Best intelligent document processing software in 2026

Three different purchases hide behind one acronym

Intelligent document processing, or IDP, covers three different purchases. A cloud API from Microsoft, Amazon or Google reads a page and returns fields, and a developer builds everything else. A packaged platform such as ABBYY Vantage, Rossum, Nanonets or Docsumo adds the inbox, review screen and ERP export. The newest group, Mistral OCR, LlamaParse and Reducto, turns any page into clean text or JSON for around a cent a page. Buying the wrong kind costs more than buying the wrong brand.

I build and run production AI agents that read invoices, purchase orders, delivery notes and timesheets for operational teams, so my bias is towards buying the reading and owning the checking. No vendor has paid to be here. Prices are list prices in US dollars, checked on vendor pages on 25 September 2026, so confirm them before you budget.

1. Azure Document Intelligence in Foundry Tools

Microsoft's document service began as Form Recognizer and is now called Azure Document Intelligence in Foundry Tools. It reads printed and handwritten text and tables, ships prebuilt invoice and receipt models, and trains custom models on your layouts. Its sibling, Azure Content Understanding, generally available since its November 2025 API, adds language-model analysers for invoices and purchase orders with confidence scores.

It wins when you run on Microsoft and have a developer, and is the wrong buy if nobody will build the inbox, review screen and ERP posting. List prices are 1.50 dollars per 1,000 pages for Read, 10 dollars per 1,000 for the prebuilt invoice model and 30 dollars per 1,000 for custom extraction, with 500 free pages a month. The old v2.1 API retires on 15 September 2027.

2. Amazon Textract

Textract is the AWS document reader. Analyze Expense is the part that matters here, pulling supplier, totals and line items from invoices and receipts, while Analyze Document handles forms, tables and queries. Since March 2025 AWS has also sold Bedrock Data Automation, which uses generative models and your own output schema.

It wins for AWS shops, and on price. Analyze Expense costs 1 cent per page for the first million pages a month, plain text 1.50 dollars per 1,000 pages, and Bedrock Data Automation 1 cent per page, or 4 cents with your own schema. New AWS customers get a three-month free tier with 100 Analyze Expense pages a month. Without engineers it is the wrong buy, since the inbox, review queue and posting are yours to build.

3. Google Document AI

Google Document AI offers OCR, form and layout parsers, prebuilt invoice and expense parsers, and a Custom Extractor that runs on Gemini models and works from a field list and a few examples rather than a large training set. In June 2026 Google added validation rules to it in preview.

It wins for Google Cloud teams with varied layouts and a developer. The prebuilt Invoice parser costs 10 cents per document of up to ten pages, while the Custom Extractor costs 30 dollars per 1,000 pages, under a third of the price on one-page invoices. Google retires extractor versions about a year after release and says some Gemini 3 versions are not covered by data residency, so budget for re-testing. The first 1,000 OCR pages are listed free, then 1.50 dollars per 1,000.

4. ABBYY Vantage

ABBYY is one of the oldest names in document capture, and Vantage is its low-code IDP platform, with pre-trained skills for invoices, purchase orders and transport documents, and a human review step. Vantage 3.0, released in January 2026, added prompt-based extraction through Azure OpenAI, redaction and a dashboard of touchless rates.

It wins for mid-size and larger organisations with many document types that want one supported platform with review built in. It is the wrong buy for a small team with one invoice flow, because you pay for breadth you will not use and often for a partner to implement it. ABBYY publishes no prices, so request a trial through its marketplace and expect a custom quote.

5. Mistral OCR

Mistral OCR is an API that turns PDFs and scans into clean text, tables and structure, and returns the fields you define as JSON. OCR 4 and 4.1, released in June and July 2026, add bounding boxes, block labels and confidence scores, and a score is what lets you send a doubtful field to a person.

It wins when you have a developer and want good reading of messy scans at the lowest price. OCR 4.1 lists at 4 dollars per 1,000 pages, about half that in batch, and 5 dollars per 1,000 with annotations. The free plan includes 10 dollars a month of API credits. It is the wrong buy for anyone who wants a finished product, because there is no inbox, review screen or ERP connector.

6. Nanonets

Nanonets is a no-code platform where you build a document workflow from blocks, such as extraction, classification, validation and export, connected to email, cloud storage and, on higher plans, your ERP. It has priced per block run since 31 January 2025.

It wins for small teams starting without a sales call. New accounts get 50 dollars in credits, down from 200 dollars as recently as May 2026, then 100 dollars a month buys 100 credits. Blocks cost 2 cents for simple steps, 10 cents for validation and 30 cents for extraction, and Nanonets puts a typical invoice at four to six blocks, under 2 dollars. That suits a few hundred documents a month and is the wrong buy at tens of thousands on one layout, where a cloud API costs around a cent a page.

7. Rossum

Rossum is built for transactional documents, invoices above all, plus sales orders, packing lists and certificates of analysis. It runs on its own transactional language model, Aurora, around a validation screen where a person confirms what the model is unsure of. On 12 May 2026 Coupa, the spend management company, announced it had acquired Rossum.

