--- title: "AI invoice processing in practice: how it works in production" description: "AI invoice processing in practice: LLM-based extraction, order matching, discrepancy handling and bookkeeping handover. Lessons from a live German system." url: https://milcomp.hu/en/news/ai-invoice-processing/ locale: en published: 2026-07-26 modified: 2026-08-09 author: Márton Molnár --- AI automation # AI invoice processing in practice: how it works in production How does AI process incoming invoices, from the e-mail to the bookkeeping? Lessons from a system of ours running in production in Germany: architecture, limits, payback. 26 July 2026 Updated 9 August 2026 4 min read Written by Márton Molnár Founder and developer at Milcomp. Handling incoming invoices is the job somebody does at every company and nobody enjoys anywhere: saving attachments from e-mail, retyping the data, checking it, forwarding it to the accountant. In this article I show what the same process looks like with artificial intelligence, and not in theory: I built the system described here, and it runs in production on the [German entsorgo waste-management platform](https://milcomp.hu/en/projects/entsorgo/), processing the invoices of a nationwide partner network. ## Why traditional OCR is not enough Automating invoice processing is not a new idea: OCR (text recognition) systems have existed for decades. Their practical weakness is that they depend on templates. As long as the invoice is a clean, typed PDF they do fine, but a photographed, crumpled document with a handwritten note or an unusual layout stops them, and every new supplier format needs new training. Large language models (LLMs) are the breakthrough here: they do not match templates, they READ the document, so they extract the supplier, the line items, the amounts and the due dates even from a format they have never seen before. The most common misconception OCR does not *understand* an invoice, it reads characters. Rule-based extraction breaks on the first new supplier layout, and the error usually surfaces in the bookkeeping. LLM-based extraction is not a better OCR - it is a different way of solving the problem. ## How a production system is built The AI alone is not enough; reliability comes from the process built around it. This is how I built the chain: - **Intake:** invoices arrive by e-mail or upload, every document gets an identifier, and nothing can be lost. - **Extraction:** the LLM turns the invoice into structured data against a fixed schema: supplier, invoice number, dates, line items, net, VAT, gross. - **Rule-based validation:** the AI's output is never trusted blindly. Separate, conventional code checks that the line items add up to the total, that the VAT arithmetic is right, and that the tax number has a valid format. - **Matching:** the system links the invoice to its order or contract and compares it with the expected amount. - **Discrepancy handling:** anything that does not match (a different amount, an unknown item, a suspected duplicate) goes into a human approval queue; everything else flows through on its own. - **Handover to bookkeeping:** approved items leave the system ready to book, via export or API. The key phrase is **human-in-the-loop**: the human does not disappear from the process, they move to where a real decision is needed. The machine carries the routine, the human handles the exceptions. - **Definition Human-in-the-loop**: A process where the machine decides but a person approves before the decision has financial consequences. In invoice processing that is the difference between a demo and a production system: the question is not whether the AI gets it right, but what happens when it does not. ## What we learned in production At entsorgo the system processes the supplier partner network's invoices. It replaced most of the manual data entry, and the retyping errors disappeared with it; the administrators' time moved to the exceptions and the customers. As a rule of thumb: the more invoices flow through a company and the more varied their formats, the bigger the saving. With only a handful of invoices a month it is not worth it: the rollout costs more than it will ever return. ## What to check before you adopt it - **Hallucination risk:** an LLM can be wrong, so rule-based post-validation is mandatory, along with a confidence threshold (below which the invoice goes to a human automatically), and a link from every extracted field back to the original document. - **Data protection:** an invoice carries business and personal data. Clarify where the model runs, what data-processing agreement stands behind it, and whether the provider trains on your data (the right answer: no). - **Integration:** the system is worth exactly as much as it fits your existing bookkeeping flow. The goal is not another island, it is that your accountant receives checked, ready-to-book data. - **Regulation:** EU invoicing is moving toward structured formats under the [ViDA package](https://taxation-customs.ec.europa.eu/taxation/vat/vat-digital-age-vida_en), with mandatory structured e-invoicing for cross-border B2B from 1 July 2030. Anyone building a process now should design it for machine processing from day one. ## What it costs, and when it pays back The honest answer: it depends on your volume and the number of integrations. The good news is that you do not have to buy the whole system at once: I start every project with a **pilot** on a narrow set of invoices, against a measurable target (processing time, error rate, manual hours). If the pilot's numbers do not justify continuing, I say so. The payback typically appears where several hundred invoices flow through the company each month, or where invoice entry is the bottleneck at closing time. If you would like to see what this means on your own invoices, [book a consultation](https://milcomp.hu/en/contact/): bring 10-20 invoices of mixed quality and I will show you live. :) The service details are on the [AI invoice processing page](https://milcomp.hu/en/ai-invoice-processing/). The short version - Classic OCR learns a template; a language model reads meaning - which is why photographed and never-before-seen invoices go through. - The real work is not extraction, it is validation: matching to the order, handling discrepancies, handing over bookkeeping-ready data. - Nothing passes without human approval except a high-confidence match. Share - [LinkedIn](https://www.linkedin.com/sharing/share-offsite/?url=https%3A%2F%2Fmilcomp.hu%2Fen%2Fnews%2Fai-invoice-processing%2F) - [Facebook](https://www.facebook.com/sharer/sharer.php?u=https%3A%2F%2Fmilcomp.hu%2Fen%2Fnews%2Fai-invoice-processing%2F) ## Related articles - [GEO: how your company gets into ChatGPT and Google AI answers](https://milcomp.hu/en/news/geo-guide/) 3 August 2026 Your customers are asking AI before they choose a supplier. GEO (generative engine optimization) is about being the source in that answer. A practical guide, without the hype. Let's talk about your processes We map where automation actually saves money in your company, and where it does not. Request an AI assessment