Invoices into Pohoda automatically: how it works and where the limits are

The flow has five steps, not one #

Well-designed processing does not go straight from the document to the accounting system. There are two checks in between:

• The document arrives: from email, a shared drive or a scanner, as a PDF or an image.
• Extraction: the system reads the supplier, number, dates, amounts and line items from the document.
• Checking doubtful values: anything unreadable or inconsistent is flagged.
• Approval by a person: the accountant sees the document next to the extracted data and confirms or corrects it.
• Writing to Pohoda: only the approved record is handed to the accounting system.

What the system should be able to read #

The basics are the details an accountant retypes every day: the supplier and their company ID, document number, issue and due dates, variable symbol, tax base, VAT and total amount.

Invoice line items are harder. They have a varying number of rows and different layouts, and every supplier presents them differently. That is exactly where the check matters most.

Doubtful fields are highlighted, not skipped #

A system that is sure of everything is suspect. A good system tells you which field it doesn’t trust: an illegible amount, a company ID that doesn’t exist, line items that don’t add up to the total, a supplier it hasn’t seen before.

The accountant doesn’t have to check everything from scratch. They look at the flagged fields, compare them with the document and decide. The decision always stays with a person.

Where the limits are #

• Poor quality: a blurry scan, a document photographed at an angle, a stamp over the text or handwritten notes in the margin.
• Non-standard documents: foreign invoices, advance invoices, credit notes or contract documents with an unusual layout.
• Accounting decisions: which account and which period a document belongs to is decided by company rules and the accountant, not by a model’s guess.
• A new supplier: the first documents from an unknown supplier need more corrections than those from regular ones.

How to check whether it will work in your company #

Don’t rely on sample documents. Take a batch of real documents from recent months, including the worse ones: scans, photos and exceptions.

For each field, count how often it was extracted correctly, how often the system flagged it as doubtful and how often it got it wrong without warning. The last number matters most. The measured values are your own and apply to your documents.

How we approach it #

We build processing on the documents and systems a company already uses and connect it to Pohoda, Omega, iDoklad or other accounting systems via API or import. Doubtful fields are highlighted for review, nothing is written without approval by a person, and the system learns from corrections. Before live use, we try it in a pilot on real documents.