Fraud Detection

Detects fraud by scoring transactions, logins, calls, or payment signals in real time so teams can intervene before losses or bad approvals occur.

Business impact

• Fraud detection rate — Danske Bank reported a 60% increase in fraud detection.
• False positives — Danske Bank reported a 50% reduction, and Carousell reported a 75% reduction.
• Manual reviews — Dell said it reduced manual reviews of low-risk transactions by 83%.
• Fraud loss rate — Bank of America said AI models contributed to a reduction in fraud loss rate.
• Issuer authorization rate — Carousell reported a 10% increase.
• 3DS interventions — Carousell reported a 15% reduction.
• Latency — Danske Bank’s platform scored transactions in less than 300 milliseconds; Barclays reported sub-100ms performance at the 99.99th percentile.

Data requirements

• Login behaviour (tabular) — American Express used login behaviour to calibrate risk.
• Feedback from investigations (tabular) — American Express used feedback to calibrate risk scores, and Pindrop used investigation feedback to improve performance.
• Transaction data (tabular) — Emailage Adaptive continuously learns from each organization’s transaction data.
• Email, IP, phone, and address signals (tabular) — LexisNexis said its model analyzes these signals to score fraud risk.
• Caller voice (audio) — Bank of the West used voice analysis to assess caller authenticity.
• Call data (tabular) — Bank of the West used other call data alongside voice analysis.
• Device and behaviour signals (tabular) — Pindrop used device and behaviour analysis in its call-centre fraud detection.
• User personas and risk matrix inputs (tabular) — Carousell used personas and a risk assessment matrix to shape fraud interventions.

AI methods and techniques

• Machine learning — Used to score transactions, logins, calls, or payment signals and distinguish true fraud from legitimate activity.
• Risk scoring — Assigns a risk score that can trigger additional authentication, analyst review, or other intervention.
• Feedback-based calibration — Some systems use investigation outcomes to improve future scoring and adapt to changing fraud patterns.
• Entity resolution — Used in Carousell’s fraud process to connect related accounts and identify fraud rings.
• User-persona and risk assessment modelling — Used in Carousell’s process design to identify threat types and choose interventions.

AI models and model families

machine learning models, end-to-end machine-learning system, self-calibrating AI model

Industries

Real-world evidence

8 documented case studies on record.

Companies using this: AT&T, American Express, Bank of America, Barclays, Danske Bank, Bank of the West, Dell Technologies, Carousell.