The complete guide to intelligent document processing (IDP)

This guide is for finance operations, shared services, claims, onboarding and procurement teams who process documents by hand and have been told that “AI can read them now”. By the end you should be able to tell an IDP platform from an OCR API, estimate what a document will cost you per page at list price, and design a pilot that produces numbers you can trust.

Our position: reading a page is now cheap. What costs money is everything around the read: deciding which documents a person must check, giving them a fast screen to do it, and proving the error rate to an auditor. Choose a tool for that loop, and test accuracy on your own documents, because vendor numbers are measured in different ways.

This guide is not for teams whose partners can already send structured data. If a supplier can send an electronic invoice or an EDI message, read that instead of a PDF. It is also not legal or compliance advice on KYC or data protection.

Summary

• IDP turns documents into checked, structured data. OCR is only its first step.
• Hyperscaler APIs list at $0.0015 to $0.10 per page or document, as of October 2026; IDP platforms start in the tens of thousands of dollars a year.
• Confidence scores are not accuracy. Measure field accuracy and straight-through rate on your own documents.
• Vendor accuracy claims (92.5 percent, 99.5 percent, “90 percent at the start”) use different definitions and cannot be compared.
• LLM extraction is flexible and cheap per page, but you build the review loop yourself.
• Check where documents are processed, how long they are kept and whether they train vendor models.
• The hard part of IDP is exceptions and integration; reading the text is the easy part.

1. What IDP is, and how it differs from OCR and RPA

Intelligent document processing (IDP) is software that takes in documents such as PDFs, scans, photos and emails, works out what each one is, pulls out the fields you need, checks them against rules and other systems, and sends anything uncertain to a person before the data reaches your ERP or case system. The output is structured data with a record of who or what confirmed each value.

OCR (optical character recognition) converts an image of text into machine-readable text. It does not tell you which number is the invoice total or whether this page belongs to the same claim as the last one.

RPA (robotic process automation) uses software robots to automate repetitive, rule-based tasks such as data entry, according to UiPath, one of the larger vendors in both RPA and IDP.¹1 Source: UiPath, “What is Robotic Process Automation”, accessed Oct 2026. The definition is the vendor’s own. A bot can type extracted values into a legacy screen. It cannot read a document it has never seen.

LLM extraction means sending a document to a large language model (LLM) with instructions such as “return the supplier name, invoice number and total as JSON”. On its own it is one step of a pipeline.

Approach	What it produces	Handles new layouts	Human review built in	Typical role
OCR	Raw text and positions	Yes, as text only	No	First step of IDP
IDP platform	Validated fields, audit trail	Yes, varies by model	Yes	End-to-end document intake
RPA	Keystrokes in other systems	No	No	Moves data into old systems
LLM prompt	Fields as JSON	Yes	No, you build it	Flexible extraction step

The category is now mature enough to have its own analyst coverage. Gartner published its first Magic Quadrant for Intelligent Document Processing Solutions on September 3, 2025, according to ABBYY, which was named in it.²2 Source: ABBYY, “ABBYY Named a Leader in 2025 Gartner Magic Quadrant for Intelligent Document Processing Solutions”, accessed Oct 2026. Press release dated September 8, 2025; the full Gartner report is paywalled.

2. How modern IDP works

Every IDP system runs the same five stages. Products differ in how many of them they handle for you.

Ingestion and digitization

Documents arrive by email, upload, scanner or API. The system runs OCR where a PDF has no text layer and records layout: lines, tables, checkboxes, signatures. Microsoft, for example, says the minimum text height its service reads is about 8-point text at 150 dots per inch, and recommends text-based PDFs over scans where possible.³3 Source: Microsoft Learn, “Custom document models - Document Intelligence”, accessed Oct 2026. PDFs and TIFFs of up to 2,000 pages and 500 MB are supported on the paid tier.

