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

AI invoice scoring: what the model actually looks at

The A-to-E grade attached to every invoice combines OCR confidence, fraud signals, debtor credit quality and supplier history. Here's what goes into the number.

Why a graded score, not a simple "accept / reject"

On the Tauraco marketplace, every invoice receives a grade from A to E. This score drives two things:

  • The rate offered to investors (from 80 bps for an A to 400+ bps for a D)
  • Visibility on the marketplace (A's surface first, E's are never listed)

A binary "accept / reject" pushes you to either reject too many viable deals or mis-price risk. The A-to-E grade lets you fund more broadly without diluting portfolio quality.

The 4 dimensions of the model

1. OCR extraction quality (~15 % weight)

The Tauraco OCR extracts: invoice number, dates (issue/due), amounts (net/VAT/gross), party identifiers (SIRET / SIC / OHADA RC), bank accounts. A poorly scanned document, an algorithmically-generated PDF with inconsistencies, or an unexpected format pushes confidence down. Confidence < 75 % triggers automatic human review.

2. Proactive fraud signals (~30 % weight)

Several detectors run in parallel:

  • Duplicate: same invoice already factored elsewhere (content hash
    • fuzzy match on number and amounts)
  • Date coherence: issue > due, due date in the past, abnormal gap
  • Amount coherence: VAT not matching local rates, total ≠ net + VAT
  • Supplier = buyer: classic fraud signal
  • Fictitious counterparty: SIREN that doesn't exist or has been struck off
  • Amount-dressing pattern: suspiciously round amounts, abnormal numeric run

Any alert with score ≥ 60 plus a "hard" flag (e.g. supplier = buyer) triggers automatic rejection before human review.

3. Debtor credit quality (~35 % weight)

The dominant factor in the final score. We combine:

  • Public solvency data (Banque de France, Greffes, OHADA BIB)
  • Payment history on Tauraco when available
  • Sector and macro exposure
  • Size (revenue, headcount)
  • Country and sovereign risk

A listed blue-chip or recognised public operator gets a very high implicit credit score. An SMB in a struggling sector pulls the grade down.

4. Supplier history on the platform (~20 % weight)

Suppliers with a clean track record on Tauraco see their grade marginally improved. Those with an incident (late payment, dispute) see their next invoice drop one notch.

A score is reviewable: not a black box

Scoring isn't opaque. Each invoice ships with a breakdown of contributions: "grade B because OCR confidence 92 % + debtor grade A

  • supplier on first deal". Operators can request a review with justification.

Observed metrics

  • Default rate on A portfolio: 0.4 % over 14 months
  • Default rate on B portfolio: 1.6 %
  • Fraud-detection precision: 97 % (~3 % false positives)
  • Average scoring latency: 42 seconds

Going deeper