Document and data extraction: how AI reads invoices, receipts and statements, and how accurate it is 

Document extraction reads a PDF, photo or email, pulls out the fields you need, such as supplier, date, amount and tax, and suggests where to post them in your accounting software. Clean, typed documents extract well. Handwriting, poor photos and unusual layouts need a person to check. Every result is reviewed before posting.

What is document extraction?

Document extraction is the use of software to read a document and turn it into structured fields a computer can work with.

Without it, someone opens each supplier invoice, reads the supplier name, the date, the amounts and the tax, and types them into the accounting system. With it, the software reads the document first and shows a person the result to confirm. Typing turns into checking.

The same approach works for receipts, bank statements and forms. The fields change, but the idea stays the same: read the document, find the values, suggest where they go, and let a person approve.

How does it work from PDF, photo or email?

The document reaches the system by whichever route suits your team. The system then reads it and prepares a draft entry.

1. Receive: the document arrives by email, by upload or as a photo from a phone.
2. Read: the system turns the page into text. For a photo, this includes recognising the characters.
3. Understand: AI works out which text is the supplier, the date, the amounts and the tax, using the layout and the wording.
4. Suggest: the system proposes where to post the entry, such as an expense account, based on your past entries and your rules.
5. Review: a person checks the draft against the document, corrects any errors and approves it.
6. Post: the approved entry goes to your accounting software.

Step two is often called OCR, and it produces only text. Step three is where AI earns its place, because it can handle many layouts without a separate rule written for each one.

How accurate is it, and what does "accurate" mean here?

Here, accurate means the right value in the right field. A system that reads every word on the page but puts the invoice date in the due-date field is not accurate. One wrong digit in an amount can matter more than ten correct words.

Accuracy varies from one set of documents to the next. It depends on scan quality, layout, language and how consistent your suppliers are. A vendor's published accuracy figure describes their test documents, not yours.

The honest way to find the accuracy for your business is to test it. Take a sample of your own invoices, run them through and count the fields that came out wrong. Do this before you rely on the system, and repeat it from time to time.

Why is a review step not optional?

AI can be wrong without showing any sign of it. It may read a smudged 3 as an 8 and present the result with full confidence. In accounting, an error nobody notices turns into a wrong balance, a missed payment or a tax problem later on.

Review also gives the system something to learn from. When a person corrects a field, that correction can improve later suggestions, and over time fewer items need fixing.

• Check the supplier, date, total and tax against the document.
• Check that the invoice is not already in your books.
• Take extra care with unusual amounts and new suppliers.
• Keep the original document linked to the entry, so anyone can verify it later.

Even with review, the work is much faster than typing. The saving does not come from removing the person. It comes from changing that person's job from typing to checking.

What does it post into your accounting software?

It produces a draft entry with the supplier, date, amounts, tax and a suggested account. Whether that entry can go straight into your software depends on what the software allows. Tally, Xero, QuickBooks and MYOB each have their own ways of bringing in data, and the options vary by product, version and region.

Where a direct connection exists, the approved entry can be sent across. Where it does not, the system can create a file in a format your software can load. Either way, the entry waits for a person, such as your accountant, to approve it before it touches your books.

Bring your accountant in early. Their preferences on coding, tax treatment and approval shape the whole setup, and they are the one who has to trust the result.

Which documents are easy and which are hard?

Start with the easy documents. Prove the method, build trust, then add harder types one at a time. Where you can, ask suppliers to send typed PDFs. It is a small request, and it makes the whole process more reliable.

How do you set up the review loop?

The review loop is the routine that keeps extraction honest. It has four parts, and none of them is complicated.

• A queue: every extracted document waits in one list for a person to approve.
• Flags: the system marks uncertain fields, such as a blurred amount or a missing tax figure, so the reviewer looks there first.
• Corrections: when the reviewer fixes a field, the fix is saved and used to improve later suggestions.
• A sample check: even once trust has grown, a person still checks a regular sample of approved items.

Name one person as owner of the loop, and write the name down. Agree how quickly documents are reviewed, so nothing sits for weeks. Look over the corrections each month. If the same mistake keeps coming back, the fix belongs in the rules, not with the reviewer.

How do you test it on your own documents?

Pick a sample from the last few months with a mix of easy and hard documents. Include your most common suppliers and a few awkward ones. Run them through and compare every field with the original.

Count the wrong fields by type. You may find that dates are fine but tax amounts often go wrong, or that one supplier's layout fails every time. That tells you where to add rules and where a person needs to look closely. Repeat the test after any change, and again whenever a new supplier appears.

Keep a record of the results. They are the only accuracy figures that apply to your business, and they show whether a change actually helped. Share them with your accountant, so the review step is built around the mistakes that really happen.

Frequently asked

Owners usually ask how accurate extraction is, how it differs from OCR and whether it works with their accounting software. The answers follow. The short version: test with your own documents, and keep a person checking every result.