AI-powered fraud detection for a Canadian fintech

The challenge

A challenger bank's rules-based fraud system was flagging 8% of legitimate transactions, forcing customers to call in to unlock cards. Support costs were rising and NPS was falling — while true fraud was still slipping through.

Our approach

• 01

  Built a labeled training set from 18 months of transaction and dispute data.

• 02

  Shipped a gradient-boosted model behind a feature flag alongside the existing rules — shadow-scoring for six weeks.

• 03

  Instrumented explainability so risk analysts could see why any transaction scored high.

• 04

  Progressive rollout with automatic fallback to rules if model drift exceeded thresholds.

Outcomes

• 40%

  fewer false positives
• 27%

  more true fraud caught
• $3.2M

  annual chargeback savings
• +18

  NPS points