Classification and splitting

Classification decides the document type; splitting decides where one document ends and the next begins in a scanned batch. Both Azure AI Document Intelligence and Google Document AI sell classification and splitting as separate, priced steps.⁴4 Source: Microsoft, “Azure AI Document Intelligence pricing” and Microsoft, Azure Retail Prices API, accessed Oct 2026. Dollar figures from the Retail Prices API for East US, pay-as-you-go (S0) and cloud commitment tiers; the pricing page shows them only by region selection. ⁵5 Source: Google Cloud, “Document AI pricing”, accessed Oct 2026. Standard prices shown; savings plans list lower rates.

Extraction

Extraction returns each field’s value, usually with a page location and a confidence score. Three methods exist, and most products mix them:

• Templates. Rules or zones per layout. Fast and precise on fixed forms, and brittle when a supplier changes its invoice design. Google’s documentation suggests about three training documents for a template-based custom extractor on fixed-layout forms.⁶6 Source: Google Cloud, “Custom extractor overview”, accessed Oct 2026.
• Trained models. Machine learning models trained on labeled examples. Microsoft says five labeled examples of a document type are enough to start a custom extraction model, and that one neural model can cover layout variations that would each need their own template.³
• Generative and LLM-based models. A foundation model fine-tuned or prompted to find fields in any layout. Google recommends its generative option first for variable layouts such as contracts, invoices, bank statements and payslips, and says these models typically need fewer training documents.⁶

Validation

Validation checks the extracted values against rules and reference data: does the total equal the sum of the line items, does the purchase-order (PO) number exist in the ERP, is the supplier’s tax ID on file, is the certificate’s expiry date in the future. Rules catch errors that a confident model will never flag.

Human-in-the-loop review

Human-in-the-loop (HITL) review sends uncertain fields or documents to a person, who confirms or corrects them on a screen that shows the value next to the page. Two details matter:

• What triggers review. Microsoft describes field confidence as an estimated probability, so 0.95 means a prediction is likely correct 19 out of 20 times, and suggests using it to decide between auto-accept and human review where accuracy is critical.⁷7 Source: Microsoft Learn, “Interpret and improve model accuracy and confidence scores”, accessed Oct 2026. Good triggers combine confidence thresholds, failed validation rules, missing required fields and a random sample of “confident” documents.
• Who builds the screen. Textract’s managed review option, Amazon Augmented AI (A2I), can trigger review on confidence ranges, missing keys or random samples, but AWS says it is closed to new customers and will get no new features.⁸8 Source: AWS, “Use Amazon Augmented AI with Amazon Textract”, accessed Oct 2026. A2I integrates with the AnalyzeDocument operation. A new AWS build needs its own review tool.

LLM-based extraction

General-purpose models now read documents directly. Anthropic’s API converts each PDF page into an image and also passes the page’s extracted text, so the model sees layout, charts and tables; it estimates 1,500 to 3,000 text tokens per page plus image costs, with up to 600 pages per request on its largest context windows.⁹9 Source: Anthropic, “PDF support” and Anthropic, “Pricing”, accessed Oct 2026. The 600-page limit drops to 100 pages when the request’s context window is under 1M tokens. Google’s Gemini API counts each document page as 258 tokens and accepts PDFs up to 1,000 pages or 50MB.¹⁰10 Source: Google, “Document understanding” (Gemini API) and Google, “Gemini Developer API pricing”, accessed Oct 2026. Worked-example token counts are our assumptions.

The appeal: no training, any layout, and fields you redefine by editing a prompt. The limits: no reliable page location or calibrated confidence score by default, and the occasional plausible value that is not on the page. A document is untrusted input: OWASP lists indirect prompt injection, where instructions hidden in files or other external content change the model’s behavior, as the top risk for LLM applications.¹¹11 Source: OWASP Gen AI Security Project, “LLM01: Prompt Injection”, accessed Oct 2026. A supplier invoice with white-on-white text telling the model to change the bank account is a cheap attack to attempt.

For where document extraction sits among broader automation, see the complete guide to AI agents in enterprise operations.

