What is intelligent document processing, and when is it worth it?

Intelligent document processing (IDP) is software that reads business documents and turns them into checked, structured data. It works out what each document is, extracts the details you need, validates them against your own rules and passes the result to your systems, sending anything it is unsure about to a person. It combines text recognition, AI models and a workflow with human review. Invoices, onboarding forms and statements make good first candidates.

How IDP differs from OCR and templates

• OCR (optical character recognition) turns an image of a page into text. It doesn’t know which number is the invoice total and which is the VAT registration.
• Templates tell software where to look on a known layout: the total is always in this box. They work until a supplier changes its invoice design, and then they quietly break.
• IDP, sometimes called AI document processing, uses AI models that read layout and meaning, so it can find the total on a layout it has never seen. It also reports how confident it is in each field, which is what makes a review step possible.

IDP still uses OCR underneath for scanned pages. The difference is what happens after the text is read.

How intelligent document processing works

1. Capture. Documents arrive by email, upload, scanner or shared folder, and are converted into a consistent format.
2. Classify. The system decides what each document is: invoice, credit note, bank statement, passport, application form. Files holding several documents are split.
3. Extract. It pulls out the fields that matter for that type, such as supplier, date, amount and line items.
4. Validate against your rules. The extracted data is checked. Do the line items add up to the total? Does the purchase order exist? Is the supplier on your approved list? These checks are written in code, not left to the model.
5. Human review of low-confidence cases. Anything that fails a rule or falls below a confidence threshold goes to a person, who sees the document beside the extracted data and corrects it.
6. Hand-off to systems. Clean data goes into your accounting package, CRM or case system, with a record of what was extracted, checked and changed.

In the early weeks, the corrections people make in review are the most valuable output. They show where the system goes wrong and which rules are missing.

Good first documents

• Supplier invoices. Invoice automation is a common starting point: plenty of volume, a familiar structure and simple checks against purchase orders and totals.
• Onboarding forms. Client or investor application forms, identity documents and proof of address, where the job is to extract the details and list what is missing.
• Statements. Bank, broker or custody statements that need to be read and reconciled against your own records.

A good first document type arrives often, follows a recognisable shape and has rules you can write down. If your team can’t say what makes a document correct, the software can’t either.

What makes it fail

• Poor scans. Skewed, faint or photographed pages, handwriting, and stamps over key fields. Fix capture at the source where you can.
• Edge cases. Credit notes that look like invoices, several documents in one PDF, foreign currencies, documents in other languages. Most of the effort goes on these, not on the typical document.
• No review loop. A system with no confidence threshold and nobody checking the exceptions will quietly pass errors into your books.
• No rules. Extraction without validation only moves the checking work further down the line.
• Testing on the easy cases. Judge it on a sample of real documents from a busy month, messy ones included.

What it looks like on regulated documents

At Alter Domus, where I was Head of AI & Automation, document AI classified regulated documents across 13 categories and 73 sub-categories, straight through. The hard part of work like that is rarely reading the text. It is agreeing the categories with the people who use them, deciding what happens to a document that fits none of them, and showing that the results hold up on real volumes. More of that track record is on the AI for financial services page, and the operations view of AI in asset management shows where document work fits in a wider back office.

Build or buy?

Buy when your documents are common types, such as invoices, receipts or identity documents, and an established product already handles them and connects to your accounting or case system. Some accounting packages include invoice and receipt capture, so check what you already pay for first.

Build, usually on a general AI model and your existing workflow tools, when the documents are specific to your sector, the rules are complex, the output has to fit an existing process, or you need tight control over where the data goes. Regulated documents and fund paperwork often fall here.

Many projects end up in between: a bought capture and extraction service with your own rules, review screen and integrations around it. Whichever route you take, ask:

• Does it give a confidence score for each field, and can we set the threshold?
• What does the review screen look like for the person correcting errors?
• Where are documents processed and stored, and for how long?
• Can we export our data and rules if we leave?
• How does the price change as volume grows?

The guide to what AI automation costs covers the running costs to plan for.

Who it suits, and who it doesn’t

IDP pays back when a team spends a real part of its week reading documents and re-typing what they say. If you receive a handful of documents a week, a person is cheaper and more accurate. If every document is different and each one needs expert judgement, use AI to prepare a summary for the expert instead.

Document workflows are a core part of my agents and automation work. If your team re-types the same kinds of document every week, book a free 30-minute AI consultation and we can look at whether IDP would pay back for you.