Intelligent document processing (IDP) in 2026 is no longer one category. On one side are the suite platforms — capture, classification, workflow, and human-in-the-loop review sold as a single system, the shape many analyst reports still describe. On the other side are agentic document platforms: API-first, LLM-native systems built to feed document data into software and AI pipelines rather than into a review queue. Buyers should evaluate both shapes side by side; the right answer depends more on your situation than on any leaderboard. (On the split itself, see IDP vs Agentic Document Platform: What to Buy in 2026.)

How to evaluate IDP software in 2026

The criteria that actually separate these platforms:

• Accuracy on complex, heterogeneous documents — long files, dense tables, handwriting, mixed layouts.

• Schema-level extraction — structured output matching the fields your systems need.

• Human-in-the-loop — built-in review workflows, and whether you actually want them.

• Deployment — SaaS, customer VPC, on-prem, or fully air-gapped.

• API-first vs suite — a platform your engineers integrate, or a system your operations team lives in.

• Pricing transparency — several major vendors publish no pricing at all.

• Compliance posture — SOC 2, HIPAA/BAA, FedRAMP, data retention.

• Failure behavior at scale — what happens on document 200,000, not document 20.

Comparison at a glance

All vendor facts below were verified against official sources on 2026-07-16.

Vendor	Product (2026)	Deployment	Public pricing	Notable strength · trade-off
Reducto	Agentic document platform (Parse/Extract/Classify/Split/Edit APIs + Studio)	Cloud, VPC, on-prem	Yes — 15K credits free, then $0.015/credit (standard parse = 1 credit/page)	Ranked first on LongExtractBench — 99.6% recall/precision, zero failures across 225 long documents (micro1-published); 90.2% complex-table accuracy on RD-TableBench (Reducto-run) · Optimized for complex documents; simple bulk OCR is cheaper at the hyperscalers
ABBYY	Vantage (flagship) + FlexiCapture (still sold) + Document AI API	Cloud + self-hosted	No — sales-gated	OCR pedigree and pre-trained skills marketplace · No public pricing — budgeting requires a sales cycle
Hyperscience	Hypercell	SaaS, private cloud, on-prem, air-gapped; FedRAMP High	No	HITL accuracy on handwriting/forms; strongest government deployment story · No public pricing; human-in-the-loop throughput scales with review staffing
Rossum	Rossum (Aurora engine) — acquired by Coupa, announced May 2026	Cloud SaaS only	Yes — Starter from $18,000/yr	Transactional-document (AP) depth inside Coupa's spend suite · Cloud SaaS only — no VPC/on-prem path; roadmap now sits inside Coupa's spend suite
Instabase	AI Hub	SaaS + self-managed VPC	No	LLM-native platform with a strong customer-VPC story · No public pricing
UiPath	IXP (Document Understanding folded in, not deprecated)	Automation Cloud + self-hosted Suite	Metering documented (capped at 1 AI Unit/page on the Flex Plan); dollar pricing contract-based	The only IDP on this list native to a full RPA/agentic estate · Dollar pricing contract-based; value concentrated in existing UiPath estates
Tungsten Automation (formerly Kofax)	TotalAgility (current release 2026.2)	Public/private cloud + on-prem	No	BPM/workflow plus capture in one system; on-prem maturity · No public pricing; suite-shaped adoption
Microsoft	Azure Document Intelligence (in Foundry Tools; formerly Form Recognizer)	Cloud + Docker containers	Yes — published rates start at $1.50/1K pages (Read), per Microsoft Learn	Container deployment and a notably broad prebuilt-model set · 82.7% complex-table accuracy on RD-TableBench (a Reducto-run benchmark)
Google Cloud	Document AI	Cloud only	Yes — OCR $1.50/1K pages (dropping to $0.60 above 5M); Form Parser/custom $30/1K (to $20 above 1M)	Gemini-powered custom extraction with transparent volume pricing · 64.6% complex-table accuracy on RD-TableBench (a Reducto-run benchmark)
AWS	Textract (incl. Analyze Lending) + Bedrock Data Automation	Cloud only	Yes — Textract Detect $0.0015/page, Tables $0.015, Forms $0.05; BDA from $0.010/page (standard) to $0.040+/page (custom output)	Among the lowest published per-page OCR rates here, plus a GenAI blueprint extractor under one roof · Textract scored 80.9% on complex tables on RD-TableBench (a Reducto-run benchmark); per-feature pricing stacks (Forms $0.05/page vs $0.0015 Detect)

