Intragentic
All guides

Check incoming invoices instead of typing them up

How AI reads incoming invoices, matches them against order and delivery note and presents them to you for approval. Process, costs, metrics.

Published
Reading time
5 minutes

If you want to automate incoming invoices, start with how they are handled today. Take a made-up joinery workshop in Schwabach with one part-time office administrator. Invoices arrive there as PDFs by email, on paper with the delivery, through supplier portals and occasionally already as XRechnung. The administrator types supplier, amount, due date and job number into a spreadsheet and looks for the order and delivery note in the binder. In the evening the owner signs off the pile, and once a month the documents go to the tax adviser.

A survey by the German Confederation of Skilled Crafts (ZDH) shows how big the pile can get. In 2025 the businesses surveyed received around 1,200 incoming invoices on average, only about half of them as e-invoices (E-Rechnung). According to the ZDH, most invoices are processed by hand. Applied to our made-up joinery, that comes to around 100 a month.

As of 21 October 2026, suppliers with more than 800,000 euros turnover in the previous year must send e-invoices from 1 January 2027. From 2028, in principle all suppliers must. Small amounts up to 250 euros gross and invoices from small traders under the VAT exemption (Kleinunternehmer) may still arrive as PDF or on paper. So the inbox stays mixed. A delay to these deadlines has been called for, but not decided.

The process with AI

  1. Inbox. All invoices land in one mailbox, paper is scanned, and portal invoices are added.
  2. Read and match. The AI reads PDF, scan and XML, which is the data record of an e-invoice. It checks that the mandatory details under Section 14(4) of the German VAT Act (UStG) are complete. It looks for the order and delivery note and flags differences in quantity, price or bank details.
  3. Approval. This is where the AI’s work ends: it suggests, the owner decides. In an approval list she sees every invoice above her threshold and every invoice with a difference. The business sets the threshold itself. She approves or sorts out the difference with the supplier. Before anything goes to accounting or the tax adviser, she confirms the complete list.
  4. Output. Only then comes the export in the format your tax adviser accepts, for example for DATEV. Every invoice is archived with a log of the checks, and the data record stays unchanged.

What the AI does not decide: it pays no invoices, and transfers are still made by hand. It makes no judgement on input tax deduction, account coding or booking; that stays with the tax adviser. It does not write to suppliers. Approval by a person is our rule, not a legal requirement.

What it takes

It takes your invoices in every form, plus orders and delivery notes from your stock management system or a spreadsheet, and your supplier list. We connect the systems you already use: your email inbox, DATEV Unternehmen online or an export, Microsoft 365 or your industry software. If it is unclear where orders are kept or where the booking data goes, we start with the sprint. Otherwise we go straight to the prototype.

Invoices must be kept for eight years (Section 14b UStG). For e-invoices, the data record as received is usually enough. If the readable image carries extra details such as booking notes, it is kept as well. This is set out in the GoBD, the Federal Ministry of Finance’s bookkeeping rules (margin no. 131). You settle with your tax adviser how your business implements and documents this; we supply the groundwork.

This is business data, and personal data only where a supplier is a sole trader. Your instance and the models run on servers in Germany. We set out the processing in a data processing agreement under Art. 28 GDPR. We do not train AI models on your invoices.

What it costs

All prices are net. In the AI sprint for €4,900 we spend five working days clarifying systems, purchasing and thresholds. The prototype, from €14,900, runs for four weeks on your real invoices and covers reading, matching and the approval list. For everyday use, the production build follows from €29,000, with accounting export and supplier master data. We credit the full price of the sprint against it.

Operations cost from €490 a month. We measure what each invoice costs in model use and computing time during the prototype, and report it every month once the system is running. We give no figure in advance.

Whether the Digitalbonus Bayern, a Bavarian grant for small businesses, can help is open. It does not fund measures required by law (implementation notes, Vollzugshinweise No. 1), and whether that applies here is unsettled. If you want to try, the rule is apply first, order second, and your business submits the application itself through its ELSTER business account. This is not legal, tax or funding advice.

How you can tell it pays off

We measure four values: minutes per invoice, the share needing no correction, days from receipt to approval, and early payment discounts missed. We take the baseline in the sprint or at the start of the prototype. The comparison follows later in the prototype, with the same kinds of invoice. We do not promise savings in advance; you see them in these values, or you do not.

If incoming invoices travel through your office by hand, give us a short description of their route and of who approves them today. In a first call we work out where to start.

Sources

All links accessed on 21 October 2026.

Let’s talk about your process.

A first call takes 30 minutes and costs nothing. Afterwards you know which step fits.

Request a first call

Assistant

Fixed answers, no AI