Invoices, delivery notes and drawings arrive as finished fields. Your team only looks at what is flagged.
The system reads arrangement, tables and stamps along with the text. A new supplier needs no new template.
Amounts above a limit, tolerances and contract deadlines are always released by someone on your team, even when the system is sure.
Accuracy per field, share processed without a touch and review rate, measured on your documents before go-live and continuously in operation.
45 %
Invoices are becoming structured, step by step. Delivery notes, mill certificates, drawings and contracts are not. They keep arriving as PDF, scan or photo, and someone retypes them.
Bitkom, Dec. 2024, just before receiving became mandatory: fewer than half of German companies could receive e-invoices (1,103 companies with 20 or more employees)Classic OCR needs a template per sender. With every new supplier, every new title block and every table running over two pages, the upkeep starts again.
92.7 %
On scans, even the best model tested gets roughly every fourteenth field wrong. Writing that into the ERP without a confidence score and a review list means finding the errors in accounting or at the machine.
Fraunhofer IAIS and Lamarr Institute, 2025: best of eight multimodal models, fields read correctly on scanned invoices (96.5 % on clean PDFs)Recognise and read
Every document is first assigned its type, because an order confirmation needs different fields from a mill certificate. Then the system reads text and arrangement together, across all pages, with a location for every field.
Items, continued from page 1
Item 4 Round bar 1.4301 Ø 60, 24 mtable, 98 %
Delivery week 44 instead of week 42differs
Stamp and sales initialsrecognised
Check
Read is not the same as right. Every document runs through the rules for its type, and whatever does not fit or is uncertain lands with your team, with its location.
Totals against line items, mandatory invoice details, certificate values against the standard, prices against the order.
Every field has its own threshold, derived from your test documents. Below it, nothing is taken over silently.
Your team sees the document, the flagged spot and the suggestion side by side and decides with one click.
Measure
Before go-live, your real documents become a test set, and we measure per field against it. In operation the same figures keep running, so you see when a new layout or a poor scanner pulls quality down.
400 documents from three months
Invoices as PDF and paper, delivery notes, certificates6 types
Filled in and checked by hand: the right result per fieldreference
Binding per document type: amounts, quantities, tolerancesalways released
Approaches
For reading there are ready-made cloud services, language models that read the page as an image, and open source tools. Measuring on your documents shows which fits; often it is a combination.
Our stack, open source, in Germany
Managed, Microsoft Azure
Managed, Google Cloud
Managed, AWS
The model reads the page image
What the system takes on
What your team takes on
Document Intelligence
Should the data behind it be put in order first, master data, duplicates, text for knowledge systems? Then data & documents.
Does each document need a decision about what happens next? Then an AI agent on the fields read.
Does the checked record always take the same route into the ERP? Then workflow automation is enough.
From the first call to a system in operation. Each phase stands on its own; none ties you to the next.
free
30 minutes. We work out where AI fits for you and whether an AI strategy is worth it.
€1,990
AI strategy for mid-sized companies: which of your processes suit AI, what the build costs and brings back, and which ones are not worth the effort.
from €5,000
AI pilot at a fixed price: one use case, fixed scope and acceptance criteria in writing up front. Built in your systems, tested with your data.
from 20% a year
Running AI systems: hosting in Germany, monitoring, model and prompt maintenance, ongoing measurement, and the next use case on the same foundation.
Frequent questions
The questions that almost always come up.
Intelligent document processing does not just read documents as text, it understands them: it recognises the document type, reads layout, tables, stamps and handwriting, extracts the fields and checks them against your rules. Every field carries a confidence score, anything uncertain lands on a review list.
Classic OCR turns pixels into text, and a template per supplier says where each field sits. Every new layout needs a new template. Current models read text and arrangement together: they recognise tables, title blocks and stamps even in documents they have never seen. We use both, depending on which performs better on your documents.
We measure before we name a figure. A batch of your real documents becomes a test set. On it we report per field how often it is read correctly, what share runs through without a touch and how much lands on the review list. Those three figures are also the acceptance criteria.
We read them directly from the structured data, without OCR. Since 2025 a PDF invoice no longer counts as an e-invoice in Germany; paper and PDF will stay for a long time for delivery notes, certificates, drawings and invoices from smaller suppliers. Both run through the same pipeline and the same checks.
The system reads them as well as it can and says honestly how sure it is. If a field falls below the agreed threshold, it goes to the review list with its location. A blurred photo is flagged, not guessed.
Your team. Amounts above a limit, tolerances, deadlines and termination clauses always go to a person, even when the system is sure. You decide which fields those are, not us.
In German data centres or in your own infrastructure, with a data processing agreement under Art. 28 GDPR in place before the first document flows. Your documents are not used to train third-party models.
The next step
30 minutes, free and without obligation.