Report

AI Feature Teardown Report

Carbon · Mono · Copilot Money

A product-led teardown of AI/ML systems in three fintech products, focused on how each system makes decisions, where it breaks down, and what that implies for product and governance design.

Carbon — Credit scoring

Analysed how Carbon's underwriting model uses alternative data (transaction history, device signals) to score thin-file borrowers, and where that scoring breaks down for informal-sector income.

Mono — Transaction enrichment

Looked at how Mono classifies and enriches raw bank transaction data for downstream fintech consumers, and the tradeoffs between enrichment accuracy and latency at API scale.

Copilot Money — Personalised categorisation

Reviewed the on-device/cloud split for transaction categorisation and how the product handles category correction feedback loops.

Five Cross-Cutting Patterns

1. Alternative data sources substitute for missing formal credit history, with real accuracy tradeoffs.
2. Enrichment and categorisation systems all rely on a correction feedback loop to improve over time.
3. None of the three expose their AI decision logic to end users — a governance gap, not just a UX one.
4. Latency constraints at API scale push teams toward simpler, faster models over marginally more accurate ones.
5. All three treat model output as advisory by default, with human or user override built in.