Invoice processing for European SMEs
They only check the ones the system isn't sure about. Everything else lands in your accounting system, validated.
German, French, Dutch and English invoices. PDF, scan or email.
New: voice agents that answer your phone line| Bezeichnung | Menge | Betrag |
|---|---|---|
| Faltkarton 40×30×20 | 800 | 680,00 |
| Stretchfolie 500 mm | 40 | 460,00 |
| Palettenversand | 1 | 100,00 |
01The problem
Someone opens each PDF, finds the invoice number, the date, the net amount, the VAT, and types them into the accounting system. Then the next one.
At 3 to 5 minutes per invoice, that is about 20 hours a month. It is also where the errors come from: a transposed digit, a wrong VAT rate, the same invoice paid twice because it arrived by post and by email.
02The work
This is the review screen your team works in. Invoices arrive by upload or a forwarding address. Each one is read, checked and sorted before anyone opens it.
60-second walkthrough
25 mixed invoices from upload to export into the accounting system. No cuts, no edits.
| Supplier | Date | Lang | Gross | Status |
|---|---|---|---|---|
Müller Verpackung GmbH RE-2026-04817 · 1.475,60 € | 14.09.2026 | DE | 1.475,60 € | Ready |
Transports Lefèvre SARL F-88213 · 3.912,00 € | 13.09.2026 | FR | 3.912,00 € | Needs review 20 % TVA on cross-border service |
Van den Berg Logistiek B.V. 2026-1142 · 846,70 € | 12.09.2026 | NL | 846,70 € | Ready |
Kessler Bürobedarf KB-55120 · 312,40 € | 10.09.2026 | DE | 312,40 € | Duplicate Same as KB-55120, received 02.09. |
Nordic Paper Supply Ltd INV-30417 · 1.120,00 € | 10.09.2026 | EN | 1.120,00 € | Ready |
Brandt & Söhne KG 2026/0932 · 5.604,25 € | 09.09.2026 | DE | 5.604,25 € | Needs review Lines ≠ net (26,05 €) |
Every invoice gets one of three states. Ready means every check passed. Needs review means a person should look. Duplicate means we have already seen this supplier, number and amount.
The original document on the left, the extracted fields on the right. Click a field to see where on the page it came from.
Checks run in plain code, not in the model. If the line items do not add up to the net amount, the invoice stops here.
Sample data. Supplier names and numbers are fictional.
03How it works
Drag PDFs into the dashboard, or forward supplier emails to a dedicated address. Scans and phone photos work too.
Supplier, VAT ID, invoice number, dates, line items, net, VAT and gross. Each field gets a confidence score and a link to its position on the page.
Lines must sum to net. Net times rate must equal VAT. The VAT ID must be valid. Supplier, number and amount must not match an earlier invoice.
Invoices that pass go straight to your system. The rest wait in a queue with the reason shown. One click to approve, fix or reject.
Output
Clean, typed data. Dates in ISO format, amounts as numbers, VAT rates as decimals. We map it to whatever your accounting system imports: an API call, a CSV, or a DATEV-format export.
{
"supplier": {
"name": "Müller Verpackung GmbH",
"vat_id": "DE284519736"
},
"invoice_number": "RE-2026-04817",
"invoice_date": "2026-09-14",
"due_date": "2026-10-14",
"currency": "EUR",
"net": 1240.00,
"vat_rate": 0.19,
"vat": 235.60,
"gross": 1475.60,
"checks": {
"lines_equal_net": true,
"vat_matches_rate": true,
"duplicate": false
},
"status": "ready"
}04Voice agents
Most calls to a wholesaler or logistics firm are the same few questions. The voice agent answers them from your own systems, in the caller's language, and hands everything else to a person with a written summary.
Guten Abend, hier ist der digitale Assistent von Kessler Handel. Wie kann ich helfen?
Good evening, this is Kessler Handel's AI assistant. How can I help?
Ich warte auf die Lieferung zu Bestellung 11873.
I'm waiting for the delivery on order 11873.
Die Ware ist gestern rausgegangen und kommt am Freitag, dem 2. Oktober. Soll ich Ihnen die Sendungsnummer per SMS schicken?
It shipped yesterday and arrives Friday, 2 October. Shall I text you the tracking number?
Ja, bitte. Und auf der Rechnung fehlt unser Rabatt.
Yes, please. And our discount is missing from the invoice.
Die SMS ist unterwegs. Die Rabattfrage gebe ich an die Buchhaltung weiter, Sie bekommen morgen bis 12 Uhr einen Rückruf.
The text is on its way. I'm passing the discount question to accounts; you'll get a call back by 12:00 tomorrow.
05Built for Europe
06Services
We map one workflow and show where automation pays off.
One working agent, invoice or voice, integrated with your tools.
Monitoring, improvements and support.
07Process
A typical engagement takes about eight weeks from first call to production. You can stop after any step.
You describe the process. We tell you honestly whether it is a good fit for automation.
You send ten real invoices, redacted if you prefer. We run them and show you the extracted data and the flags.
We test on a larger sample, map the target system, and give you a fixed price for the pilot.
We build the agent, the rules and the integration. You see progress every week.
Your team keeps its current process while the system runs alongside. We compare every result.
The system takes over. Your team works the review queue. We monitor accuracy each month.
08About
I have spent three years building enterprise ERP and invoicing systems with .NET, Angular and SQL Server: multi-tenant invoicing, credit notes, delivery notes, bank reconciliation. More recently, LLM integration with function calling inside those same systems.
That work taught me what business data looks like in practice. Supplier names spelled four ways. Credit notes that reference invoices from last year. Totals that are off by a cent because of rounding per line. A model alone does not handle that. A model plus strict rules plus a person for the edge cases does.
StackVerx is small on purpose. You talk to the person who writes the code.
09FAQ
It will, sometimes. That is why the model is not the last step. Every field has a confidence score, and every invoice passes arithmetic and consistency checks written in plain code: lines sum to net, VAT matches the rate, the VAT ID is valid, the invoice is not a duplicate. Anything that fails or scores low goes to your review queue. Corrections are logged and used to improve extraction for that supplier.
Your choice. An EU cloud region, such as Frankfurt, or your own infrastructure. We sign a data processing agreement before the first document, you set the retention period, and we do not use your documents to train models.
Anything with an API or an import format. For most clients that means a direct API connection, a CSV import, or a DATEV-format export. We confirm the exact route during the audit, before you commit to a build.
Three to six weeks. It depends mostly on how many document types you have and how the integration works. A single invoice type into a system with a good API is at the short end.
No. We need one person who knows the process and can give us about two hours a week during the pilot, plus access to the target system. Your team uses a web dashboard.
Yes. The voice agent says so in its first sentence, as the EU AI Act requires, and callers can ask for a person at any time. If you record calls, callers hear a recording notice before the conversation starts.
It either gets its own number or takes calls you forward to it, for example after 18:00 or when every line is busy. That works with most office phone systems, including SIP-based and cloud telephony. Transfers to a person go back to your existing numbers.
Pricing is on request, because it depends on volume, document types and integration. The audit ends with a fixed price for the pilot and an estimate of monthly running costs per invoice, so you can compare it with the hours you save.
Redact anything you like. We delete them after the demo unless you ask us to keep them.