AI document processing
Paperwork in,
clean data out.
AI reads invoices, delivery notes and forms, matches them against your data and passes checked fields to your systems. Uncertain cases go to a person.
Discuss your projectFree initial conversation · Directly with Daniel Gläser
- 01Document
- 02Extract fields
- 03Check matching
- 04Hand-over
Document types
Which documents are a good fit.
Good candidates arrive often, contain similar details and are typed in by hand today.
Incoming invoices
Invoice number, date, supplier, amounts, VAT rates and order reference from PDFs, scans or photos. How structured e-invoices fit in is covered in the next section.
Delivery notes
Extract line items, quantities and delivery note numbers and match them against the order and goods receipt. Discrepancies are flagged for clarification.
Order confirmations
Compare prices, quantities and delivery dates with your purchase order. Changed terms stand out before the invoice arrives.
Contracts
Pull out parties, terms and notice periods and add them to your system as reminders. The legal assessment stays with you.
Job applications
Capture contact details and file documents in order. Filtering or scoring applicants with AI is not part of this, as the EU AI Act sets strict requirements for it.
EU AI Act guideForms and email attachments
Applications, claim reports, inspection records or orders sent by email. The attachment is recognised, extracted and linked to the right case.
E-invoicing
E-invoices: where AI helps and where it does not.
Since 1 January 2025, businesses in Germany must be able to receive e-invoices from other domestic businesses. Transition periods for issuing them run out by the end of 2027 at the latest. Only a structured format under the European standard EN 16931 counts as an e-invoice, for example XRechnung or ZUGFeRD in the permitted profiles. A plain PDF is not an e-invoice.
Structured e-invoices are machine-readable without AI. The data arrives as XML and can be imported and validated directly. I use a conventional import for this, not AI.
AI helps with everything unstructured: PDF and paper invoices during the transition, small-amount invoices and invoices from foreign suppliers, which are exempt from the requirement, plus delivery notes, order confirmations and free text in emails. A good intake handles both: XML directly, everything else with AI.
Process
From intake to hand-over.
Intake
Documents arrive from a mailbox, a scan folder or an upload. The original is stored unchanged before anything is extracted.
Classify
AI recognises the document type, such as invoice, credit note, delivery note or reminder, and splits combined PDFs into individual documents.
Extract fields
A language model with vision or a specialised text recognition service reads the agreed fields. Each value is stored with its location in the document, so checking is quick.
Check plausibility
Rules check the result: does the order number exist? Do line items and VAT add up to the total? Do the bank details match the known supplier? Has this invoice number been received before?
Review when uncertain
If a field is missing or a rule fires, the document goes to a review list. Your staff see document and values side by side and correct them on the spot.
Hand-over
Checked data goes to accounting, ERP or DMS, for example DATEV or an inventory system. Depending on the system, this happens via an interface or an export. I check beforehand which options your system offers.
GoBD & data protection
Retention and data protection are part of it.
Extraction does not replace the original. Under the GoBD, the German rules for digital bookkeeping, incoming electronic documents are generally kept in the format they arrived in and are not altered. Since the 2025 GoBD update, keeping the structured part of an e-invoice is usually sufficient. Data obtained by text recognition and then checked is retained alongside the original.
The workflow belongs in your procedural documentation: who captures, checks and passes on documents, and how. I provide the technical description for it. The tax assessment is for you to agree with your tax adviser.
HR documents such as applications or employment contracts contain sensitive details, sometimes including health information. They get tight access rights, short retention periods and, where needed, a local model. Providers, legal basis and data processing agreements are reviewed for the specific use.
Measurement & technology
Measure first, then decide.
Pilot with your own documents
We put together a sample of your real documents, for example invoices from your main suppliers and a few difficult scans, and record the correct values by hand once. Against these we measure the recognition rate per field and document type.
Model chosen by data type
Depending on how sensitive the data is, recognition runs on a cloud service under a data processing agreement or on a local model on your own server.
More on local AICost factors
Document volume, the number of document types and fields, scan quality, the number of connected systems and the share of cases needing review determine effort and running costs.
Limits
Poor scans, handwriting and unusual layouts lower the rate. AI can mix up digits, so rules check every amount. Error-free recognition is not a realistic goal. What matters is how much typing and checking you no longer have to do.
Working together
A clear path to everyday use.
Review your documents
Record document types, volumes, required fields and target systems, and put together a sample.
Measure in a pilot
Test extraction and plausibility rules on your documents and evaluate the rate per field.
Set up review and hand-over
Build the review screen, interface or export, and decide where originals are stored.
Monitor operation
Analyse review cases, refine rules and add new suppliers or document types.
Before we begin
Good questions. Clear answers.
Can AI extract invoice data automatically?
Yes, from scans and photos too. Invoice number, date, supplier, amounts and VAT rates can be captured well when plausibility rules check the result. A pilot shows how well it works with your documents.
Do I need AI for XRechnung and ZUGFeRD?
No. These formats already contain the invoice data in structured form and can be imported without AI. AI is worthwhile for PDFs, scans, delivery notes and free text.
Does it work with DATEV, ERP or DMS?
Depending on the system, via an interface or an export. I check beforehand with you, and if needed with your tax adviser, which options your DATEV setup, ERP or DMS offers.
How high is the recognition rate?
It depends on document quality, layouts and fields. A blanket figure would not be honest. In the pilot we measure the rate per field on your own documents before you decide.
Is AI document capture compatible with the GoBD?
The GoBD require, among other things, unaltered originals, traceability and procedural documentation. The workflow is designed around these. The overall tax assessment rests with you and your tax adviser.
What does AI document processing cost?
Document volume, the number of document types, integrations and review effort determine cost. We start with a pilot whose effort is agreed in advance. Running costs can then be estimated from real volumes.