Tauraco.ai
FREN
My accountInvest
MarketplaceInvestorsInsightsTauraco.io →
FREN
My accountInvest
2 min readBy Sega Diarrah

Factoring anti-fraud: 7 patterns we systematically block

Factoring concentrates specific fraud risks: fictitious invoices, cross-factor duplicates, over-invoicing. Here are the 7 patterns our AI detects.

Why factoring is a prime target

A factored invoice means cash out the door within 24-48 hours against a paper claim. That's exactly the profile a fraudster looks for: fast liquidity, limited traceability, multiple parties (supplier, buyer, factor) which complicates coordination.

Global factoring fraud losses are estimated at 0.3 % to 0.8 % of factored volume depending on the market. Our target at Tauraco: stay under 0.1 %, by blocking upstream.

The 7 systematically detected patterns

1. Cross-factor duplicates

A supplier sells the same invoice to two different factors. Detection: content hash + fuzzy match on (number, amount, issuer) against a shared registry. Tauraco shares this history with other players through a sector-wide initiative.

2. Same-platform duplicate

Simpler variant — the same invoice proposed twice. Near-instant detection via hash(pdf) + (number, issuer, amount). Blocked upstream.

3. Supplier = buyer

An invoice issued by a company to itself via a subsidiary. Detection through SIREN / SIC / OHADA RC cross-check + ultimate beneficiary analysis. "Hard" flag → automatic rejection.

4. Fictitious buyer

The recipient does not exist (SIREN inexistent, OHADA RC not found, invalid address). Detection via public APIs (INPI, Banque de France, OHADA registries, court clerks). Pre-pipeline blocking.

5. Over-invoicing patterns

Suspiciously round amounts (€10,000, €50,000), numeric sequence in invoice numbers (FACT001, FACT002, FACT003 all issued the same day to the same customer). These patterns are rare in real activity; combined they trigger manual review.

6. Phantom suppliers

A supplier never seen before, recently incorporated, no tax history, no digital footprint (no website, no LinkedIn), mailbox-only address. Strong negative score boost, mandatory manual review for the first deals.

7. Date manipulation

Backdating to game scoring ("fresher" receivable) or postdating to artificially extend due date. Detection via PDF metadata reconciliation, SFTP/email/upload timestamp, and supplier history.

What happens when a signal fires

Three response tiers:

  • Soft alert: isolated signal, score ≥ 30. Note on the file, no block. Operators can see the alert.
  • Hard alert: critical signal (e.g. supplier = buyer), or combination of multiple soft alerts (score ≥ 60). Automatic block, compliance officer review.
  • Trusted shutdown: pattern indicating organised activity (multiple accounts, shared IPs, structurally similar documents). Account suspension, reporting to Tracfin / national FIU depending on country.

What we do NOT have in the model

For transparency: no discriminating signals (gender, name origin, place of birth, etc.). The model is trained purely on behavioural and financial features.

Going deeper