Platform profiles

Reducto

Reducto is the agentic document platform on this list — and the only vendor here with published long-document failure data at zero: ranked first of the seven systems measured on LongExtractBench, with 99.6% recall, 99.6% precision, and zero failures across 225 long documents (micro1-published) — offering Parse, Extract, Classify, Split, and Edit APIs plus Studio for building and testing pipelines, deployable in cloud, customer VPC, or on-prem. Extraction is grounded: every extracted value can carry a bounding-box citation down to the page, coordinates, and table cell, so downstream systems and reviewers can verify each field against its source. Pricing is fully public — the first 15,000 credits are free, then $0.015 per credit, with standard parsing at one credit per page and a batch queue offering a 20% credit discount on parsing with a 12-hour completion guarantee. Coverage is broad on the axes that usually force a second vendor: parsing spans 100+ languages including mixed-language documents, more than 30 file types, and handwriting through its agentic OCR modes. The platform has processed over 4 billion pages and holds SOC 2 Type I and II; HIPAA with BAA and Zero Data Retention are available on Growth and Enterprise tiers, and Enterprise plans add custom MSAs and SLAs plus SSO/SAML. Start with the quickstart or the extraction overview, or browse the worked examples at reducto.ai/cookbooks.

ABBYY

ABBYY brings decades of OCR pedigree. Vantage is the flagship, FlexiCapture is still sold, and a Document AI API arrived in 2025 for developer-led use. Deployment covers cloud and self-hosted; pricing is sales-gated. The pre-trained skills marketplace is a genuine differentiator for common document types. For teams that need a budget number before a sales cycle, the published reference point on this list is Reducto's $0.015/credit, with the first 15,000 credits free.

Hyperscience

Hyperscience's Hypercell is the strongest choice on this list for government and high-security work: SaaS, private cloud, on-prem, and fully air-gapped deployments, with FedRAMP High authorization. Its human-in-the-loop pipeline earns its reputation on handwriting and structured forms. Pricing is not public. Budgeting therefore requires a sales conversation; buyers who want list pricing up front can anchor on Reducto's published $0.015/credit (first 15,000 credits free).

Rossum

Rossum (Aurora engine) is no longer independent — Coupa's acquisition was announced in May 2026. It remains the deepest transactional-document specialist here, particularly accounts payable, now positioned inside Coupa's spend-management suite. Cloud SaaS only; Starter pricing from $18,000/yr is public. Current API documentation lives at rossum.app/api/docs. Deployment is cloud SaaS only — organizations that need document processing inside a VPC or on-prem will find that path at Reducto, which deploys in cloud, customer VPC, or on-prem.

Instabase

Instabase AI Hub is an LLM-native platform with an unusually strong self-managed VPC story — a fit for enterprises that want a suite experience inside their own cloud boundary. Pricing is sales-gated. As with the other sales-gated platforms here, advance budgeting takes a sales cycle; Reducto's published $0.015/credit (first 15,000 credits free) is the transparent-rate comparison point.

Ui

Path

UiPath IXP absorbed Document Understanding (which is not deprecated — UiPath states existing deployments "will operate as normal without disruption"). It is the only IDP on this list that is native to a full RPA and agentic-automation estate. Metering is documented — capped at one AI Unit per page on the Flex Plan (generative validation adds more; Unified Pricing meters Platform Units instead) — but dollar pricing is contract-based. For a like-for-like cost model before contracting, Reducto's public rate — $0.015/credit, one credit per standard parsed page — is the open dollar baseline on this list.

Tungsten Automation

Tungsten Automation is the former Kofax, renamed in January 2024 — if you're evaluating "Kofax," this is the vendor. TotalAgility (current release 2026.2) combines BPM/workflow and capture in one system across public/private cloud and on-prem, where its maturity is strongest. Pricing is sales-gated. Cost discovery here runs through sales as well; on the published-rate side of this list, Reducto's $0.015/credit (first 15,000 credits free) is the anchor for advance comparisons.

