Intelligent Document Processing: How AI Reads Business Paperwork

What intelligent document processing is

Intelligent document processing is software that reads business documents the way a clerk would, pulls out the fields that matter and passes them on to the next step.

Every business receives paper that someone has to read: supplier invoices, credit notes, delivery notes, order confirmations, receipts, forms from new clients. The information is all there, but in a shape a computer cannot use until someone types it in. Intelligent document processing does the reading. The name is often shortened to IDP, and you will also see AI document processing, document processing automation and AI document automation used for much the same thing.

Whatever the label, and people search for it as automation document processing too, the job breaks into the same parts: collect the document, work out what it is, read the fields, check them, and send the result somewhere useful. The hard part is rarely the reading. It is knowing when a reading should not be trusted.

What makes it intelligent rather than plain scanning is that it understands which number is the total, which date is the invoice date and which name is the supplier, even when the layout is new. What makes it useful is what happens next: the fields are checked against what you already know before anything is entered. The table shows common documents, the fields usually taken from them and what they are checked against.

OCR, templates and AI: three ways to read

Documents are read in three ways: templates for layouts that never change, OCR for images, and an AI model for everything else.

Each method has a place, and good intelligent document processing uses all three rather than one. Our supplier bills robot, running today for a client, works exactly this way: known suppliers by template, scans by OCR, the rest with a model's help.

• Templates: for a sender whose layout never changes, the system knows where each field sits. Fast and very reliable, but it breaks when the sender changes the layout.
• OCR, optical character recognition: turns a scan or a photo into text. It is the first step for any image, not the whole job, because text alone does not say which number is the total.
• AI model: reads the document the way a person would and labels the fields, even on a layout it has never seen. Flexible, but it can be confidently wrong, which is why the checks that follow matter.
• Rules after reading: whichever method read the document, fixed rules decide what to do with the result. That part should never be left to the model.

Checks that catch errors

In intelligent document processing, the checks after reading matter more than the reading itself, because they decide whether a field is trusted or sent to a person.

No reading method is perfect. A smudged digit, a total printed twice, a supplier who changed name: each can produce a field that looks right and is not. The defense is to compare every extracted field with something independent. Do the lines add up to the total? Is the supplier on your list? Has this invoice number been entered before? Does the delivered quantity match the order?

A document that passes every check can be drafted with confidence. One that fails any check is not guessed at; it goes to a person with the document attached and a plain note of what did not agree. That rule, that anything the system cannot handle exactly goes to a person, is what separates document processing automation you can rely on from automation you have to watch.

Where the extracted data goes

Extracted data should land where your team already works, usually as a draft: a bill in the accounting system, a row in a register or a filed PDF.

A note on the related term AI document automation. It is sometimes used for the opposite direction: producing documents, such as invoices, purchase orders and contracts, from data you already hold. The two belong together. A business that reads its incoming paperwork cleanly usually wants its outgoing paperwork built from the same lists, and our invoicing and purchase order robots, both running today for a client, do that side.

Reading a document is only useful if the result goes somewhere. In our set-up, it goes into the client's own accounts: drafts in the accounting system, rows in the client's own sheets, PDFs filed in the client's own folders. Nothing is moved into a separate system the team would have to learn. Here is the route one supplier invoice takes from mailbox to draft:

1. The invoice arrives in the mailbox or a shared folder.
2. The system decides what kind of document it is: invoice, credit note, statement or something else.
3. It reads the fields, by template, OCR or model, depending on the document.
4. It checks each field against what is already known.
5. If every check passes, it creates a draft bill with the PDF attached and files the PDF under the supplier.
6. If any check fails, it sends the document to a person with a note of what was unclear.
7. It writes every action to the audit record, and a person approves the draft.

Handling scans and poor quality

Poor scans and phone photos can usually be read, but they should be checked more carefully and handed back when they cannot be read cleanly.

Real paperwork is rarely tidy. Delivery notes are photographed on a loading bay, receipts are crumpled, faxes are still sent, and some files hold several documents at once. OCR copes with much of this, and an AI model can often make sense of what OCR returns. But the lower the quality, the more the checks matter, and the more often a document should come back to a person rather than be drafted.

Two habits help. Ask the people who send you documents to send PDFs rather than photos where they can. And when a document is unreadable, have the system say so plainly instead of returning a best guess. A clean handback costs a person a minute; a wrong figure entered quietly can cost much more when it is found at the end of the quarter.

Starting with one document type

Start intelligent document processing with the one document type that arrives most often and can be checked against something you already hold.

For most businesses that is the supplier invoice: it arrives constantly, the fields are the same every time and there is always a supplier list and a total to check against. Delivery notes, receipts and client forms follow the same pattern once the first type is working. The self-check on this page helps you test a document type before you start.

Be clear, too, about what intelligent document processing does not do. It does not decide how a cost should be treated, approve a bill, pay a supplier or reply to the sender. It reads, checks and drafts. Every decision about the result stays with a person, and every action is written to an audit record that shows what was read and what was done with it.

As for what runs today: supplier invoices and credit notes are read and drafted for a client, and past invoices are read into a price history of every product line. Expense receipts, delivery notes linked to a shipment register, and checking client or supplier documents for what is missing or expired are offered and built to order. All of it runs on our servers against your own accounts, rented monthly, for one monthly figure per workflow stated in the quote. A person approves every result, and a kill switch stops any robot in one command.

Questions