Software

AI Invoice Processing & ERP Integration

Reduced invoice processing time from 5-10 minutes to 10 seconds using AI vision and ERP automation.

10 secPer invoice (was 5-10 minutes)

The Challenge

A software and ERP consulting company processed 300-400 incoming invoices per month from their own vendors and suppliers. Each invoice required 5-10 minutes of manual work: opening the PDF, identifying the supplier and invoice number, extracting line items and tax amounts, looking up the supplier in their WeClapp cloud ERP to find the correct customer number and cost center, and manually creating the invoice record with all extracted data. The invoices were entirely in German with varying layouts — from simple single-page bills to complex multi-page invoices with multiple tax rates. Staff made data entry errors on roughly 5% of invoices, which cascaded into accounting reconciliation issues at month-end. The monthly close process was consistently delayed by 2-3 days due to invoice backlog.

Our Solution

We built an AI-powered invoice processing pipeline that takes a scanned or digital PDF invoice and produces a complete, verified ERP entry in about 10 seconds. The pipeline starts when an invoice PDF lands in a designated Google Drive folder (via email forwarding rule or manual upload). Make.com detects the new file, sends it to PDF.co for high-quality image conversion, then passes the image to GPT-4o Vision with a carefully engineered prompt that extracts 15+ fields: supplier name, invoice number, date, line items, individual and total amounts, tax rates, IBAN, and payment terms. The extracted data is then matched against the WeClapp ERP supplier database using a fuzzy matching algorithm on supplier name and tax ID. Once matched, the system creates the invoice record in WeClapp via API with all line items, attaches the original PDF, and logs every step to an audit spreadsheet with extraction confidence scores. Invoices with any field below 85% confidence are routed to a review queue rather than auto-processed.

Detailed Approach

The key technical challenge was achieving reliable data extraction from highly varied German invoice formats. We tested three approaches: template-based OCR (too rigid — broke when layouts changed), generic OCR + regex (missed too many edge cases with German number formatting), and vision AI (winner — handled layout variation gracefully).

Our GPT-4o Vision prompt went through 8 iterations, each tested against a validation set of 50 invoices. The final prompt uses a structured output format (JSON schema) that forces the model to extract each field explicitly, with special handling for German number formats (comma as decimal separator, period as thousands separator) and multi-tax-rate invoices where different line items carry 7% vs. 19% VAT. We also built in a "reasoning" field where the model explains its extraction logic, which helps with debugging edge cases.

The WeClapp matching module queries the ERP's supplier API endpoint, searching by tax ID first (most reliable), then falling back to fuzzy name matching using trigram similarity. For new suppliers not yet in the ERP, the system creates a draft supplier record pre-populated with extracted data and flags it for human approval. The Make.com scenario includes comprehensive error handling: PDF conversion failures retry 3 times, API timeouts trigger delayed retry, and any unrecoverable error sends a Slack notification with the invoice attached for manual processing.

Key Results

🕐 We collected 50 sample invoices representing the full range of supplier formats for prompt engineering

✓Processing time per invoice reduced from 5-10 minutes to ~10 seconds
✓Data entry error rate dropped from 5% to 0.3% (review-queue catches)
✓Monthly close process accelerated by 2.5 days — invoices now processed same-day
✓93.8% of invoices fully auto-processed without human intervention
✓Staff reallocated 60+ hours/month from data entry to strategic finance work
✓Complete audit trail for every invoice — supports tax compliance requirements

We collected 50 sample invoices representing the full range of supplier formats for prompt engineering. Development and testing took 3 weeks. A 2-week validation period processed invoices through both manual and automated paths to verify accuracy before going fully automated.