Automate Invoice Processing and Payment Workflows

This skill designs automated invoice processing workflows spanning the full invoice-to-payment lifecycle: intake (email, portal, EDI, API), classification, OCR/ML data extraction, validation, approval routing, ERP integration, and compliant archival. Target performance metrics anchor the design - 80%+ straight-through processing for standard invoices, 99%+ accuracy on critical fields, under 24-hour processing time, and under 15% exception rate.

Extraction is covered at two levels: a classic OCR pipeline (InvoiceOCR) using pytesseract on preprocessed images (denoising, adaptive thresholding via OpenCV) with regex patterns for invoice number, PO number, amount, and date; and an ML-based extractor using Azure's DocumentAnalysisClient with the prebuilt-invoice model, keeping only fields above a confidence threshold. Validation is implemented as a rule chain (InvoiceValidator) checking vendor existence/status, PO three-way match with a configurable amount-variance tolerance, duplicate detection, tax calculation, and line-item consistency, each returning a PASS/WARNING/FAIL result.

def validate_po_match(self, invoice_data):
    po = self.po_db.get(invoice_data.po_number)
    tolerance = Decimal('0.05')
    amount_diff = abs(invoice_data.total_amount - po.amount) / po.amount
    if amount_diff > tolerance:
        return {'status': ValidationResult.WARNING, 'message': f'Amount variance: {amount_diff:.2%}'}
    return {'status': ValidationResult.PASS}

Workflow orchestration uses a state machine (RECEIVED -> PROCESSING -> PENDING_APPROVAL -> APPROVED/REJECTED -> PAID) with SLA hours per state and amount-based approval-level routing (supervisor under $1,000, manager under $10,000, director+finance above). ERP integration is shown for SAP via pyrfc, calling BAPI_INCOMINGINVOICE_CREATE with header, line-item, and GL-account data. Exception handling attempts auto-resolution for known exception types (missing PO, amount mismatch, duplicate) before escalating to manual review with processor notification.

Use this skill when designing or extending an automated accounts-payable invoice pipeline that needs OCR/ML extraction, business-rule validation with three-way matching, amount-based approval routing, ERP posting (e.g. SAP), and exception handling with escalation. It applies to both greenfield AP automation builds and adding specific capabilities (validation rules, ERP integration, exception auto-resolution) to an existing pipeline.

It is not the right tool for one-off manual invoice entry at low volume where automation overhead isn't justified, or for non-invoice document processing (contracts, receipts without PO matching needs) that doesn't fit the invoice-specific validation and approval model here.

Input: incoming invoice documents (PDF/image) from any intake channel, along with vendor and PO reference data for validation. Output: structured InvoiceData extracted via OCR or ML, a validation result per business rule, a routed approval workflow with SLA-tracked state transitions, an ERP-posted invoice record (e.g. via SAP BAPI), and exception cases for anything that couldn't be auto-resolved.

Built on pytesseract/OpenCV for classic OCR, Azure DocumentAnalysisClient (prebuilt-invoice model) for ML-based extraction, and pyrfc for SAP BAPI integration (BAPI_INCOMINGINVOICE_CREATE). Workflow orchestration integrates with an approval matrix and notification service; scalability guidance references message queues and containerized, cloud-native deployment.

Finance systems engineers and RPA/automation developers building or extending accounts-payable automation - particularly teams needing OCR/ML extraction, three-way match validation, SLA-tracked approval routing, and ERP posting with auditable exception handling.