.png)
A payments company processing $2.1 billion annually reduced fraud losses by 73% and false positives from 6.8% to 1.1% in 9 weeks.
.png)
A US payments infrastructure company processing $2.1 billion in annual transaction volume across 4,800 merchant accounts was running fraud detection through a 112-rule static rules engine and a six-person manual review team working a 48-hour queue. The rules engine was flagging 6.8% of all transactions — the majority legitimate — and was structurally blind to coordinated fraud ring activity. Confirmed fraud losses were running at $4.2 million annually, a figure the fraud team knew was understated.
Verttx replaced the rules engine with a two-component real-time detection system combining a gradient boosting transaction classifier and a graph neural network for fraud ring detection. The system went live across full production volume 9 weeks after the initial discovery call.
Of the 14,200 transactions flagged monthly, approximately 11,400 were false positives — valid transactions from legitimate cardholders incorrectly held by the rules engine. Three of the platform's top 20 merchants had cited false positive rates in offboarding conversations. The platform's largest merchant, a 340-location restaurant group generating $18.4 million annually, was actively evaluating alternatives.
The 48-hour review queue had become a systematic vulnerability. At 473 flagged transactions per reviewer per day, analysts had roughly 4 minutes per decision. False negative rates were 34% higher on Friday afternoons than Monday mornings — a pattern sophisticated fraud actors had begun to exploit by timing high-value transactions to coincide with reviewer fatigue. Two merchant accounts were approaching Visa's 1.0% chargeback monitoring threshold, above which the platform would face fines and potential programme placement.
After reviewing the transaction data environment and fraud pattern distribution, Verttx designed a two-component system. The first is a gradient boosting transaction classifier evaluating each transaction across 280 features — velocity signals, device fingerprints, merchant context, card and issuer signals, and cross-merchant behavioural patterns. The second is a graph neural network mapping relationship clusters across cards, devices, merchants, IP addresses, and email domains to surface coordinated ring fraud that individual transaction classifiers cannot detect.
Both models run in parallel, returning a fraud probability, decision recommendation, and plain-language reason code within 180 milliseconds end-to-end — well within the platform's 500-millisecond authorisation budget. A step-up authentication tier routes 2.1% of borderline transactions to SMS OTP verification rather than manual review, eliminating the queue for the vast majority of ambiguous cases. A circuit breaker falls back to a simplified real-time rules engine if either model service is unavailable.
The system ran in shadow mode for two weeks before cutover, processing 6.2 million transactions in parallel with the existing engine. The fraud team validated 400 sampled disagreement cases, confirming the new system was more accurate in 91.3% before approving cutover. Post-deployment monitoring tracks score distribution drift, false positive rates by merchant category, and ring detection recall — automated alerts trigger retraining reviews if any metric deviates more than 8% from the deployment baseline over a rolling 14-day window.
In the first full quarter of operation, confirmed fraud losses fell by 73% — from an annualised $4.2 million to $1.13 million. The graph network surfaced three coordinated fraud rings that had been completely invisible to the previous system, including one active for 14 months that had generated an estimated $340,000 in losses attributed to individual card compromises rather than recognised as organised activity.
The false positive rate fell from 6.8% to 0.9%, eliminating 12,310 unnecessary transaction holds per month. The analyst team was redeployed to fraud investigation and chargeback recovery, higher-value work that had been neglected under the previous operational model. Chargeback recovery rates improved by 41%. Both merchants approaching Visa's chargeback threshold came back into compliance within 60 days. Fraud operations cost per transaction fell from $0.34 to $0.04 - an 88% reduction.
The complete detection system - both model components, the graph database, all integration code, the monitoring pipeline, and the retraining infrastructure, was transferred in full to the platform's engineering team at handover with complete technical documentation and no ongoing dependency on Verttx.
The false positive problem was destroying merchant relationships we had spent years building. Verttx had a production system live in nine weeks. The fraud ring detection alone paid for the entire engagement in the first quarter. — VP of Risk and Compliance, US Payments Platform
.png)
Confirmed fraud losses fell by 73% in the first full quarter of operation, from an annualised $4.2 million to $1.13 million. The false positive rate dropped from 6.8% to 0.9%, eliminating 12,310 unnecessary transaction holds per month. Fraud operations cost per transaction fell 88%. Two merchant accounts approaching Visa's chargeback monitoring threshold came back into compliance within 60 days. The system went live across full production volume 9 weeks after the initial discovery call.
.png)
.png)