3. Common use cases

IDP works best where volume is high, layouts vary and the extracted data feeds a system that enforces its own rules.

Invoices and accounts payable (AP). The classic case. Header fields (supplier, invoice number, dates, totals, tax) and line items are extracted, then matched against the PO and goods receipt in the ERP (two-way or three-way matching). Line items are the hard part; header fields are close to solved.

Insurance claims. Claim forms, medical bills, repair estimates, photos and correspondence arrive as a mixed pack. Classification and splitting matter as much as extraction, and handwriting is common.

KYC and customer onboarding. Know your customer (KYC) checks need data from ID documents, proof of address, bank statements and company registry extracts. AWS lists a dedicated Analyze ID API, and Google lists parsers for US driver licenses, US passports, bank statements and pay slips.¹²12 Source: AWS, “Amazon Textract pricing”, accessed Oct 2026. Rates shown for US West (Oregon); new AWS customers get a three-month free tier. ⁵ Checking that an ID is genuine is a separate product category.

Supplier certificates and compliance documents. Insurance certificates, ISO certificates, food safety audits. Few fields (issuer, holder, scope, expiry), endless layouts, and the value is in validation: is it expired, does the holder match the supplier record. A good fit for generative extraction plus a short rule set.

Contracts. Contracts are long, unstructured and full of clauses where meaning matters more than position. Extraction of key terms (parties, dates, renewal, liability caps) is feasible; judgment about whether a clause is acceptable is a contract review problem, not a capture problem. UiPath, for example, sells generative extraction for unstructured content like contracts separately from its forms and invoice models.¹³13 Source: UiPath, “Document Understanding”, accessed Oct 2026. Product claims are the vendor’s own.

Where trading partners can send structured orders and invoices, skip the document entirely; see the complete guide to EDI for ecommerce brands. For the full invoice-to-payment process around extraction, see the complete guide to accounts payable automation.

4. Accuracy, and how to measure it

Every vendor quotes accuracy. Almost none of those numbers can be compared with each other, because they measure different things.

The metrics that matter

• Field-level accuracy. The share of fields that exactly match a human-verified value, counting missing required fields as errors. Report it per field, because 99 percent on invoice number and 80 percent on line-item quantity average out to a number that describes neither.
• Document-level accuracy. The share of documents where every required field is correct. Always lower, and closer to what your AP clerk experiences.
• Straight-through processing (STP) rate. The share of documents that reach the target system with no human touch. It drives cost.
• Error escape rate. The share of straight-through documents later found to contain an error. It drives risk.
• Review time per document. How long a person spends on documents that do need review.

STP and escape rate pull against each other: lower the threshold and more errors go straight through too. The right setting depends on the cost of an error in that field: a wrong invoice date is an annoyance, a wrong bank account is a loss.

What the vendors claim

Read these as marketing figures:

• Rossum says its Aurora model averages 92.5 percent accuracy across customers.¹⁴14 Source: Rossum, “Aurora”, accessed Oct 2026. Accuracy and language figures are company claims; Rossum also claims zero risk of hallucination, which we could not test.
• Hyperscience says its customers “regularly achieve” 99.5 percent accuracy and 98 percent automation, and claims 98 percent accuracy on handwriting.¹⁵15 Source: Hyperscience, “Platform”, accessed Oct 2026. Accuracy and automation figures are company claims.
• ABBYY says its pre-trained Vantage skills work with 90 percent accuracy at the start.¹⁶16 Source: ABBYY, “Vantage”, accessed Oct 2026. Accuracy and skills figures are company claims.

None of these pages defines the denominator in a way you can reproduce: field or document, which fields, which document types, before or after human review. Microsoft’s guidance is more useful as a yardstick: for its custom models it suggests targeting an estimated accuracy of 80 percent or higher, and close to 100 percent for sensitive cases such as financial or medical records, with human review added for critical workflows.⁷

How to measure it yourself

Build a ground-truth set before you talk to vendors: a few hundred real documents, chosen to include your worst scans, rarest suppliers and longest tables, with every required field keyed by a person and checked by a second. Hold some back from vendors. Score every candidate field by field on the same set, and record STP and escape rate at the confidence threshold the vendor recommends.

