Grow · Intermediate

AI document processing: invoices, orders and delivery notes without retyping

Every invoice, order and delivery note that arrives as a PDF or on paper is typed into a system by someone. Software can now read most of those fields. What it cannot do is decide by itself when it has read them wrongly.

6 min read

Sheets of coloured paper stacked in the trays of a shelf.

Photo: Hendrik Luehrsen, public domain (CC0)

In this guide

Short answer

AI document processing reads incoming documents, pulls out the fields you need (supplier, number, date, lines, amounts), checks them against your own data and passes them to the accounting or ERP system. People no longer type every document. They review only the fields the software is unsure of, and that review step stays.

5steps from the incoming document to the booked record
4kinds of document that suit it: invoices, orders, delivery notes, forms
3things accuracy depends on: scan quality, layout variety, handwriting

What is AI document processing?

It is software that turns a document into data and checks the result before anyone books it. It is also sold as intelligent document processing. Older systems needed a template for every supplier’s layout: the invoice number is here, the total is there. Current systems read a document more as a person does, recognising a total or a delivery address by its meaning and position, so a new layout does not need a new template.

Reading is only half of it. The useful part is what happens next: the extracted fields are compared with what your systems already know, and only records that pass go through untouched.

One kind of document needs no reading at all. A structured electronic invoice already carries its data as fields. If a supplier can send those, take them: nothing is cheaper or more reliable than data that never had to be read from a page.

Which documents suit it?

Documents suit it when they arrive in volume, carry the same kinds of field every time and end up in a system. The more of the three apply, the better.

Documents and how well they suit automatic reading
Document Typical fields Suitability
Supplier invoices Supplier, invoice number, date, net, tax, total, order reference Very good: fixed fields, sums that can be checked
Purchase orders from customers Customer, article numbers, quantities, delivery date Good, if article numbers can be matched to your own
Delivery notes Order reference, articles, quantities delivered Good, but stamps and handwritten notes need review
Forms and applications Named fields, tick boxes, signatures Good for printed text, weaker for handwriting
Contracts and free-form letters Parties, dates, terms in running text Limited: useful for finding and summarising, not for booking

Orders show where the effort lies. Reading “50 pcs gasket 40 mm” is easy. Knowing which of your article numbers the customer means is the hard part, and it depends on your master data, not on the reading software.

How does it work? The five steps

Every working set-up has the same five steps, and the third and fourth are the ones that make it trustworthy.

  1. CaptureDocuments come in from an email inbox, a scanner or an upload. Each is stored once, with its source and the time of arrival.
  2. ExtractionThe software reads the document and returns fields: header data such as supplier and number, and the lines with article, quantity and price.
  3. ValidationRules compare the fields with your data. Is the supplier known? Does the order number exist? Do net and tax add up to the total? Has this invoice number arrived before?
  4. Human reviewDocuments that fail a rule, and fields the software marks as uncertain, go to a person who sees the original and the fields side by side and corrects them.
  5. ExportChecked records go to the accounting or ERP system through its interface, with the original document attached.

The connection in the last step is ordinary integration work, and often the larger part of the project. It is described under API integrations.

Why does a review step stay?

It stays because a misread amount that is booked and paid costs more than the minutes of review ever save. No reading software is right every time, and a wrong digit looks exactly as confident on the screen as a right one.

The aim is therefore not to remove people. It is to send them only the documents that need them. Validation rules do most of that sorting: a document whose sums add up, whose supplier and order are known and whose quantities match the order has passed several independent checks. A document that fails one is shown to a person with the reason.

What does accuracy depend on?

It depends on the documents far more than on the software, which is why a figure from a brochure tells you little about your own inbox.

  • Scan quality. A PDF created by the sender’s software contains real text. A skewed photo of a crumpled delivery note does not.
  • Layout variety. Twenty regular suppliers are easier than five hundred occasional ones.
  • Handwriting and stamps. Printed text reads well. Handwritten quantities and stamps over printed text read worse.
  • Line items. Header fields are easier than long tables that run over several pages.
  • Your master data. If the same supplier exists three times in the system, no software can pick the right one.

Test with your own documents before you decide: a few hundred real ones from the past months, the difficult ones included.

How we do it

We begin with one kind of document and a sample from your own archive. We set up extraction, write the validation rules together with the person who books the documents today, build the review screen and connect the export. Everything is reviewed at first, and the rules are tightened from what the corrections show.

How we helpAI data processingWe set up the reading, checking and hand-over of invoices, orders and other documents into your accounting or ERP system, with a review step for uncertain fields.See how we process documents with AI

How much time does it save? An illustrative example

It saves most of the typing time and adds a little review time. The figures below are an illustrative example with invented numbers, to show how to do the sum for yourself.

A company receives 500 supplier invoices a month. Typing one in by hand takes 4 minutes: 2,000 minutes, or 33 hours 20 minutes a month. With document processing, assume 350 invoices pass every check and get a glance of 0.2 minutes each (70 minutes), while 150 are flagged and take 1 minute each to correct (150 minutes). Together that is 220 minutes, or 3 hours 40 minutes.

Illustrative example: minutes a month for 500 invoices

The difference in this example is 1,780 minutes, close to 30 hours a month. Your own share of flagged documents will differ, and it is the number to measure in a trial. The same pattern of rules and hand-overs carries over to other routine tasks, described under workflow automation.

  • One kind of document is chosen, with its volume a month counted.
  • The fields that must be extracted are listed, lines included.
  • A sample of real documents is collected, the bad scans too.
  • Supplier, customer and article master data have been cleaned of duplicates.
  • The validation rules are written down in plain sentences.
  • It is decided who reviews flagged documents and how quickly.
  • The receiving system has an interface for importing records.

Questions we hear often

No. OCR turns a picture of a page into text. Document processing goes further: it works out which piece of text is the invoice number or the total, checks the fields and hands them to your system.

Partly. Clear printed text reads well, handwriting noticeably worse, so handwritten fields should always go to review.

It does if the system can import records through an interface or a file format. Check this first, because the export is often the larger part of the work.

It depends on your documents: scan quality, the variety of layouts and the state of your master data. Measure it in a trial with your own invoices instead of relying on a published figure.

Yes. The work shifts from typing to reviewing flagged documents and to the cases that need judgement.

Related guides

Want this sorted without doing it yourself?

Tell us what you are working on. A person replies within one working day.

Talk to us

All guides