Automatic reading of invoices, contracts and orders: how it works in practice

What the document reading function does

A document enters the system by one of several routes: an email attachment, a photo taken on a phone, a file dropped onto the screen, or a scanner. AI reads out the data you need: supplier or customer, number, dates, line items, amounts, order number, delivery address. The system fills in the right fields, and a person sees the original and the extracted data side by side.

All that is left is to check and approve. After approval the data goes on: to the accounting program, the order, the contract record.

Where it pays off

The effect shows where there are hundreds of documents a month, each waiting for someone to open it.

• Purchase invoices from suppliers who send PDFs or scans. Number, amounts, due dates, matching to a purchase order.
• Customer orders in PDF and email. Business customers often send an order from their own system as a file. AI turns the file into a ready order awaiting approval.
• Contracts. Parties, dates, amounts, notice periods, penalties. The data goes into the contract record and the system reminds you about deadlines.
• Field reports and job sheets. A photo of a handwritten report becomes data in the system before the technician is back at the office.
• Consignment and customs documents. Tracking numbers, weights, collection dates, matching to an order.

At a few dozen documents a month the function does not pay back. A person enters them faster than they can check the reading. I will say so plainly if that is your situation.

What the approval screen looks like

On the left the original document, on the right the extracted fields. A field AI is unsure about is highlighted, and that is where checking starts. Clicking a field highlights the place in the document the value came from.

Below the fields are the results of simple rules: do the line items add up to the total, does the supplier exist in the database, is the invoice number a repeat of a previous one, does the order number exist. The rules need no AI and catch mistakes AI could let through, for example a net amount entered as the amount due.

A person corrects what needs correcting and approves. For a document that looks like dozens before it, that takes a few seconds. The system records every correction, so I can see which fields cause trouble.

What happens with errors

By default AI writes nothing into accounting or into an order without approval. Three layers protect against a mistake: AI reports how confident it is about each field, rules check that numbers and database records agree, and a person approves.

How much control you need, we agree at rollout. If you want an exception to the default, we write it down together: for example, low-value documents from regular suppliers pass on rules alone, while everything above an agreed amount waits for a person. That is your decision, and the system enforces it.

Electronic invoicing and document reading

In Poland, for example, invoices from domestic companies now arrive through a mandatory national e-invoicing system as structured data, so there is nothing to read there and the system pulls them in directly. Similar schemes are spreading across Europe. Reading makes sense for everything those schemes do not cover: customer orders, contracts, reports, consignment documents and invoices from suppliers abroad.

If your accounting program cannot accept data from outside, the extracted documents can still be entered into it. How, is on the page about what to do when a program has no connection.

An example from my work

I worked on the systems of a British repair management platform for heavy industry and aerospace. The platform had automatic document recognition: repair quotes and claim documents from which it extracted data and let the customer find costs they had been overcharged. My work was the screens for running repairs. The scale and the requirements were larger than in a typical small or mid-sized company, but the principle is the same: a document goes in as an image and comes out as data to check.

How we start

We pick one type of document, the one you have most of. We count how many there are a month and how many minutes each takes. I build the reading as part of a module, meaning a self-contained part of the system the team already works in, for example the invoices or orders module. After a month we compare the numbers.

The bigger picture of where AI makes sense in a company is on the main page of the guide. The other pages of the guide cover answering customer questions and searching company knowledge.

Describe in the form below which documents come in, how many a month, and where the data should go. I will get back to you.