It wins for accounts payable and order desks handling thousands of documents a month, above all on Coupa. Starter is published from 18,000 dollars a year with unlimited seats, and there is a 14-day trial. Master data matching and duplicate detection, the checks that make an extraction safe to post, sit in the custom-quoted Business plan and above. If you are not on Coupa, ask about the roadmap for your ERP before a multi-year deal. It is the wrong buy below a few hundred documents a month.

8. Docsumo

Docsumo is a no-code extraction platform with pre-trained models, an AI document reviewer and validation rules. Its pages lean towards lending and insurance, listing bank statements, pay stubs and tax forms beside invoices.

It wins for finance teams whose documents look like a lender's file, and for testing before any sales call, since the free trial covers up to 1,000 pages over 14 days. Paid plans are custom quotes on page volume, with setup fees on top, and cross-document validation and master data lookup sit in Enterprise. It is the wrong buy if your volume is mostly purchase orders and delivery notes, because its prebuilt models concentrate on financial documents.

9. LlamaParse and Reducto

These are developer platforms for feeding documents into AI systems. LlamaParse, from LlamaIndex, turns more than 130 file formats into markdown or JSON and adds extraction with citations and confidence scores. Reducto sells parse, extract, split and classify endpoints, and on 1 September 2026 launched its own r-1 parsing model in preview at a flat cent a page.

They win when you are building your own pipeline. LlamaParse gives 10,000 free credits a month, paid plans start at 50 dollars a month, credits cost 1.25 dollars per 1,000, and agentic parsing uses 10 credits a page, about 12.50 dollars per 1,000 pages. Reducto gives 150 dollars of free usage, then charges 10 dollars per 1,000 pages to parse and 20 to extract. Both are the wrong buy without a developer. With one and a tight budget, the open-source Docling, started at IBM Research, and LlamaIndex's LiteParse cost only the machine they run on.

What changed in 2025 and 2026

Reading a page got cheap, with plain text at 1.50 dollars per 1,000 pages from all three clouds and full layouts from Mistral and Reducto for a cent or less. Coupa bought Rossum in May 2026, ABBYY shipped Vantage 3.0 in January 2026, and the Mistral OCR list price went from 1 dollar per 1,000 pages in March 2025 to 4 dollars in June 2026. AWS launched Bedrock Data Automation, Microsoft moved its service under Foundry Tools and made Content Understanding generally available in November 2025, Google discontinued its legacy processors on 30 June 2026, and Nanonets cut its sign-up credit. Lists older than a year get several of these wrong.

The enterprise suites

Gartner published its first Magic Quadrant for intelligent document processing on 3 September 2025 and its second on 8 September 2026, and ABBYY, Hyperscience, Tungsten Automation and UiPath each announced they were named Leaders in both. UiPath sells Document Understanding through IXP and meters extraction at 0.2 platform units per page, with no published unit price. Hyperscience sells its Hypercell platform to large enterprises and government, and Tungsten Automation, called Kofax until January 2024, sells TotalAgility for cloud or on-premises use. Like Instabase, none publishes a per-page price.

Reading the page is the easy part, checking it is the hard part

Every tool above can read a clean invoice. What decides whether you save time is what happens after the read, and I learned that building payroll and order-entry systems. The expensive errors were rarely a misread digit. They were correct readings matched to the wrong pay rate, customer or product.

The checking comes in three layers. First, arithmetic. Quantity times unit price equals each line total, lines plus VAT equal the invoice total, and daily hours on a timesheet add up to the weekly figure after breaks. Second, matching against records you already hold. The supplier exists with unchanged bank details, the purchase order is open and agrees with the invoice and delivery note, each customer description maps to one catalogue item in the right unit of measure, and the worker and pay rate are current in payroll. Third, routing. Anything that fails goes to a named person with the page image beside the flagged field and a plain reason, never to a generic error folder.

Keep those checks in deterministic rules against your own master data, not in another model prompt, and notice where vendors put them. Rossum keeps master data matching for its Business plan, Nanonets puts ERP and database integrations in Growth, Docsumo keeps cross-document validation for Enterprise, and the Google extractor validation rules were still in preview in September 2026. Reading is the entry tier and checking is the upsell. Price the tier you will need, and track the share of documents a person still touches each week, not a headline accuracy figure.

Comparison at a glance

How to pick by volume and team

Below a few hundred documents a month, do not buy a platform yet. Try free tiers on your worst documents, the crumpled delivery note and the handwritten timesheet, not the vendor's sample invoice. Nanonets and Docsumo show what a finished tool feels like, and Azure, AWS or Mistral show the raw reading for pennies if someone can write a little code.

From a few hundred to tens of thousands a month, the choice is build or buy. With a developer or partner, a cloud API or Mistral OCR plus your own checks is cheapest and keeps the rules in your hands. Without one, Rossum, Nanonets or ABBYY give you the inbox and review screen, so ask each which plan includes matching against your master data. Above that, or if you already run UiPath, Tungsten or Hyperscience, the enterprise suites earn their implementation cost.

At every size, run a month of real documents through your shortlist, count how many a person still touches and why, and decide on that number. If a vendor cannot show you what happens to the documents it gets wrong, it is not ready.