Invoice processor

Invoice Processor

Extracts structured financial data from invoice PDFs, scanned images, and email attachments — capturing vendor details, line items, totals, and payment terms — and routes invoices through configurable approval workflows with validation and exception handling.

When to Use

• User provides an invoice PDF or image and needs the data extracted
• Accounts payable workflow requires automated invoice ingestion
• Multiple invoices need to be batch-processed and entered into a system
• An invoice needs to be validated against a purchase order or contract
• Invoice data needs to be exported to an accounting system (QuickBooks, Xero, SAP, NetSuite)
• User asks to automate the invoice approval routing process
• Duplicate or fraudulent invoices need to be detected

Process

1. Ingest the document:

   • Accept: PDF, JPEG/PNG scan, TIFF, email attachment, or forwarded email body
   • Detect document type: standard invoice, credit memo, pro forma invoice, receipt, or statement
   • Assess quality: if the document is blurry, skewed, or low-resolution, attempt deskewing and enhance contrast before extraction; flag if quality is too low for reliable extraction

2. Extract structured fields using OCR and NLP:

   Header fields:

   • Invoice number / reference ID
   • Invoice date and due date
   • PO number (if referenced)
   • Vendor name, address, tax ID / VAT number, contact email/phone
   • Bill-to name and address
   • Ship-to name and address (if different)
   • Payment terms (e.g., "Net 30", "Due on receipt")
   • Currency

   Line items (for each line):

   • Description / item name
   • Quantity
   • Unit price
   • Discount (if applicable)
   • Line total (quantity × unit price − discount)

   Footer fields:

   • Subtotal
   • Tax amount and tax rate(s)
   • Shipping / handling charges
   • Total amount due
   • Bank details / payment instructions (IBAN, ACH, wire)

3. Validate extracted data:

   • Verify arithmetic: confirm subtotal = sum of line totals; total = subtotal + tax + shipping
   • Flag discrepancies (e.g., line items sum to $1,480 but subtotal reads $1,450) as extraction errors or document errors
   • Cross-reference against any provided PO: check vendor, amounts, and line items match
   • Detect potential duplicates: flag invoices with same vendor + amount + invoice number as a prior processed invoice
   • Validate dates: due date should be ≥ invoice date; invoice date should not be more than 90 days in the past (flag as stale)

4. Flag exceptions:

   • 🔴 High priority: arithmetic mismatch, duplicate invoice, missing mandatory fields (invoice number, amount, vendor), amounts above approval threshold
   • 🟡 Medium: missing PO reference, date anomalies, unfamiliar vendor name, unusually large line item
   • 🟢 Low: minor formatting issues, optional fields absent

5. Route for approval:

   • Apply approval routing rules based on: invoice amount, vendor, department/cost center, or expense category
   • Default rules (customizable):
     • ≤ $500: auto-approve if PO matches
     • $500–$5,000: line manager approval

     • $5,000: finance director approval

   • Output: approval task assigned to the appropriate approver with invoice summary attached

6. Export to accounting system:

   • Format the extracted data as the target system's required format (JSON, CSV, XML, API payload)
   • Map fields to accounting system schema (e.g., vendor → supplier_name, total_due → invoice_amount)
   • Include: extracted values, confidence scores, and flags for human review

Output Format

Extracted Invoice Data

{
  "document_type": "Invoice",
  "extraction_confidence": 0.97,
  "flags": [],
  "header": {
    "invoice_number": "INV-2025-00482",
    "invoice_date": "2025-06-01",
    "due_date": "2025-07-01",
    "payment_terms": "Net 30",
    "po_number": "PO-8821",
    "currency": "USD",
    "vendor": {
      "name": "Acme Design Studio",
      "address": "123 Creative Blvd, Austin, TX 78701",
      "tax_id": "82-1234567",
      "email": "[email protected]"
    },
    "bill_to": {
      "name": "Globex Corporation",
      "address": "456 Corporate Ave, New York, NY 10001"
    }
  },
  "line_items": [
    { "description": "Logo Redesign", "qty": 1, "unit_price": 2500.00, "total": 2500.00 },
    { "description": "Brand Style Guide", "qty": 1, "unit_price": 1500.00, "total": 1500.00 }
  ],
  "totals": {
    "subtotal": 4000.00,
    "tax_rate_pct": 8.25,
    "tax_amount": 330.00,
    "shipping": 0.00,
    "total_due": 4330.00
  },
  "payment_instructions": {
    "bank": "Chase Bank",
    "account": "****4821",
    "routing": "021000021"
  },
  "validation": {
    "arithmetic_check": "PASS",
    "duplicate_check": "PASS",
    "po_match": "PASS"
  },
  "routing": {
    "approval_required": true,
    "approver": "[email protected]",
    "reason": "Amount $4,330 exceeds $5,000 manager threshold — routed to Finance Director"
  }
}

Exception Report

🔴 EXCEPTION: INV-2025-00483
Issue: Arithmetic mismatch — line items sum to $1,480.00 but subtotal reads $1,450.00. Difference: $30.00.
Action: Flagged for manual review. Do NOT auto-approve until vendor confirms correct total.
Assigned to: [email protected]

Examples

Example Input

Extract all data from this invoice PDF and check if the math is correct. Route it for approval.

Example Output

✅ Extraction complete (confidence: 96%)

Invoice: INV-2025-00482 | Vendor: Acme Design Studio | Total Due: $4,330.00
Due Date: July 1, 2025 | PO Match: ✅ PO-8821

Line Items:
1. Logo Redesign × 1 — $2,500.00
2. Brand Style Guide × 1 — $1,500.00
Subtotal: $4,000.00 + 8.25% tax ($330.00) = $4,330.00 ✅ Math verified.

No duplicates found. Routing to [email protected] for approval ($4,330 > $5,000 threshold).

Boundaries

• Do NOT auto-approve invoices without human sign-off when amounts exceed the defined threshold.
• Flag all arithmetic mismatches and duplicate invoice candidates for human review — never process them silently.
• Do NOT store invoice data, including bank account or tax ID details, beyond the immediate processing task.
• Treat all extracted financial data as confidential — do not log or expose vendor banking details in plain-text outputs that could be intercepted.
• If extraction confidence is below 85% for critical fields (total amount, vendor name, invoice number), flag for manual re-entry rather than proceeding with low-quality data.
• Do NOT make payment on behalf of the user — output is for review and system ingestion only.