentText costs US$1.50 per 1,000 pages. Calling AnalyzeDocument with forms, tables and queries together can reach US$70 per 1,000 pages on the same document. That is a 47-times difference driven purely by which call your developer wrote.
For invoices and receipts specifically, AnalyzeExpense at US$8 to US$10 per 1,000 pages is both cheaper and generally more accurate than assembling the same result from AnalyzeDocument forms analysis. Many teams do not know this API exists.
How much does Rossum cost?
Rossum publishes a Starter plan floor of US$18,000 per year. It includes API access, document ingestion by email, API or manual upload, a 12-month document archive with search, the Rossum Aurora document AI, and unlimited seats. Business and Enterprise tiers are quote-only.
Rossum is not priced per page, and comparing it to the hyperscalers on unit cost misses the point.
The Starter tier buys a working accounts-payable product: a validation interface, a review queue, an audit trail and unlimited user seats. On Azure, Google or AWS you receive an API response and build all of that yourself.
What sits above Starter matters for enterprise buyers. Business adds custom business logic, master data matching, duplicate detection and SAP and Coupa integrations. Enterprise adds single sign-on, a sandbox environment, preferred cloud location and custom branding. Both require a sales conversation, and neither price is published.
What does 8,000 invoices a month actually cost?
For a Hong Kong logistics firm processing 8,000 invoices a month at an average of two pages each, or 16,000 pages, published API rates give annual figures of roughly US$1,920 on AWS AnalyzeExpense, US$1,920 on Azure prebuilt invoice, and US$5,760 on a Google custom extractor. Rossum starts at US$18,000.
Run the arithmetic before the vendor demo, not after.
Worked example: 16,000 pages per month, invoices only
--- AWS Textract AnalyzeExpense at US$10 per 1,000 pages: US$160 per month, about US$1,920 per year.
--- Azure prebuilt invoice model at US$10 per 1,000 pages: US$160 per month, about US$1,920 per year.
--- Google custom extractor at US$30 per 1,000 pages: US$480 per month, about US$5,760 per year.
--- Rossum Starter: US$18,000 per year, flat, regardless of whether you send 16,000 or 60,000 pages.
The gap between US$1,920 and US$18,000 looks decisive. It is not, and the next section explains why.
What is not included in the per-page price?
The API fee is the smallest line in a document automation budget. Integration engineering, an exception-handling interface, validation staffing, master data matching and ongoing accuracy monitoring typically dominate total cost. A US$1,920 annual API bill routinely sits inside a project costing many times that in the first year.
Five costs that do not appear on any pricing page:
--- Integration engineering. Connecting extraction output to your ERP, matching to purchase orders and handling failures is bespoke work in every deployment.
--- The exception interface. Roughly one document in ten will need human review. Someone has to build the screen where that happens, or buy a product that ships with one.
--- Validation staffing. Extraction at 95% field accuracy still means a person checks the flagged 5%. That headcount is the real operating cost.
--- Master data matching. Matching an extracted vendor name to the right vendor record is a separate problem from reading the text, and it is where most projects stall.
--- Accuracy drift monitoring. Document layouts change. Extraction quality degrades quietly unless someone measures it.
If you are building a broader picture of AI spend across the organisation, our framework for managing enterprise AI cost covers how to structure that budget.
Where does each option genuinely lose?
Every option in this comparison has a condition under which it is the wrong choice. Azure's free tier truncates documents at two pages. Google's volume discount is unreachable for mid-market volumes. Textract's pricing punishes the wrong API call severely. Rossum's US$18,000 floor is poor value below roughly 3,000 documents a month.
Stated plainly, because a comparison where one option wins every row is an advertisement.
Azure loses when you need to evaluate multi-page documents cheaply. The two-page free-tier cap makes honest pilot testing of contracts and statements impossible without paying.
Google loses for mid-market buyers. Its structural advantage activates above 5 million pages a month. A 500-employee Hong Kong firm will not reach it, and pays the same as everyone else.
AWS loses on cost predictability. The 47-times spread between the cheapest and most expensive call on the same page means a single developer decision can multiply your bill. Governance overhead is real.
Rossum loses on entry price. Below roughly 3,000 documents a month, US$18,000 per year is expensive for what a competent integration on a hyperscaler API would deliver. It also does not publish Business or Enterprise pricing, so budgeting above Starter requires a sales cycle.
And UD has limits worth naming. UD is a Hong Kong technology implementation partner, not a reseller of these platforms. We do not set their prices and cannot obtain vendor discounts on your behalf. If your requirement is a pure accounts-payable product with no integration to existing systems, buying Rossum or a comparable AP platform directly is a reasonable path and you do not need an implementation partner for it.
Which option should you choose?
Choose by document volume and internal engineering capacity. Below 3,000 documents a month with in-house developers, use Azure prebuilt models or AWS AnalyzeExpense. Above 3,000 documents a month without engineering capacity, a packaged platform like Rossum earns its floor price. Above 5 million pages a month, Google's volume tier wins.
Four buyer profiles and the honest verdict for each:
--- Under 3,000 documents a month, with developers. Azure prebuilt or AWS AnalyzeExpense. Annual API spend under US$2,000. You build the review interface.
--- Over 3,000 documents a month, no spare engineering. A packaged platform. The workflow, validation screen and audit trail are the product, and rebuilding them costs more than US$18,000 of engineering time.
--- Already committed to one cloud. Use that vendor's service. Cross-cloud data movement and a second security review will cost more than the price difference.
--- Above 5 million pages a month. Google Document AI. The drop to US$0.60 per 1,000 pages is worth six figures annually at that scale.
What is the right next step?
Before requesting any vendor quote, establish three numbers: documents per month, average pages per document, and the fields you actually need extracted. Without those, every quote you receive will be built on the vendor's assumptions rather than your volumes, and the comparison will not be meaningful.
Then run a paid pilot on 500 of your own documents, not the vendor's samples.
Measure field-level accuracy on your real layouts, and count how many documents need human correction. That percentage, multiplied by your document volume and a fully loaded hourly rate, is your true operating cost. The API fee is rounding.
Document automation is one of the few AI investments where the arithmetic is genuinely knowable in advance. Most organisations skip it and buy on demo quality instead. Technology cycles come and go, and the firms that come through them with working systems are the ones that had someone experienced sitting beside them, asking for the volume numbers before the contract. We understand AI. We understand you. With UD by your side, AI never feels cold.
Reviewed by the UD enterprise AI team, Hong Kong. Prices are vendor list rates current as at August 2026 and should be verified before purchase.
Work Out Your Real Number First
Before you compare quotes, find out where document work is actually costing your team. UD's AI Staff Solution puts trained AI employees on admin, accounting and back-office document workflows, and we'll walk you through every step, from volume assessment and platform selection to integration and accuracy tracking. 28 years serving Hong Kong enterprises.
