What is the best software to extract data from documents in 2026?

Azure AI Document Intelligence is Microsoft's document extraction service, and it runs on models. The prebuilt ones already understand an invoice, a receipt, an ID or a generic layout: send the PDF, get JSON back, with a confidence score on every field. For your own forms, you train a custom extraction model on a handful of labeled samples, and a classification model can sort a mixed pile before extraction.

Pricing is public and per page, which is rarer than it should be. Read (plain text) costs $1.50 per 1,000 pages, dropping to $0.60 past one million. Prebuilt models cost $10 per 1,000 pages, custom extraction $30, classification $3. Training runs $3 an hour, with the first 10 hours free. Run 10,000 one-page invoices a month through the invoice model and you're looking at $100.

You also pick the region, including Western Europe if you serve EU customers, and containers let you run extraction outside Microsoft's cloud when sensitive documents can't leave your infrastructure.

The catch: it's a service, not an app. There's no review screen your AP clerk can open on day one. You build it, or you wire Azure into an automation tool. With no developer around, Parseur gets you running faster.

Tip: start on the free F0 tier. 500 pages a month is plenty to push 30 of your real invoices through the prebuilt model and count correct fields before you spend a cent.

Mistral is a French AI lab. Its document model, Mistral OCR 3, reads PDFs and images, handwriting included, and keeps the structure: headings, tables, columns. Reading costs $2 per 1,000 pages, $1 in batch mode if you can wait for results, and annotations, which return the fields as JSON, cost $3 per 1,000 pages: the cheapest field extraction in this ranking.

Let's be clear on an extraction page: Mistral OCR 3 reads very well, but you describe the fields you want and you check the output. No prebuilt invoice model, no review screen. It's for teams with a developer who would rather pay very little per page and build the rest.

A newer version, Mistral OCR 4.1, is listed at $4 per 1,000 pages. On Hacker News, feedback on recent versions is split: some users are happy with it on their documents, others caught the model inventing whole sentences mid-page. That's the risk with any generative model: it doesn't leave a box empty, it fills it.

Tip: on invoices, check the math automatically. If line items don't add up to the subtotal, or subtotal plus tax doesn't equal the total, route the invoice to human review. That one check catches a good share of generative mistakes.

Rossum is a full platform for vendor documents. Invoices come in by email, API or upload, AI extracts the fields, and your team validates on a screen built for it: low-confidence fields are flagged, and a fix takes one click. Rossum learns from those corrections, which matters when every vendor has its own layout.

It's also the only platform at this level that publishes an entry price: Starter from $18,000 a year, with unlimited seats, extraction, the validation screen, a 12-month archive and API access. Higher tiers, on quote, add master data matching, duplicate detection, ERP connectors, SSO, a sandbox and a 3-year archive on Ultimate. Contracts run one year minimum; go over your volume and you either move up a plan or pay per page.

Data sits in Frankfurt by default, with a backup in Ireland, and Rossum also runs US and Japan regions. It holds ISO/IEC 27001 certification.

The catch: the ticket. At $18,000 a year, Rossum only pays off with real invoice volume and a team that spends its days keying them in. For a few hundred documents a month, it's overkill.

Tip: during the 14-day trial, send invoices from your 20 messiest vendors, not your cleanest. Count how many fields you fix in week one, then week two. That's where you see whether the learning actually works.

Amazon Textract is AWS's document extraction service, split into features with their own price tags: plain text, tables, forms (key-value pairs), Queries in plain English ("what's the PO number?"), an expense API for invoices and receipts, one for IDs and one for lending packets.

That's where the bill can surprise you. Plain text costs $1.50 per 1,000 pages. Turn on Forms, Tables and Queries together and you're at $70 per 1,000 pages, 47 times more. For invoices, Analyze Expense at $10 per 1,000 pages already returns vendor, totals and line items, and is often enough.

Developers note that Textract carries surprising inertia: many teams wired it in years ago and never revisited the choice. It's a solid service, but price it against Azure or Mistral before you start a new project.

The catch: like Azure, it's for developers, with no ready-made review screen, and your bill depends on which features you switch on.

Tip: never enable every feature by default. Start with Analyze Expense on invoices and add Queries only for the one or two fields it misses.

Parseur starts from a simple idea: forward your emails and PDFs to a dedicated address and it pulls out the fields. You can guide it with templates (show it once where each field lives) or let the AI find fields on its own. Data then flows to Excel, Google Sheets, a webhook, or automation tools like Zapier and Make.

Pricing is public and tied to pages per month: 20 free pages to try, then Micro at $49 a month for 100 pages, Mini at $89 for 300, Starter at $129 for 1,000, Premium at $269 for 3,000 and Pro at $499 for 10,000. Paying yearly lowers the bill: $468 a year for Micro, $1,188 for Starter.

Data handling is a strong point: servers in the EU, an ISO 27001 certified data center, and adjustable retention. Handy if you have European customers.

The catch: price at volume. At 10,000 pages a month you pay $499, while the same extraction costs $100 on Azure's invoice model. Parseur charges for simplicity, which is fair as long as you don't have a developer. Public reviews are few, but they mostly praise the support.

Tip: set retention to the shortest option on day one. Once the data is in your spreadsheet, there's no reason to keep invoices stored with a third party.

Nanonets no longer sells pages, it sells workflows. You chain blocks: intake, extraction, checks, export. Each block run burns credits: $0.02 for a simple block, $0.10 for standard AI, $0.30 for complex AI like extraction. Nanonets estimates a typical invoice runs through 4 to 6 blocks, for under $2 end to end.

Starter is free: $50 in credits with no card, up to three users, then $100 a month for 100 credits. Credits never expire, which is rare here. Growth, on quote, brings volume discounts of up to 40%. Enterprise adds SSO, HIPAA and SOC 2, private cloud or on-prem, and data residency in the US, EU or APAC.

The catch: cost per document. Under $2 an invoice is fine for a few hundred a month, far less so at 10,000. One public review also complains that files are kept for AI training: read the terms before sending anything sensitive.

Tip: price your own workflow before you pick. Count the blocks a typical invoice actually runs, multiply by monthly volume, and compare with a per-page API.