Extraction from PDF and scan
Including email attachments and photographed pages. Layout changes at the supplier do not break it.
Product · Document Intelligence
Documents become data and decisions.
Invoices, contracts, forms and reports are read out, checked against your rules and handed to the system that needs them. The same pipeline runs backwards too: from checked data and your templates come finished documents, with a log for every run. The reading happens on a local model, and your paperwork stays in the house.
In preparation
The reading, the checking and the generating all run today in our own installations. What is still being finished is the rule editor for your own team. Until then we maintain the rules, and pilot customers decide which ones get built first.
- Local model, no cloud
- Servers in Germany
- Rules in plain language
- A log for every run
One invoice, read and checked
invoice-2026-0417.pdf
Fields read
- Invoice number
- 2026-0417
- Net amount
- 4,180.00 EUR
- VAT rate
- 19 percent
- Service period
- not stated
Rules applied
- Mandatory details under section 14 VAT Actcomplete
- Routing ID check digitvalid
- Service period presentfor review
The third line is the point. The invoice is not rejected and not waved through, it lands on a review list with the passage that should have contained the date.
What actually eats the time in document handling
Rarely the typing. It is the checking, the asking back and the moment six weeks later when nobody can reconstruct why this invoice was approved.
| By hand, or with a rigid template reader | Document Intelligence | |
|---|---|---|
| New supplier layout | The template no longer matches and the fields land in the wrong place | Fields are recognised by label and context, not by a fixed coordinate |
| Unreadable passage | Somebody types in what they think it says | The field stays empty and the document goes on a review list |
| Checking | Lives in the head of the person who has done it for years | Written as rules, in plain language, in one place |
| A failed check | Says invalid, and the search begins | Names the rule, the value and the spot in the document |
| Evidence six weeks later | Reconstructed from memory and an email thread | Every run logged with rules, result and the file |
| Where the documents go | Uploaded to whichever online tool was quickest | Read on a local model, nothing leaves the building |
From the inbox to the booked record in five steps
The pipeline is deliberately boring. Every step produces something you can look at, which is what makes the fifth one trustworthy.
- 1
Arrival
automatic
A mailbox, a folder, an upload or an interface. The document arrives where it already arrives today, and nobody changes their habits.
- 2
Read out
seconds
Text, tables and single fields are extracted, from scans as well. Every value carries the spot in the document it was taken from.
- 3
Check
instant
Your rules run across the values. The result is readable: which rule, which value, passed or not, and where to look.
- 4
Decide
with you
Clean goes through. Doubtful goes on the review list with the document excerpt beside it. Nothing is accepted quietly.
- 5
Hand over
automatic
Into your accounting, your ERP, your database, or back out as a finished document. Every run is in the log.
The same route runs in reverse: checked data plus your templates produce documents, in series if needed, each one logged.
Two directions, one pipeline
Reading and producing are usually bought as two systems. They are the same problem seen from both ends, and splitting them means maintaining your rules twice.
Rules in plain language
Mandatory details, value ranges, deadlines, check digits, comparison against your own records.
Readable check results
Which rule, which value, which spot in the document. Never just a red cross.
Review list
Everything uncertain in one work queue, with the excerpt next to it and a tick when it is done.
Documents from templates
Word and PDF generated from checked data, in your layout, individually or as a series.
Draft and binding version
A draft carries a visible watermark on every page. The binding version demands complete entries and is logged.
Handover to your systems
Structured file, interface or straight into the database. We connect to what you run.
Audit trail
Who, when, which rules, which result, which file. The question six weeks later has an answer.
Why the rules have to be readable
A check result that only says failed is worth nothing, because nobody knows what to do next. So every rule carries its plain wording, the value it tripped over and the spot in the document. On an incoming invoice that means naming the mandatory detail that is missing under section 14 of the German VAT Act, and where it should have been.
The second half of the same rule: the model does not fill a gap. If a value is not in the document, the field stays empty and the document goes on the review list. An invented invoice number does not look like a mistake, it looks like an invoice number.
Mandatory details under section 14 VAT Act
complete
Which rule, which value, which spot in the document. Never just a red cross.
Service period present
for review
Everything uncertain in one work queue, with the excerpt next to it and a tick when it is done.
One price, three sizes
The packages differ in how many documents run through per month. Everything else is in all of them.
Setup
from €3,900 once
Operation
from €250 a month
Small
Documents per monthup to 500
FitsIncoming invoices at a company with a manageable supplier list.
Become a pilot customerMedium
Most chosenDocuments per monthup to 5,000
FitsSeveral document types with their own rule sets.
Become a pilot customerLarge
Documents per monthunlimited
FitsShared service centres and anyone processing for several entities.
Become a pilot customerIn every package
- Rule set written from your own documents
- Review list with document excerpts
- Reading and generating in one pipeline
- Audit trail for every run
- Handover into your existing systems
- Hosting, updates and support
These prices apply from general availability. Pilot customers get different terms and decide which rules are built first.
Local model, no cloud. KI-Server
Questions before the first ten documents
How does it recognise the fields?+
By label, position and context rather than a fixed coordinate, which is why it survives a supplier redesigning their invoice. Critical fields are checked against a rule on top of that.
What happens with bad scans?+
Whatever is legible gets read, the rest stays empty and goes on the review list. A guessed amount would be worse than an empty field.
Which rules can we define?+
Anything you can write as a sentence: mandatory details, value ranges, deadlines, comparison against your own records, arithmetic like check digits. For electronic invoices the checks against EN 16931 and the German VAT Act are already built.
Do the documents stay in the house?+
Yes. Reading runs on a local model, on our server or on yours. No document is sent to a cloud provider.
How do the results reach our accounting?+
As a structured file, through an interface or straight into your database. Which route your system supports is something we settle before setup.
Can it produce documents as well?+
Yes, that is the same route in reverse. Checked data plus your templates produce finished documents, with draft marking and a log.
What is not finished yet?+
The rule editor for your own team. Today we maintain the rules for you, and from general availability you will do it yourself. The reading, the checking and the generating already run.
What happens if we stop?+
You keep the rule set, the extracted data and the log. It sits in your database and runs on open technology.
Send us ten documents.
Ten invoices, forms or contracts, ideally the ugly ones. You get back which fields we read with confidence, which rules would catch, and where we would have to pass.
Become a pilot customer
No obligation, and the documents never leave our machine.