5. The vendor landscape by type

The market splits into three groups. Many companies use more than one.

Hyperscaler document services

AWS Textract, Azure AI Document Intelligence and Google Document AI sell OCR, prebuilt models (invoices, receipts, IDs) and custom models through APIs, billed per page. They are cheap and run in your existing cloud account, but they are developer tools: you build the intake, validation rules, review screen and integration.

Real limits: Textract says its handwriting, invoice and receipt, identity document and Queries processing are English only.¹⁷17 Source: AWS, “Amazon Textract FAQs”, accessed Oct 2026. Its managed review option is closed to new customers.⁸ Azure’s neural and generative models give no accuracy score at training time.⁷ Google also bills custom processors for hosting, at $0.05 per hour, or $438 a year per deployed version.⁵

IDP platforms

ABBYY Vantage, Hyperscience, UiPath Document Understanding and IXP, Rossum and similar products (Tungsten Automation and Infrrd also sit in this group) package the whole loop: intake channels, classification, extraction, validation rules, review screens, model retraining from corrections and connectors to ERPs and RPA. They are priced as annual contracts.

• ABBYY Vantage offers pre-trained skills for more than 150 use cases and runs in ABBYY’s cloud in Europe, the US or Australia, on premises, or in a private Azure instance, the company says.¹⁶ Limit: no public list price that we could find.
• Hyperscience deploys on premises, in a customer’s private tenant on AWS, Google or Azure, or as SaaS, the company says.¹⁵ Limit: a high entry price for small volumes (chapter 6).
• UiPath Document Understanding and IXP handle structured and semi-structured documents, with generative extraction for unstructured ones, feeding UiPath robots and agents.¹³ Limit: best if you already run UiPath; document features are not in its Basic plan.¹⁸18 Source: UiPath Documentation, “IXP licensing and pricing FAQ” and UiPath, “Pricing”, accessed Oct 2026. UiPath’s docs note that consumption is charged per operation at runtime; rates have changed across versions, so confirm with UiPath.
• Rossum is built around transactional documents such as invoices and orders, with its own model, Aurora, that the company says supports 276 languages and learns from each correction.¹⁴ Limit: a one-year minimum contract and a narrower focus.¹⁹19 Source: Rossum, “Pricing”, accessed Oct 2026.

LLM-based approaches

General-purpose models from Anthropic, Google, OpenAI and others, called directly or through a thin wrapper. The cloud providers are also folding LLMs into their services: Azure lists a custom generative extraction meter, and Google offers a fine-tuned foundation model as the first choice for variable layouts.⁴⁶

Type	Examples	Pricing basis	You build	Best for
Hyperscaler API	Textract, Azure, Google	Per page or document	Review, rules, integration	Teams with developers
IDP platform	ABBYY, Hyperscience, UiPath, Rossum	Annual contract, volume	Integration only	Regulated, high-volume ops
LLM direct	Claude, Gemini, GPT	Per token	Everything else	Variable docs, prototypes

6. Pricing models and real list prices

List prices in US dollars, as of October 2026. Cloud prices vary by region; check before budgeting.

