Reading invoices with AI: what works today and where people stay in the loop

Entering invoices by hand is the least valuable hour in any finance team's week. It is also where many errors begin: a transposed digit in an invoice number, a VAT amount typed into the net field, a purchase recorded in the wrong month. AI document reading has become good enough to remove most of that typing. It has not become good enough to remove the accountant, and the businesses getting the most from it are the ones that are clear about the difference.

What AI does well today

Current models read invoices the way a person does, by understanding the layout rather than relying on a fixed template for each supplier. On a typical purchase invoice they reliably pick out:

• the supplier's name and tax identification details;
• the invoice number, the invoice date and the due date;
• the net amount, the VAT amount, the VAT rate and the gross total;
• the currency and the payment terms;
• a short description of what was bought.

They do this from scanned PDFs, photographed receipts and invoices in Albanian, English, German and other languages, and they do it in seconds.

Checks that make the output trustworthy

Reading is only half the job. Every extracted invoice should pass simple, mechanical checks before anyone relies on it:

• Arithmetic. Net plus VAT equals the gross total, and the VAT amount matches the stated rate.
• Known rates. In Kosovo the rate on a domestic invoice should be 18%, 8% or zero. Anything else is flagged.
• Duplicates. The same supplier and invoice number seen twice is stopped before it is booked twice.
• Period. The invoice date falls in the month being closed, or the difference is explained.
• Confidence. Each field carries a confidence score, and low-confidence fields are highlighted instead of guessed.

These checks are not intelligent, which is exactly why they are useful: they catch the errors a model makes with confidence.

Where people stay in the loop

The judgement calls stay with a person: whether input VAT can be deducted, whether an item is an expense or an asset, whether a foreign supplier's invoice needs a different treatment, and whether something simply looks wrong. A model can point at these questions. It should not answer them on its own, and it should never post to the books without a person's approval.

What changes for the finance team

When reading is automated, the accountant's month changes shape. Less time goes into typing and more into review, reconciliation and the conversations that matter: why a cost line moved, which supplier invoices inconsistently, what the VAT position will be on the 20th. The close gets faster because documents are processed as they arrive rather than in a batch at month end.

The goal is not an accountant who checks the AI. It is an AI that does the typing so the accountant can do the accounting.

How we run it at Ura

Clients upload invoices and receipts through the Ura client portal. Each document is read automatically, checked against the rules above and marked as ready or as needing review. An accountant reviews every flagged field, confirms the VAT treatment and approves the entry before it reaches the books. Documents are stored privately and opened only by the people who work on the client's account.

Getting started

If you want to try AI invoice reading in your own team, start small: one document type, one month and a person reviewing every result. Measure how often the reviewer has to correct something, fix the causes, and widen the scope only when the error rate is one you would accept from a new colleague.