An average Slovak company with €1M+ revenue processes 300–1,500 incoming invoices a month. Each means four minutes of re-typing, matching and filing. At a thousand invoices, that is 67 hours a month — a full working month of one person.
What an AI agent can do today
Extract the header and line items from a PDF including ugly scans, match the invoice to the order and delivery note, check VAT rates and limits, post the document into the accounting system and — only when something does not add up — call a human. Straight-through rate: 85–95 % of documents depending on input quality.
What it costs and returns
Deployment on top of existing systems (Pohoda, Money, SAP B1, custom ERP): from €14,900 including savings measurement. Inference costs: cents per document. At a thousand invoices a month, payback is 8–11 months — and the error rate drops below the best human's best day.
Why it's not just better OCR
Classic OCR reads text; an LLM agent understands context — it knows that “due” on an invoice means the payment date, recalculates VAT the supplier forgot, and matches a line item even under a different name than on the order. That understanding is what moves the success rate from 60 % (OCR) to 85–95 %. The difference is not scanning, it is judgement.
Control and audit: the accountant becomes an approver
The goal is not to replace the accountant but to move them from re-typing to reviewing exceptions. Every document carries a confidence score and anything below the threshold goes to a human for approval; everything is logged and auditable after the fact. In an hour the accountant now checks what used to take an hour to re-type — and makes decisions instead of doing mechanical work.
The best first step: two weeks of measuring the current state. If it comes out under 20 hours a month, we will talk you out of automating — it makes no sense. Above 40 hours, the decision is arithmetic, not faith.