Intelligent document processing for invoices and contracts

Intelligent document processing is the full pipeline that turns a document into checked, usable data: classify it, extract the fields, validate them and post them where they belong. OCR is one step inside it. OCR turns an image of text into characters. It does not know that the number at the bottom right is the invoice total, or that the total should equal the sum of the lines.

Current IDP pairs a document AI service, such as Azure Document Intelligence, Amazon Textract or Google Document AI, with language models that read a page much as a person would. The model handles the variety. The validation rules decide whether to trust it.

For the short definition and related terms, see the glossary entry on intelligent document processing. For a worked example on supplier invoices, from inbox to accounting system, read the invoice processing pipeline blueprint.

Accurate enough to post most clean invoices without review, and never accurate enough to skip validation. The real answer depends on your documents, so we measure it on them before you commit to a build.

During Prove we run extraction on 100 to 300 of your labeled documents and report accuracy for each field separately, because a single blended score hides the field that matters. Header fields such as supplier name, invoice number, date and total are usually the easiest. Line items are harder, especially multi-page tables, merged cells and descriptions that wrap. Invoices with several tax rates are another common source of errors.

What lowers accuracy:

• Poor scans, such as phone photos taken at an angle, faxes and low-resolution images.
• Handwriting. Current vision models read neat handwriting reasonably well. Scrawl, stamps over text and crossed-out values go to review.
• Unusual layouts, such as sideways pages or invoices with several separate tables.
• Number and date formats from other countries, for example 1.234,56 or 03/04 meaning April.

What raises it: asking for fewer fields, validating against your supplier master and purchase orders, and sending each document type to whichever method scored best on it.

The figure your finance team will care about is the straight-through rate: the share of documents that post with no human touch and every value correct. We report it next to field accuracy and agree a target with you during Map.

It stops and asks a person. Every extracted value carries a confidence signal, and every document passes validation rules before anything is posted. If either check fails, the document goes to the review queue with the doubtful fields marked.

Typical validation rules for supplier invoices:

• Line items add up to the subtotal, and subtotal plus tax equals the total.
• The supplier exists in your vendor master, matched on tax ID or bank details.
• The invoice number has not been seen before for that supplier, which catches duplicates.
• Any purchase order number exists, is open and matches the amount within your tolerance.
• Bank details match the ones on file. A change always triggers a hold, because altered bank details are a common sign of invoice fraud.

The reviewer sees the page on one side and the fields on the other. Clicking a field jumps to where it sits on the page, and the reviewer corrects or approves it. Every correction is logged. Once a month we review the log, and where a pattern shows up, such as one supplier's layout tripping the same field, we adjust the prompt or add a rule. That loop is human-in-the-loop review doing its job.

Thresholds are yours to set. A finance team might let invoices under $2,000 from known suppliers post on their own and review everything above that.

Building one document family to production costs $12,000 to $40,000 and takes 6 to 12 weeks. A document family is one kind of document with one destination, such as supplier invoices into NetSuite or employment contracts into a contract register. Additional families on the same pipeline cost less than the first, because intake, review and monitoring already exist.

The price moves with:

• Layout variety. 20 suppliers with similar invoices sit at the low end. 800 suppliers in 6 languages sit at the top.
• Line items. Header fields alone are simpler than full line-item extraction with product codes matched to your catalog.
• The destination. Posting to Xero or QuickBooks Online through their APIs is straightforward. An on-premise ERP with batch imports takes longer.
• Review design. Using an approval tool you already have costs less than a custom review screen with role-based access.
• Data residency and retention rules.

Running costs are separate and paid by you: the extraction service charges per page, model calls are billed per use and hosting is usually small. Published per-page prices for IDP software run from about $30 to $560 per 1,000 pages, depending on vendor and volume, according to a September 2026 comparison by Parseur, itself an IDP vendor (Parseur IDP pricing (opens in a new tab)). We estimate your monthly running cost from your own volumes during Map, so you can set it against the hours spent keying documents today.

When the data already exists in structured form. If suppliers can send e-invoices as UBL or Peppol XML, or your customers can fill out a web form, parse that directly. It costs a fraction of an IDP build and has no extraction errors. Ask your 10 largest suppliers before you commission anything.

IDP is also hard to justify when:

• Volume is low. A few dozen documents a month rarely repay a $12,000 build. A shared inbox and a checklist may be enough.
• Every document is unique and needs judgment, such as one-off contracts under negotiation. We can pull key terms into a register, but deciding whether a clause is acceptable is a lawyer's job. The contract review blueprint shows where that line sits.
• People mainly want to ask questions across thousands of policies or manuals. That calls for a knowledge system that answers with citations.

Not included in an IDP project:

• Scanning hardware, mailroom services or digitizing a paper archive.
• Records management and retention policy design.
• Legal advice on contract terms.
• Per-page software fees and model usage, which you pay directly at cost.
• Cleanup of your vendor master or chart of accounts beyond what validation needs.