Expense Categorization

Receipt OCR, GL mapping, policy compliance, and anomaly detection for business expenses.

Workflow

1. Receipt Extraction (OCR)

Use tesseract (local) or Vision API for image receipts; pdfplumber for PDF receipts.

Key fields to extract:

• Vendor name, date, total amount, line items
• Payment method (last 4 digits if visible)
• Tax amount (HST/GST/sales tax)
• Tips/gratuity (separate from subtotal)

# Tesseract OCR on receipt image
tesseract receipt.jpg stdout --psm 4 | python3 scripts/parse_receipt.py

# Or use Claude vision directly for complex layouts

For complex or handwritten receipts → use vision model with prompt in references/ocr-prompt.md.

2. GL Code Mapping

Map extracted expense category to chart of accounts. See references/gl-mapping.md for:

• Standard QBO GL codes for common expense types
• IRS-aligned categories (meals 50%, travel, home office, etc.)
• Crypto/DeFi expense categories

Matching logic:

1. Exact vendor name match (known vendor list)
2. MCC code match (credit card transactions)
3. Keyword match on description/line items
4. Fallback: prompt user to select category

3. Policy Compliance Check

Apply policy rules before approval routing. See references/policy-rules.md for standard rules.

Core checks:

• Per diem limits: Meals >$75 require itemized receipt; travel per diem by city
• Receipt threshold: Receipt required for any expense ≥$25 (IRS standard)
• Time limit: Receipts must be submitted within 30/60/90 days (configurable)
• Duplicate detection: Same vendor + amount ± $1 within 7 days = flag
• Split transactions: Same vendor, sequential dates, amounts just below approval threshold = flag

4. Anomaly Detection

Flag for human review:

• Amount > 2× historical average for that vendor/category
• Weekend or holiday transactions (especially travel/entertainment)
• Round-number amounts (potential personal purchase)
• Vendor in restricted list (casinos, adult entertainment, competitors)
• Missing required fields (date, vendor, business purpose)
• Out-of-state purchases for office supply categories

5. Output Format

{
  "receipt_id": "REC-20260315-001",
  "vendor": "Delta Air Lines",
  "date": "2026-03-15",
  "amount": 487.50,
  "currency": "USD",
  "gl_code": "6200",
  "category": "Travel - Air",
  "policy_status": "approved",
  "flags": [],
  "confidence": 0.94,
  "requires_review": false,
  "notes": "Business purpose required for reimbursement"
}

Batch Processing

For expense report batches:

# Process folder of receipts
import glob
receipts = glob.glob("receipts/*.{jpg,png,pdf}")
results = [categorize(r) for r in receipts]

# Summary stats
flagged = [r for r in results if r["requires_review"]]
total = sum(r["amount"] for r in results)
by_category = group_by(results, "category")

Output batch summary as CSV or feed directly to QBO via qbo-automation skill.

Common Patterns

Credit card statement import:

1. Parse CSV/OFX from bank
2. Match known vendors → auto-categorize
3. Unknown vendors → ML classification or prompt
4. Export mapped transactions to QBO

Expense report approval routing:

• Auto-approve: policy-compliant, under $250, no flags
• Manager approval: $250–$2,500 or single flag
• Finance review: >$2,500, multiple flags, or restricted category

Mileage reimbursement:

• Extract start/end locations + business purpose
• Calculate at current IRS rate (check references/irs-rates.md)
• Map to GL 6210 (Auto/Mileage)

Integration Points

• qbo-automation: Push categorized transactions directly to QBO
• crypto-tax-agent: Route DeFi/crypto expenses for cost basis tracking
• kpi-alert-system: Trigger alerts when department spend exceeds budget
• invoice-automation: Cross-reference receipts with vendor invoices

Negative Boundaries

• Not for PTIN-backed tax work — categorization ≠ tax advice; defer to licensed preparer
• Not for payroll — employee expense reimbursement != payroll processing
• Not a real-time feed — batch review with human sign-off before posting to GL
• Not for legal contracts — use contract-review-agent for vendor agreements
• Confidence \x3C0.7 → always route to human review, never auto-post