AI Document Processing

An invoice arrives as a PDF by email, someone reads off the supplier, amount and order number and types them into the ERP. Delivery notes, orders and contracts go much the same way. That costs time, invites errors and ties people to work that is easy to automate. Classic text recognition alone is often not enough, because every document is laid out differently.

We build solutions that recognize, classify and extract documents, check the values and hand them over in structured form to the ERP and document storage. Clear cases run through automatically; unclear ones go to the responsible staff for review.

From incoming mail to the ERP

In English this is called intelligent document processing, and it means a chain in which every step has a clear job. Documents arrive by email, scan, upload or from a portal. Text recognition (OCR) reads text from scans and images, and layout recognition finds paragraphs, tables, headers and footers. For digitally created PDFs the text is already there. Then the system works out whether it is looking at an invoice, a delivery note, an order or a contract, and splits bulk scans into individual documents. Next comes extraction of the fields needed, such as invoice number, amounts, due dates, contracting parties and line items. Rules and cross-checks assess whether the values are plausible, and only then does the checked data go to the ERP, accounting or document management.

What e-invoicing changes and what it does not

In Germany the e-invoicing mandate applies to transactions between domestic businesses. An e-invoice has a structured format following the European standard EN 16931, in practice XRechnung, a pure XML file, or ZUGFeRD, a PDF with embedded XML data. A plain PDF or a scan does not count.

Since January 1, 2025, companies have had to be able to receive e-invoices. For issuing, a phased plan applies. Until the end of 2026, paper invoices remain permitted, as do PDFs with the recipient’s consent. From January 1, 2027, companies with more than 800,000 euros total turnover in the previous year must issue e-invoices, and for everyone else it applies from January 1, 2028. There are exceptions, among others for small-amount invoices up to 250 euros and for tickets; invoices to private individuals are not affected. Tax details such as retention belong with your tax advisor, and we do not offer tax advice.

For processing, this means that structured e-invoices can be read directly from the data and need no AI recognition. Checking, account assignment and handover remain. AI recognition is still needed for PDF invoices arriving during the transition period and from suppliers that have not switched, for invoices from abroad and for small amounts. Add all documents with no e-invoicing mandate, that is delivery notes, orders, order confirmations, contracts, forms and correspondence. A good solution therefore reads both, e-invoices directly and the rest through recognition. As the share of structured invoices grows, the emphasis shifts to the other documents.

Extracting with templates, document models or language models

There are three common approaches to extraction, and they can be combined. With OCR plus rules and templates, the position of a field is fixed. That works for forms that always look the same and fails on new layouts. Pretrained document models, which large cloud providers offer as a service, know invoices and receipts, return fields with confidence scores and usually cope well with unknown layouts. Language and vision models read the page as text and image and fill a predefined schema. They are flexible with unusual documents and contracts but can invent values and cost more per page.

In practice one service typically delivers OCR and layout, a model extracts the fields, and rules check the result. Structured output is what matters. We define a schema with required fields, data types and line items that the model has to stick to, and we store where on the page each value was found. The reviewer sees immediately where a number came from.

Which service works depends on document types and languages, on quality on your own documents, cost per page and the cloud you already use. Open tools that run locally are an option too. Since documents often contain personal data, we clarify with your data protection officer up front whether processing in the EU is enough or whether it has to run in-house. More on this is in the topic on GDPR-compliant AI.

When a person decides

An extracted value is only usable after it has been checked. For invoices, at least these checks apply.

• Do line items, tax and total match arithmetically?
• Does the supplier exist, and do VAT ID and bank details match the master record? A different IBAN is flagged separately.
• Is there a matching order, and do quantities and prices agree with the order and goods receipt?
• Has this supplier invoice number already been processed?
• Are date, currency and tax rate valid?

When there is doubt, a person decides. For each field we set a confidence threshold. If a value falls below it or a check fails, the document goes to a review screen showing the document and the recognized fields side by side. We set the thresholds on real data, because a wrong amount is more critical than a misread free-text field. We prefer to start with more review and loosen after measuring. Letting too much through automatically means posting wrong amounts. We store corrections and use them for tests and improvements. Without people clearly responsible for the review screen, doubtful cases pile up.

Integration and recognition rate

Handover happens via API. The ERP receives the checked data as a structured record, for example a posting proposal with account assignment, and the document management system gets the original with the extracted details as metadata. Repeated transfers must not create duplicates, and errors during handover must be reported.

Measurement shows whether a solution works. We assemble a test set from real documents, with bad scans, stamps, faxes and special cases, and check the correct values by hand. A demo with clean PDFs says little about that. In our experience, tables with many rows and page breaks are hard, header data is easy, and bulk PDFs have to be split first. We measure accuracy per field, the share of documents that pass without correction, the share going to review and, most important, the errors that slip through unnoticed. The average hides a lot here, because amounts and bank details matter more than free text. In operation we keep measuring, since new suppliers, new layouts and new model versions change the rate.