AI Document Processing: From 15% to 100% Capture Rate

A mid-market trading company with 120 employees faced a familiar challenge: 1,200 incoming invoices monthly, only 15% processed digitally via EDI from large suppliers, the rest arriving as PDFs or scanned paper - all keyed in manually. That meant over 85 hours of hands-on work each month, error rates around 6%, and constant bottlenecks during month-end close. We solved it with Azure AI Document Intelligence and Power Platform.

The Problem: 85% Manual Data Entry

Before we tackled automation, the daily reality looked like this:

Invoice Volume:

• ~1,200 invoices per month
• ~180 invoices from major vendors via EDI (15% digital)
• ~1,020 invoices via PDF email or paper (85% manual)

Manual processing was a serious time drain:

1. Open email or scan paper document
2. Read supplier name
3. Note invoice number and amount
4. Type into SAP Business One
5. File in folder structure

The cost: time, errors, and missed opportunities:

• Average processing time: 4-5 minutes per invoice
• Monthly effort: ~85 hours (more than one FTE)
• Error rate: ~6% (wrong amounts, incorrect cost centers, duplicates)
• Bottleneck: Month-end pileups, missed payment deadlines, lost early payment discounts (Skonto)

This wasn't just expensive-it was error-prone. The AP team was constantly under pressure, and month-end close suffered.

The Solution: Four Phases of Intelligent Automation

We built a system in four phases where AI and humans work in perfect tandem.

Phase 1: AI Document Intelligence (Azure)

All incoming invoices-whether email or scanned-land in a SharePoint library. A Power Automate flow triggers Azure AI Document Intelligence with a pre-built invoice model.

The system automatically extracts:

• Vendor name and vendor number
• Invoice number and date
• Net and gross amounts
• VAT rate
• Payment terms
• Bank details / IBAN
• Line items

Accuracy:

• For standard invoices: 92-95% correct
• For known vendors (trained custom model): 97-99% correct

Phase 2: Matching & Validation (Dataverse + Power Automate)

Extracted data is now intelligently validated across multiple checks:

1. Vendor Matching: The recognized vendor number is cross-checked against master data in Dataverse. If unknown, similar names are suggested.

2. Purchase Order Reconciliation: Invoice number is matched against open purchase orders and goods receipts (3-way matching).

3. Plausibility Checks:

   • Is the amount within the expected range?
   • Is the tax rate correct?
   • Has this invoice already been processed (duplicate detection)?
   • How much time remains on the early payment discount window?

4. Confidence Scoring: Each invoice receives a confidence score:

   • > 95%: Auto-approved for posting
   • 80-95%: Routed to human review (with pre-filled form)
   • < 80%: Flagged for manual processing

Phase 3: Human-in-the-Loop (Power Apps)

For invoices requiring human review, a custom Model-Driven App provides:

• Extracted data and original PDF displayed side-by-side
• All identified matches and potential conflicts highlighted
• Pre-filled fields for rapid correction
• One-click approval or inline editing

Key advantage: Every correction feeds back into the AI model automatically. The system learns continuously-known edge cases get recognized faster over time.

Phase 4: ERP Integration & Audit Trail

Approved invoices are automatically posted to SAP Business One via API. In parallel:

• A DATEV export is generated for the tax advisor
• The original PDF is archived in SharePoint with full audit trail
• A booking report is created for control purposes

The Results: 86% Less Effort, 0.8% Error Rate

The numbers tell the story:

ROI: The system paid for itself in 4 months.

What changed for the AP team:

Instead of typing invoices all day, the team now focuses on genuine exceptions: unusual formats, international invoice structures, or discrepancies requiring judgment. Routine work is gone-less stress, higher quality.

And month-end? No more stacks of invoices. They're processed throughout the day, payments happen on time, and early payment windows are never missed.

Technology Stack

• Azure AI Document Intelligence (pre-built + custom models)
• Power Automate (orchestration and workflows)
• Microsoft Dataverse (validation rules, master data)
• Power Apps (Model-Driven App for AP team)
• SAP Business One (API integration)
• SharePoint (document management and archival)
• Power BI (monitoring and reporting)

What's Next

If you're facing a similar situation-lots of manual documents, high volume, error-prone processes-this approach is worth exploring. The combination of AI and human-in-the-loop is robust and scalable.

Related automations:

• Invoice Approval Automation
• DATEV XML Integration
• AI Document Processing: Hype vs. Reality
• Data Extraction for Renewable Energy Projects