Case Study Investment Banking

Detecting coordinated fraud rings invisible at sampled volume

A Tier-1 investment bank identified a coordinated fraud ring operating across 2,400 accounts using micro-transaction patterns that were individually below every threshold — patterns that only emerged when LLM reasoning was applied to 100% of transaction volume.

$47M

Fraud ring value identified and neutralized

18% → 100%

Real-time transaction coverage

2,400

Coordinated accounts identified in initial cluster

8 months

Duration fraud ring had been operating undetected

94 minutes

Time from full deployment to first cluster detection

1000×

Inference cost reduction vs prior LLM evaluation

The Organisation

Global Tier-1 Investment Bank · Investment Banking

The Challenge

The bank's fraud detection system was analyzing approximately 18% of transaction volume in real time, with the remainder batch-processed on a 24-hour delay. Sophisticated fraud actors had adapted their techniques to exploit this gap — using low-velocity, low-value transactions spread across thousands of accounts to stay beneath detection thresholds. The fraud team had noted unexplained losses accumulating over an 8-month period but had been unable to identify the source.

The Approach

LLM reasoning was applied to 100% of transaction events in real time. The model was configured to identify coordinated behavioral patterns across accounts — not just individual transaction anomalies — surfacing network-level fraud signatures that are invisible at sampled coverage.

"The individual transactions were unremarkable. Under $50 each, spread across thousands of accounts. The pattern only exists at full volume. At 18% coverage it was statistically invisible."

Head of Financial Crime Intelligence

Key Finding

The coordinated ring was operating a structured layering scheme — aggregating small values across thousands of accounts before consolidation. The LLM identified the network topology in 94 minutes after full coverage was established. The ring had been active for 8 months, accumulating $47M before detection.

Results at a Glance
Fraud ring value identified and neutralized $47M
Real-time transaction coverage 18% → 100%
Coordinated accounts identified in initial cluster 2,400
Duration fraud ring had been operating undetected 8 months
Time from full deployment to first cluster detection 94 minutes
Inference cost reduction vs prior LLM evaluation 1000×
Get in Touch

Talk to us about your data.

Tell us about your event stream and we'll show you what full LLM reasoning coverage looks like for your environment.

Or book a call directly →