Design Spec Review — AffyScore Pitch Site — 1 June 2026
Behavioural credit intelligence for South African lenders
The credit signal the bureau doesn't have.
Bank statements in. Explainable decision pack out — behavioural score, affordability, fraud checks, reason codes, and outcome recommendation. Programmatic risk scoring via API or portal.
Half your declines aren't bad credit. They're no credit.
50%
Thin file
of declined applications have insufficient bureau data. Not bad history — no history.
20–45 min
Manual grind
per application reading statements, categorising transactions, plugging numbers into spreadsheets. Qualified staff doing work a machine should do.
0
Invisible signals
bureau data points from bank statement behaviour. Income regularity, cash buffer, gambling spend, returned debits — none of it reaches the credit decision.
See it run
Watch it work.
2-minute walkthrough: bank statements uploaded, decision pack delivered.
Loom demo — coming soon
How it works
Bank statements in. Decision pack out.
Three months of statements. Six stages. One complete, defensible credit decision.
01
Intake
Operator upload or customer self-upload via tokenised WhatsApp link. Batch up to 30 statements.
02
Extract
Six SA banks parsed by fingerprint — FNB, Standard Bank, ABSA, Nedbank, Capitec, Discovery. Regex extraction with AI vision enhancement for scanned and complex documents.
03
Tamper Check
18 document tampering checks across metadata, font/layout, mathematical, and sequence anomalies. Advisory — informs the human, never blocks.
04
Classify
Every transaction categorised into 17 NCR categories. Counterparty identification. Salary detected by recurrence.
05
Affordability
Regulation 23A on statement income. Disposable income, max instalment, norm-floored, audit-stamped.
06
Score + Recommendation
300–850 behavioural score across four weighted families. Reason codes in plain English. Outcome recommendation to support the lender's decision.
The output
Line-by-line. Every transaction. Fully categorised.
Download sample output from a real extraction. Every transaction is there — date, description, amount, balance, category, confidence score, and counterparty.