Hyperscalers: per page

Service	What it does	List price
Textract Detect Document Text	OCR	$1.50 per 1,000 pages
Textract Analyze Expense	Invoices, receipts	$0.01 per page
Azure Read	OCR	$1.50 per 1,000 pages
Azure prebuilt (invoice, ID, etc.)	Prebuilt models	$10 per 1,000 pages
Azure custom extraction	Trained models	$30 per 1,000 pages
Google Enterprise Document OCR	OCR	$1.50 per 1,000 pages
Google Invoice parser	Prebuilt invoices	$0.10 per 10-page block
Google Custom extractor	Trained or generative	$30 per 1,000 pages

Textract prices are for the first million pages a month in US West (Oregon), with lower rates above that.¹² Azure charges $3 per 1,000 pages for classification, gives 500 free pages a month, and charges $3 an hour for neural model training beyond 10 free hours; monthly commitment tiers lower the unit price.⁴ Google lists its custom classifier and splitter at $5 per 1,000 pages, and its invoice parser charges $0.10 for each document of up to 10 pages.⁵

IDP platforms: annual contracts

Platform pricing is mostly quote-based. What is published:

• Rossum lists its Starter plan from $18,000 per year, with unlimited seats; higher plans are not priced publicly, and the minimum contract is one year.¹⁹
• Hyperscience lists its SaaS product on AWS Marketplace at $50,000 for a 12-month contract, with 24- and 36-month terms also offered and AWS infrastructure costs possibly extra.²⁰20 Source: AWS Marketplace, “Hypercell - SaaS” (sold by Hyperscience), accessed Oct 2026. Listed as “HS Private Cloud Professional”; private offers may differ.
• UiPath meters documents in its own units, up to 1 AI Unit per page for classification plus extraction (generative validation adds one more), or 0.2 Platform Units per page on unified pricing. Its public pricing page does not give a unit’s cash price.¹⁸
• ABBYY does not publish Vantage list prices that we could find.

LLMs: per token

Anthropic lists Claude Haiku 5.5 at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens.⁹ Google lists Gemini 2.5 Flash-Lite at $0.10 input and $0.40 output. Both halve prices for batch processing.⁹¹⁰

A worked example

Take 10,000 two-page invoices a month (20,000 pages). Illustrative, at list price:

• Textract Analyze Expense: 20,000 × $0.01 = about $200 a month.¹²
• Azure prebuilt invoice: 20 × $10 = about $200 a month.⁴
• Google Invoice parser: 10,000 documents × $0.10 = about $1,000 a month.⁵
• Gemini 2.5 Flash-Lite, assuming 258 tokens per page, a 1,000-token instruction and 500 output tokens per invoice: about 15.2 million input tokens and 5 million output tokens, so roughly $3.50 a month at standard rates.¹⁰
• Rossum Starter: from $1,500 a month on an annual contract, before any volume pricing.¹⁹

The model call is rarely the largest cost. If 20 percent of 10,000 invoices need two minutes of review, that is about 67 hours of work a month, before rework. Cutting review time per document can be worth more than slightly better reading.

7. Security, privacy and data residency

Ask four questions of every option: where is it processed, how long is it kept, can it train the vendor’s models, and can it run inside your own environment.

AWS. AWS says its AI services may store and use customer content for service improvement, including model training, possibly in another region. An AWS Organizations opt-out policy stops this and deletes content held for that purpose; the Textract FAQ confirms Textract is covered.²¹21 Source: AWS, “AI services opt-out policies”, accessed Oct 2026. ¹⁷ Set the opt-out before go-live.

Microsoft Azure. Microsoft says Document Intelligence processes data in the region where you created the resource, stores inputs and results temporarily for 24 hours after an analysis completes, then deletes them, and offers an API call to delete them sooner.²²22 Source: Microsoft Learn, “Data, privacy, and security for Document Intelligence”, accessed Oct 2026. For sites that cannot connect to the cloud, Azure offers disconnected containers, which require an access request and a commitment plan.²³23 Source: Microsoft Learn, “Use Document Intelligence containers in disconnected environments”, accessed Oct 2026.

Google Cloud. Document AI asks you to choose a processing and storage location: the us or eu multi-region, or one of a short list of single regions. Capabilities vary by region.²⁴24 Source: Google Cloud, “Document AI regional and multi-regional support”, accessed Oct 2026.

