AI invoice processing in practice: how it works in production

Handling incoming supplier invoices is the work that somebody does in every company and nobody enjoys anywhere: save it from the e-mail, retype it, check it, forward it to the bookkeeper. This article shows what the same process looks like with artificial intelligence - and not in theory: the solution described here runs in production on the German waste-management platform entsorgo, across a nationwide partner network.

Why traditional OCR is not enough

Automating invoice processing is not a new idea; OCR-based systems have existed for a long time. Their practical problem is that they are template-bound. While the invoice is a tidy, typed PDF they do well, but a photograph taken with a phone, a creased page, a handwritten note in the margin or an unfamiliar layout stops them, and every new supplier format has to be trained in. Large language models are the breakthrough here: they do not look for a template, they UNDERSTAND the document, so they extract the supplier, the line items, the totals and the due dates from a format they have never seen before.

How a production system is built

The AI on its own is not much; reliability comes from the process built around it. Our chain looks like this:

• Intake: invoices arrive by e-mail or upload, every document gets an identifier, and nothing can be lost.
• Extraction: the model turns the invoice into structured data against a fixed schema: supplier, invoice number, dates, line items, net, VAT, gross.
• Rule-based validation: we do not take the model's output on faith. Separate, conventional code checks that the line items add up to the total, that the VAT arithmetic holds, and that the tax number is well formed.
• Matching: the system attaches the invoice to its order or contract and compares it with the expected amount.
• Discrepancy handling: anything that does not agree - a different total, an unknown item, a suspected duplicate - goes into a human approval queue. Everything else passes on its own.
• Handover to bookkeeping: approved items go on in a bookkeeping-ready form, by export or over an API.

The key phrase is human-in-the-loop. The person does not disappear from the process; they move to where a decision genuinely has to be made. The machine takes the routine, the human takes the exceptions.

What we learned in production

At entsorgo the system processes supplier invoices from the partner network. It replaced most of the manual data entry, and the transcription errors that came with it went too; the administrative team's time moved to exceptions and to customers.

What to check before you adopt it

• Hallucination risk: a model can be wrong, so rule-based post-validation is mandatory, as is a confidence threshold below which an invoice goes to a human automatically, and the ability to open the original document next to every extracted field.
• Data protection: an invoice carries business and personal data. Establish where the model runs, what data-processing agreement sits behind it, and whether the provider trains on your data. The right answer is no.
• Integration: the system is worth exactly as much as it fits your existing accounting process. The goal is not another island; it is that your bookkeeper receives finished, validated data.
• Regulation: on the outgoing side, Hungary's NAV Online Invoice reporting is already strict (our developer guide covers it), and EU e-invoicing is moving towards structured formats. If you are designing a process now, design it for machine processing from the start.

What it costs, and when it pays back

The honest answer is that it depends on volume and on the number of integrations. The good news is that you do not have to buy the whole system at once: we start every project with a pilot on a narrow set of invoices, against a measurable goal - processing time, error rate, manual hours. If the pilot's numbers do not justify continuing, we say so. Payback typically appears where invoice entry is a bottleneck at close, or where the volume is high enough that the manual work is a standing cost rather than an occasional one.

If you would like to see what this would mean for your invoices, ask for a consultation: bring 10 to 20 invoices of mixed quality and we will run them in front of you. Service details are on our AI invoice processing page.