AI Invoice Extraction and DATEV Export Automation

Background

Carrivo is a car subscription brand in Germany, backed by a major vehicle importer and automotive group. Beyond its primary OEM relationship with Stellantis, it receives invoices from multiple OEMs, each arriving in its own PDF structure.

Some invoices are text based. Others are scanned images with no machine readable text at all. Every one of them had to be read by a person before the accounts team could produce a DATEV import file.

The challenge

• Invoices arrived from multiple OEMs in incompatible PDF structures
• Some documents were scanned images with no extractable text
• Operations staff read each invoice manually to pull vendor name, amounts, account codes and invoice numbers
• The DATEV import file follows a strict 124 column EXTF v7.0 specification, semicolon delimited and ISO-8859-1 encoded
• Any small formatting error broke the entire import
• The real difficulty was reliable extraction of structured accounting data from unstructured documents, not generating the file itself

Goals

• Extract structured accounting data from varied OEM invoice formats without per supplier setup
• Handle scanned and image based invoices alongside text based PDFs
• Map extracted fields onto the full 124 column DATEV structure automatically
• Keep a person in the loop wherever confidence is low, before anything reaches DATEV
• Produce a fully compliant EXTF v7.0 file ready for direct import
• Keep all processing within EU data residency requirements

What JRNY Services delivered

JRNY built the DATEV PDF Ingest module inside the Carrivo Fleet Ingest Platform as a five stage pipeline.

Upload

PDF invoices or ZIP archives, single or batch.

Extract

Readable PDFs are parsed directly, while scanned pages fall back to Google Cloud Vision OCR.

Map

An AI agent maps invoice fields onto the 124 column DATEV structure, with no per supplier configuration required.

Review

Confident runs export automatically. Uncertain fields are flagged for manual review and inline editing in the dashboard.

Export

A DATEV EXTF v7.0 file is generated, ready for direct import.

Supporting features

• Vendor Map for quick lookup of recurring supplier details
• Multi-language interface in English and German
• Full DATEV EXTF v7.0 compliance, including delimiter and encoding rules

Technology stack

• Frontend React and TypeScript for a dashboard interface with inline editing
• Backend n8n Cloud Pro for workflow orchestration and AI agents
• AI extraction n8n AI agents for structured data extraction
• OCR Google Cloud Vision API as fallback for scanned PDFs
• Database BigQuery on Google Cloud for record storage and run tracking
• Output format DATEV EXTF v7.0, 124 column, semicolon delimited, ISO-8859-1
• Infrastructure Google Cloud Frankfurt for EU data residency compliance

The hardest part

Extracting dependable accounting data from documents that were never designed to be machine read. Every OEM structures its invoices differently and some arrive as images, so the module had to combine direct parsing, OCR fallback and AI field mapping, then be honest about its own uncertainty rather than pushing a guess into an accounting system.

Results

• Around 85 percent of the manual time removed from invoice processing
• Most invoices processed automatically, with only uncertain cases flagged for review
• Individual invoices processed in seconds rather than minutes of manual reading
• DATEV import errors eliminated
• Full EXTF v7.0 compliance maintained across every export
• The accounts team moved from manual reading and file building to review and approval

Key milestones

• AI extraction layer proven across multiple OEM invoice formats without per supplier setup
• OCR fallback introduced for scanned and image based documents
• Human in the loop review model established ahead of any DATEV export
• Full 124 column EXTF v7.0 mapping validated against live imports
• English and German interface delivered for the accounts team