OCR and automated invoice data capture after KSeF: where it still makes sense and how to implement it

OCR (optical character recognition) is a technology that turns an image of a document into text. It works like a very fast typist: it will copy everything it sees, but it does not know which number is the gross amount and which is the bank account number.

Intelligent document processing (known in the industry as IDP) goes a step further. The system recognises the type of document, finds specific fields on it, such as the tax ID, date and amounts, and saves them in a structured form. It then checks them against the company’s rules and passes them on.

In practice this difference is crucial. OCR alone lets you search an archive for an invoice by keyword. Only full processing lets you pass the data on for approval and posting without retyping. In the rest of this article we simply talk about automated document reading.

From 1 February 2026 every company in Poland receives invoices through KSeF, Poland’s National e-Invoicing System – the government platform for invoices between businesses. An invoice from a Polish supplier arrives there as a file with the data in separate fields: tax ID, amounts, VAT rates and line items. There is no need to read them from an image; you simply download them from KSeF into your accounting system.

Until 31 December 2026, some of the smallest suppliers may still issue invoices outside KSeF if their invoiced sales do not exceed PLN 10,000 a month including VAT. Such invoices will keep arriving by email until the end of the year. The documents listed in the table below, however, stay outside KSeF permanently.

If your company imports a lot, uses foreign cloud services or handles many warehouse documents, the stream of paperwork to read may still be considerable. If you buy almost exclusively in Poland, you may find that the problem of retyping invoices has largely disappeared.

Take a hypothetical distribution company that receives an invoice by email from a German supplier. The whole process takes ten to twenty seconds and consists of several steps that look similar in any tool.

The most important element is the confidence score. For each field, the model states how confident it is in the reading, on a scale from 0 to 1. The company sets a threshold: above it, the data moves on automatically; below it, the document goes to a person. In Microsoft’s examples the threshold is 0.65, but the right value is set by testing on your own documents.

Even a good model sometimes gets a digit in an amount wrong or mistakes the purchase order number for the invoice number. If nothing catches this, the error reaches the accounts and only comes to light when balances are reconciled. That is why most of the implementation work concerns the rules that check the extracted data.

Some checks are required by law. For a transaction above PLN 15,000 with a Polish taxpayer, paying into an account that is not on the Polish VAT white list means losing the tax-deductible cost. The exception is notifying the tax office within 7 days of ordering the transfer. So it is worth building the account check into the workflow before the invoice goes for payment.

The following checks usually catch most reading errors and fraud attempts:

Tool vendors like to quote the percentage of documents handled without human involvement. Such figures are of little value, because they depend on scan quality, the number of suppliers and how strict the rules are. It is better to calculate the effect on your own documents.

Say a company receives 150 foreign invoices and receipts and 250 delivery documents a month, 400 documents in total. By our estimate, entering one by hand takes 4 minutes, which is 1,600 minutes, or about 27 hours a month.

After automation, someone still reviews each document for a minute, and corrects one in five for another 3 minutes. That is 400 minutes plus 240 minutes, 640 minutes in total, or about 11 hours. So the company saves about 16 hours a month, which at PLN 80 an hour comes to about PLN 1,280 a month and PLN 15,360 a year.

Plug in your own document numbers, times and rate. If the result is small compared with the cost of licences and setting up the workflow, automated reading is not the first step worth taking.

In a company that works on Microsoft 365, the simplest route is AI Builder, a set of prebuilt AI models available in Power Automate. The invoice model supports Polish and several dozen other languages, and if it struggles with an unusual document, you can train it further on your own examples. From 1 November 2026, new Premium licences no longer include AI Builder credits. Reading is therefore charged in Copilot credits, and the cost depends on the number of pages processed.

For large volumes or unusual documents, developers turn to Azure Document Intelligence, the Microsoft service on which AI Builder’s reading is based. It gives more control but requires your own integration. On the other hand, many accounting and ERP systems already have built-in PDF invoice reading, which is often enough when document numbers are small.

Microsoft Copilot and other AI assistants are good at summarising a single document or answering questions about its content. They cannot, however, replace repeatable reading of hundreds of invoices with checks and a recorded history. You will find a comparison of the most common options in the table below.

The first mistake is automating without checking how many non-KSeF documents are actually left. It happens that a company buys an invoice-reading tool, and a few months later it turns out that it downloads its domestic invoices from KSeF anyway. Always start by counting documents by type.

The second mistake is having no path for exceptions. A document with an uncertain reading must go to a specific person, with a deadline and a preview of the file next to the extracted data. Without this, doubtful documents get stuck and people go back to entering data by hand. Other typical mistakes:

It is best to start with one type of document that arrives in large numbers and has a consistent layout. In trading companies this is usually foreign invoices; in manufacturing and logistics companies, delivery documents.

The following steps let you check whether automation makes sense before you spend money on a full rollout: