Key Takeaways
- Basic duplicate detection catches only 30% of duplicate payments
- 12 distinct fraud patterns require layered detection approaches
- Cumulative fraud (split invoicing) is the most costly and hardest to catch
- Item normalization enables cross-vendor price comparison
- Pure SQL detection costs $0 per check vs paid ML/API services
What is invoice fraud?
Invoice fraud occurs when businesses pay for goods or services that were never delivered, were already paid for, or were billed at inflated prices. Common patterns include duplicate invoices, shell company schemes, price creep, split invoicing to avoid approval thresholds, and PO acknowledgment manipulation. In 2024, 79% of organizations were victims of payment fraud [AFP Payments Fraud Survey 2025], with median losses of $145,000 per case [ACFE 2024].
The Association of Certified Fraud Examiners reports that organizations lose 5% of revenue to fraud annually [ACFE 2024], with billing schemes costing an average of $5,600 per month. For procurement and accounts payable, the most common schemes are billing fraud, check tampering, and expense reimbursement abuse, all of which can be detected with proper invoice validation.
Yet most AP tools only perform basic duplicate checking: "Have we seen this exact invoice number before?" That catches about 30% of duplicate payments. The other 70% slip through because fraudsters simply change the invoice number, adjust the date, or split the invoice.
Duplicate payments represent 0.1% to 0.5% of annual disbursements [AFP 2024], while 1.29% of invoices processed are duplicates averaging $2,034 each [SAP Concur 2024]. For a company processing $50M annually, that's $50K-$250K in overpayments. Most recovery audits find that 80% of these duplicates were preventable with better detection.
The 12 Fraud Patterns
Modern fraud detection requires multiple detection layers, each targeting a specific attack vector. Here are the 12 patterns that comprehensive anomaly detection should catch:
The most obvious fraud: submitting the exact same invoice number from the same vendor twice. This catches vendors resubmitting paid invoices months later, hoping AP won't notice. The median fraud takes 12 months to detect [ACFE 2024], giving fraudsters ample time to resubmit.
Fraudsters change the invoice number but forget to change the amount. Same vendor, same exact amount (within 0.5%), different invoice number, submitted within 30 days = highly suspicious.
Catches typos, OCR errors, and intentional slight modifications: INV-1234 vs INV-I234 (letter I vs number 1), or INV-1234 vs INV1234 (missing hyphen).
Breaking a large PO into multiple smaller invoices that individually look normal but together exceed the PO total. Each invoice passes validation; the fraud is only visible when you sum them up. With median fraud losses of $145,000 per case [ACFE 2024], cumulative schemes can be devastating.
Prices that deviate significantly from historical norms. Uses adaptive thresholds based on sample size: more conservative (5 sigma) with limited data, stricter (3 sigma) with robust historical data.
Ordering dramatically more than typical: 10x, 15x, or 20x the normal quantity. Could indicate theft-for-resale, phantom inventory, or simple data entry errors.
First-time vendors submitting unusually large orders. Classic shell company fraud: create a fake vendor, submit one large invoice, collect payment, disappear. With 79% of organizations experiencing payment fraud in 2024 [AFP Payments Fraud Survey 2025], shell companies remain a top threat vector.
Fraudsters often submit 1-2 small legitimate invoices to establish trust, then hit with a large fraudulent invoice. The "third invoice fraud" pattern is well-documented in procurement fraud literature.
Invoice #2: $3,800 (paid, legitimate)
Invoice #3: $47,000 (fraudulent)
Breaking one large purchase into multiple smaller POs to stay under approval thresholds. If approval is required above $5,000, submit three POs for $4,900 each.
PO-001: $4,800 to Vendor X (Monday)
PO-002: $4,700 to Vendor X (Wednesday)
PO-003: $4,900 to Vendor X (Friday)
Total: $14,400 in one week, zero approvals required.
Gradual price increases that individually seem minor but compound over time. Billing schemes lose companies $5,600 per month on average [ACFE 2024]. This is a direct comparison to the last purchase price.
April: Widget-A at $115 (+15%)
July: Widget-A at $132 (+15%)
December: Widget-A at $152 (+15%)
Each increase "only" 15%, but price up 52% YTD.
Paying significantly more than other vendors charge for the same item. Requires item normalization to compare prices across vendors who use different part numbers for identical products.
Office Depot: $89 (calls it "HP CF294A")
Staples: $92 (calls it "HP-CF294A-BLK")
New vendor: $142 (calls it "cf294a")
Item normalization links all three to the same canonical item.
Paying more than the best price you've ever achieved for an item. Different from cross-vendor: this tracks YOUR historical best, regardless of current vendor availability.
Today, Vendor A invoices Widget-X at $110, 37.5% above your historical best. Even if $110 is "fair market," you've demonstrated the ability to get it for $80.
Bonus: Quote-to-PO Mismatch
Beyond the 12 core patterns, comprehensive detection should also catch unauthorized changes between quote approval and PO issuance:
The quote gets approved at one price, but the PO is issued at a higher price. This can be internal fraud (kickbacks to the person issuing POs) or vendor manipulation.
PO-750 issued at $52.00/unit for 100 widgets.
Difference: $700 unauthorized price increase.
What Makes Detection Effective
Many AP tools claim to detect fraud, but their implementations have critical gaps:
Common Detection Gaps
| Detection Requirement | Basic AP Tools | Comprehensive Detection |
|---|---|---|
| Checks paid invoices for duplicates | Often excludes paid | Always includes |
| Cross-vendor price comparison | No normalization | Item master mapping |
| Cumulative PO validation | Per-invoice only | Sum across all invoices |
| Adaptive statistical thresholds | Fixed 3 sigma | 3-5 sigma based on sample size |
| Vendor alias detection | Exact match only | Canonical mapping |
| Early relationship tracking | No history awareness | First 5 invoice scrutiny |
The Item Normalization Requirement
Cross-vendor price comparison only works if you can match items across vendors. This requires an item master that maps vendor-specific part numbers to canonical item IDs:
Without item normalization, "HP CF294A" and "cf294a" appear as completely different products, making price comparison impossible.
Implementation Considerations
Performance Requirements
Fraud detection must be fast enough to run on every invoice without creating bottlenecks:
- Target latency: Under 100ms per invoice for all 12 checks combined
- Batch optimization: Single query for all line items instead of N queries
- Index requirements: Compound indexes on (vendor_id, company_id, item_id)
- Cost: Pure SQL with no external API calls = $0 per check
Severity Calibration
Not all anomalies deserve equal attention. Severity levels should drive workflow:
- Critical: Automatic payment hold. Human review required before release.
- High: Flagged for review but can proceed with approval. Manager notification.
- Medium: Informational. Logged for audit trail and trend analysis.
The goal is prevention, not recovery. Active fraud monitoring reduces detection time from 12 months to just 6 months [ACFE 2024]. Recovery audits typically retrieve 50-75% of duplicates found, but the effort costs 25-30% of recovery value. Prevention costs a fraction and catches 100% before payment.
Beyond Detection: Building Trust Scores
Individual anomaly flags are useful, but the real power comes from combining them into vendor and invoice trust scores:
- New vendor with first large invoice = -20 points
- Third invoice from low-history vendor = -10 points
- Price 3 sigma above historical average = -15 points
- Split purchase pattern detected = -25 points
- 100+ transactions with zero anomalies = +30 points
A composite score enables automated triage: high-trust vendors get straight-through processing; low-trust invoices get manual review. This balances fraud prevention with AP efficiency.
See All 12 Detection Patterns in Action
Upload an invoice and watch real-time anomaly detection flag potential fraud patterns. No credit card required.
Start Free TrialFurther Reading
- Why 3-Way Matching Isn't Enough: The Case for 5-Way Validation
- The $0 Anomaly Detection System: Pure SQL, Sub-100ms
References
- Association of Certified Fraud Examiners. (2024). "2024 Report to the Nations" - Organizations lose 5% of revenue to fraud annually. Median loss: $145,000 per case. Billing schemes average $5,600/month. Median detection time: 12 months (reduced to 6 months with active monitoring).
- AFP. (2025). "2025 AFP Payments Fraud and Control Survey" - 79% of organizations were victims of payment fraud in 2024.
- AFP. (2024). "Payments Fraud Survey" - Duplicate payments represent 0.1% to 0.5% of annual disbursements.
- SAP Concur. (2024). "Invoice Processing Benchmark Report" - 1.29% of invoices processed are duplicates, averaging $2,034 each.
- Institute of Finance & Management, "AP Duplicate Payment Prevention Best Practices"
- ACFE, "Vendor Fraud Schemes: Detection and Prevention"