AI agents in administration: invoices that process themselves

A company with turnover above a million euros processes hundreds of incoming invoices a month. If each one means four minutes of re-typing, matching and filing, a thousand invoices add up to 67 hours a month — almost half of one person’s working time.

What an AI agent can do today

It extracts the header and line items from a PDF, including poor-quality scans, matches the invoice with the purchase order and delivery note, checks VAT rates and limits, files the document in the accounting system and — only when something doesn’t match — calls a person. It’s a typical task for AI automation.

What share of documents gets through without human intervention depends mainly on the quality of the inputs. With clean electronic invoices it’s the majority; with poor scans, fewer. That’s why it should be measured on your own documents, not promised in advance.

Why it’s more than better text recognition

Classic text recognition (OCR) reads letters. An AI agent understands context: it knows that “due” on an invoice means the due date, recalculates VAT the supplier forgot to state, and matches a line item even when it’s named differently on the purchase order. That understanding is the difference between a tool you constantly have to correct and a tool that saves time.

What it costs and when it pays back

Connecting to existing systems such as Pohoda, Money S3, SAP Business One or your own ERP is a routine task and doesn’t require replacing them. Running the AI model costs in the order of cents per document. The payback is easy to calculate: hours saved times the hourly cost, minus running costs.

The accountant becomes an approver

The goal isn’t to replace the accountant but to move them from re-typing to checking exceptions. Every document has a confidence level, and everything below a set threshold goes to a person for approval. Every step is logged, so it can be reviewed at any time. In an hour, the accountant checks what used to take an hour to re-type, and makes decisions instead of doing mechanical work.

The best first step: measure the current state for two weeks. If it comes out at less than 20 hours a month, automation probably isn’t worth it. Above 40 hours, the decision is arithmetic, not faith.