DUNA Digital
Case study - AI accounting & partner management

Invoices are now read by AI.

For entsorgo we automated incoming-invoice processing: OCR, classification, matching against orders and partner prices - up to the fully prepared booking. Approval stays human.

app.entsorgo.de
7 stagesfrom scan to DATEV pre-booking and sevDesk handover - every booking approved by a human.
Document AIthe inbox that sorts itself.
01 - The starting point

Every partner invoice was manual work.

Invoices, delivery notes and weighing slips from disposal partners arrived as scans and PDFs - often several documents in one file, sometimes with handwritten additions. Checking, matching, entering: all by hand.

02 - The requirement

Automate without giving up control.

Accounting forgives no hallucinations. So the brief was clear: the AI prepares, checks and learns - but not a single booking leaves without human approval.

The solution

Read and sort documents

Every upload lands in the document management system and runs through an AI pipeline: OCR with Claude Vision reads even handwriting, a classifier recognises invoice, delivery note, weighing slip or dunning letter - and splits multi-document PDFs into single documents automatically.

Tools
Claude Vision OCRMayan EDMSClassification

Check and match

Around 30 fields per invoice are extracted - line items included. The system then reconciles documents that belong together (invoice, delivery note, weighing slip of the same order), finds the matching order and validates every position against the stored partner prices. Deviations get flagged, not skipped.

Mechanics
Order matchingPrice validationTrust hierarchy

Learn from corrections

Every human correction is analysed: what gets corrected repeatedly for the same partner becomes a learned pattern in a partner knowledge base - and flows back into the AI prompts with the next document. The system gets better in operation, partner by partner.

Partner management
Learning loopPartner knowledgePatterns from 3 hits

Book - with approval

Approved invoices are pre-booked with DATEV accounts and tax keys, linked to the order and handed to sevDesk automatically. A watchdog process detects stuck documents and repairs the pipeline itself - the human keeps the last word.

Integrations
DATEV pre-bookingsevDeskSelf-healing

The result

7pipeline stages automated: upload, OCR, classification, extraction, reconciliation, validation, booking preparation.
~30fields extracted per invoice automatically - individual line items included.
0bookings without human approval. The AI prepares - the team decides.

Several models, one system.

Built cost-consciously: light models read and extract, strong models classify and reconcile. German documents, German accounting logic, EU-ready architecture - and a pipeline that repairs itself on failures.

Your inbox could do this too. Let's ask it.