IDP platforms. Hyperscience and ABBYY both offer on-premises and private-cloud deployment, according to their sites.¹⁵¹⁶ Rossum offers a preferred cloud location only from its Enterprise plan up, and says customers can opt in or out of having their data used to improve its model for others.¹⁹¹⁴

LLMs. Check the provider’s retention terms and treat every document as untrusted input. The extraction step should return fields, never call tools or change records. Validate bank details and payee changes against a second source no matter how confident the extraction looks.¹¹

8. Pilot, rollout and common mistakes

A sequence that works for most teams, stated as general guidance rather than a vendor method:

Weeks 1 to 2: baseline. Pick one document type and one process. Measure today’s volume, handling time, error rate and backlog. Build the labeled ground-truth set from chapter 4.

Weeks 3 to 6: bake-off. Run two or three candidates (say, one hyperscaler, one platform, one LLM build) on the held-out set. Score accuracy, STP, escape rate and review time, and price each at your real volume.

Weeks 7 to 12: shadow, then assisted. Run the tool alongside the current process with no postings, then let it draft every document for a person to confirm. Tune thresholds and rules by field.

Month 4 onward: straight-through, with sampling. Let high-confidence documents that pass every rule post without review. Keep reviewing a random sample of them permanently to catch drift. Re-run the ground-truth set whenever the vendor changes its model.

Common mistakes

• Testing on clean samples. Vendors demo on their best documents. Test on your worst.
• Trusting confidence alone. Use business rules and random sampling as review triggers too.⁷
• Ignoring line items. Header fields are easy. If your process needs line-level matching, test it separately.
• Leaving the training opt-out at its default. Check each vendor’s data-use setting before go-live.²¹
• Pricing only the read. Hosting fees, training hours, review labor and integration often cost more than per-page fees.⁵⁴

Moving the extracted data into downstream systems is a workflow problem; see the complete guide to workflow automation tools and, for how IDP fits with the rest of your systems, the complete guide to building a B2B revenue stack.

9. The verdict

IDP is a mature category with cheap reading and expensive judgment. With developers and a few document types, a hyperscaler API plus your own review screen is the cheapest route, and an LLM is worth testing for variable layouts. With regulated, high-volume, mixed document packs and little engineering capacity, a platform’s review loop and audit trail justify the contract. Either way, decide on your own measured accuracy.

Frequently asked questions

What is the difference between OCR and IDP?

OCR converts an image of text into machine-readable characters. IDP uses OCR as a first step, then classifies, extracts, validates and sends uncertain results to a person. The output is checked data you can post.

Can an LLM replace an IDP platform?

For extraction alone, often yes, and at a very low per-page cost.¹⁰ What an LLM does not give you is the review screen, validation rules, audit trail, retraining from corrections and connectors. With engineers to build and run those, an LLM pipeline is viable; without them, a platform is usually the simpler choice.

How many documents do I need to train a custom model?

Fewer than you might expect. Microsoft says five labeled examples are enough to start a custom extraction model, and Google suggests about three for template models, 10 to 100 or more for custom models, and zero to 50 or more for its generative option.³⁶

How much does IDP cost per page?

As of October 2026 at list price, cloud OCR starts at $1.50 per 1,000 pages, prebuilt invoice models cost $10 per 1,000 pages on Azure or $0.01 per page on Textract, and custom extraction costs about $30 per 1,000 pages on Azure and Google.¹²⁴⁵ Platform contracts are annual and quote-based, with published entry points of $18,000 a year for Rossum and $50,000 for Hyperscience.¹⁹²⁰

Is my data used to train the vendor’s models?

It depends on the vendor and your settings. AWS says its AI services may use content for service improvement unless you opt out, Microsoft says Document Intelligence deletes inputs and results after 24 hours, and Rossum lets customers opt in or out of model improvement.²¹²²¹⁴ Check the setting in writing before go-live.