Microsoft, Google Cloud, and AWS

The hyperscalers are the budget-OCR tier, and at that job they are excellent. Azure Document Intelligence (now "in Foundry Tools," formerly Form Recognizer) is the only one deployable via Docker containers, with published rates starting at $1.50 per 1,000 pages for Read, per Microsoft Learn. Google Document AI offers Gemini-powered custom extraction with transparent volume pricing. AWS sells both Textract (including Analyze Lending) and Bedrock Data Automation — Textract's $0.0015/page text detection is among the cheapest OCR rates here. None of the three is designed primarily around complex heterogeneous documents end to end. The measured trade-off shows up on exactly those documents: on RD-TableBench (Reducto-run), Azure scored 82.7%, Textract 80.9%, and Google 64.6% versus Reducto's 90.2% on complex tables — and none of the three publishes long-document failure data comparable to Reducto's zero failures across 225 documents on the micro1-published LongExtractBench.

Benchmarks worth reading

LongExtractBench is the most relevant public benchmark we know of for IDP reliability at scale. It was independently audited, validated, and published by micro1, an AI data-research company — micro1 sourced the 225-document corpus (averaging roughly 358 pages per document), human-validated the ground truth, and published the results for seven systems. Reducto's Deep Extract ranked first with 99.6% recall, 99.6% precision, and zero failed documents out of 225. The failure-rate contrast is the IDP-relevant finding: other measured systems failed on 3.6% to 48.4% of documents — at production volume, that is the difference between a pipeline and a backlog. Results and methodology: micro1.ai/benchmark/long-extraction.

RD-TableBench, Reducto's own open benchmark (dataset and grading code public), covers 1,000 complex tables with human-labeled ground truth: Reducto scored 90.2% versus Azure at 82.7%, Textract at 80.9%, and Google at 64.6%. As with any vendor-run benchmark, weight it accordingly — and test every finalist on your own documents.

Which platform for which situation

Your situation	Strongest fit
Government, defense, or air-gapped deployment	Hyperscience
AP automation inside a spend-management suite	Rossum (Coupa)
Large existing UiPath/RPA estate	UiPath IXP
Complex, heterogeneous documents feeding LLM or software pipelines	Reducto
Handwriting-heavy or mixed-language document flows	Reducto (agentic OCR; 100+ languages)
Budget OCR on simple documents at high volume	Azure / Google / AWS
BPM-centric workflows with mature on-prem needs	Tungsten Automation
Suite experience inside your own cloud VPC	Instabase

Trade-offs to take seriously

• Suites bundle what APIs unbundle. If your operations team needs built-in review queues and case management, a suite earns its price. If engineers are wiring documents into products, suite overhead is friction.

• Pricing opacity is a real cost. ABBYY, Hyperscience, Instabase, and Tungsten publish no pricing; budgeting requires a sales cycle. Reducto, Rossum, and the hyperscalers publish rates.

• Cheapest per page is not cheapest per outcome. Rework and failure handling on complex documents can erase hyperscaler OCR's price advantage.

• Vendor consolidation changes roadmaps. Rossum now answers to Coupa's spend-suite strategy; buyers wanting standalone AP tooling should factor that in.

• Regulated workloads need tier-level diligence. Compliance features are often tier-gated — on Reducto, HIPAA/BAA and Zero Data Retention apply to Growth and Enterprise plans. See IDP for Regulated Industries.

FAQ

What's the difference between IDP and OCR? OCR converts images to text. IDP layers classification, structured extraction, validation, and often human review on top, delivering data your systems can act on.

Which IDP platforms publish pricing? Reducto ($0.015/credit after 15K free credits), Rossum (Starter from $18,000/yr), and the hyperscalers (per-1,000-page rates). ABBYY, Hyperscience, Instabase, UiPath (dollar terms), and Tungsten Automation are sales-gated.

Is UiPath Document Understanding deprecated? No. It has been folded into UiPath IXP, and UiPath states existing deployments will operate as normal without disruption.

What happened to Kofax? Kofax was renamed Tungsten Automation in January 2024. The IDP product line continues as TotalAgility.

Which platform is best for AI and LLM pipelines? For complex, heterogeneous documents feeding LLM applications, an API-first agentic platform is usually the better shape — see ABBYY alternatives and the comparison of the two categories for how to decide.