Sources and notes

1. Source: UiPath, “What is Robotic Process Automation”, accessed Oct 2026. The definition is the vendor’s own. ↩

2. Source: ABBYY, “ABBYY Named a Leader in 2025 Gartner Magic Quadrant for Intelligent Document Processing Solutions”, accessed Oct 2026. Press release dated September 8, 2025; the full Gartner report is paywalled. ↩

3. Source: Microsoft Learn, “Custom document models - Document Intelligence”, accessed Oct 2026. PDFs and TIFFs of up to 2,000 pages and 500 MB are supported on the paid tier. ↩ ↩² ↩³

4. Source: Microsoft, “Azure AI Document Intelligence pricing” and Microsoft, Azure Retail Prices API, accessed Oct 2026. Dollar figures from the Retail Prices API for East US, pay-as-you-go (S0) and cloud commitment tiers; the pricing page shows them only by region selection. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷

5. Source: Google Cloud, “Document AI pricing”, accessed Oct 2026. Standard prices shown; savings plans list lower rates. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷

6. Source: Google Cloud, “Custom extractor overview”, accessed Oct 2026. ↩ ↩² ↩³ ↩⁴

7. Source: Microsoft Learn, “Interpret and improve model accuracy and confidence scores”, accessed Oct 2026. ↩ ↩² ↩³ ↩⁴ ↩⁵

8. Source: AWS, “Use Amazon Augmented AI with Amazon Textract”, accessed Oct 2026. A2I integrates with the AnalyzeDocument operation. ↩ ↩²

9. Source: Anthropic, “PDF support” and Anthropic, “Pricing”, accessed Oct 2026. The 600-page limit drops to 100 pages when the request’s context window is under 1M tokens. ↩ ↩² ↩³

10. Source: Google, “Document understanding” (Gemini API) and Google, “Gemini Developer API pricing”, accessed Oct 2026. Worked-example token counts are our assumptions. ↩ ↩² ↩³ ↩⁴ ↩⁵

11. Source: OWASP Gen AI Security Project, “LLM01: Prompt Injection”, accessed Oct 2026. ↩ ↩²

12. Source: AWS, “Amazon Textract pricing”, accessed Oct 2026. Rates shown for US West (Oregon); new AWS customers get a three-month free tier. ↩ ↩² ↩³ ↩⁴ ↩⁵

13. Source: UiPath, “Document Understanding”, accessed Oct 2026. Product claims are the vendor’s own. ↩ ↩²

14. Source: Rossum, “Aurora”, accessed Oct 2026. Accuracy and language figures are company claims; Rossum also claims zero risk of hallucination, which we could not test. ↩ ↩² ↩³ ↩⁴ ↩⁵

15. Source: Hyperscience, “Platform”, accessed Oct 2026. Accuracy and automation figures are company claims. ↩ ↩² ↩³ ↩⁴

16. Source: ABBYY, “Vantage”, accessed Oct 2026. Accuracy and skills figures are company claims. ↩ ↩² ↩³ ↩⁴

17. Source: AWS, “Amazon Textract FAQs”, accessed Oct 2026. ↩ ↩²

18. Source: UiPath Documentation, “IXP licensing and pricing FAQ” and UiPath, “Pricing”, accessed Oct 2026. UiPath’s docs note that consumption is charged per operation at runtime; rates have changed across versions, so confirm with UiPath. ↩ ↩²

19. Source: Rossum, “Pricing”, accessed Oct 2026. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶

20. Source: AWS Marketplace, “Hypercell - SaaS” (sold by Hyperscience), accessed Oct 2026. Listed as “HS Private Cloud Professional”; private offers may differ. ↩ ↩² ↩³

21. Source: AWS, “AI services opt-out policies”, accessed Oct 2026. ↩ ↩² ↩³

22. Source: Microsoft Learn, “Data, privacy, and security for Document Intelligence”, accessed Oct 2026. ↩ ↩²

23. Source: Microsoft Learn, “Use Document Intelligence containers in disconnected environments”, accessed Oct 2026. ↩

24. Source: Google Cloud, “Document AI regional and multi-regional support”, accessed Oct 2026. ↩