Across card-not-present, APP fraud, account takeover, and synthetic identity attack vectors — with a false positive rate under 0.3%.
Year-one fraud losses dropped from £5.8M to £1.6M — a 72% reduction. Agentic decisions processed in under 180ms per transaction.
Cloud migration eliminated legacy data-centre contracts, reduced fraud operations headcount requirement by 60%, and cut infrastructure costs by £2.1M/year.
Discovery through live production — including full cloud migration of core banking data, FCA regulatory notification, and PSD2/Strong Customer Authentication compliance.
The exposure. The intelligence.
Rules-based fraud detection was losing the arms race with AI-powered fraud
The client — a UK challenger bank with 340,000 active customers — was operating a legacy rules-based fraud engine built in 2017. The ruleset comprised 2,400 static decision rules maintained manually by a team of six fraud analysts.
Fraud typology was evolving faster than the rules could be updated. APP fraud (authorised push payment) was costing £1.2M per quarter. Synthetic identity attacks were passing the rules completely. The false positive rate had reached 4.7% — causing significant customer friction and £380k in annual operational costs from manual review queues.
The cloud infrastructure picture was equally problematic: core banking data sat in an end-of-life on-premises data centre with a renewal cost of £1.8M, no disaster recovery SLA, and significant data residency compliance risk.
Multi-agent fraud AI with real-time graph analytics and sovereign cloud migration
TechProf designed a five-agent fraud detection system. Each agent specialises in a distinct fraud vector: APP fraud pattern recognition, synthetic identity detection, account takeover behavioural analysis, card-not-present transaction scoring, and network graph anomaly detection (identifying mule account rings in real time).
All five agents share a real-time feature store ingesting 140 transaction and behavioural signals per event. An orchestrator agent synthesises their confidence scores using an ensemble weighting model — delivering a single decision (allow, challenge, or block) within 180ms at transaction time. The model self-updates weekly against confirmed fraud ground truth without manual intervention.
In parallel, TechProf executed a full cloud migration of the core banking data estate to a UK-sovereign cloud architecture — meeting FCA data residency requirements, ISO 27001, and PSD2/SCA obligations. The migration was executed with zero customer-facing downtime via a phased blue-green deployment strategy.
Production-ready in
eight months.
3-year transaction history ingested, fraud typology analysis, FCA regulatory review, cloud migration feasibility, model training data pipeline design.
Five fraud agents trained on labelled transaction data. UK-sovereign cloud infrastructure provisioned. Data migration pipeline built and tested. Shadow detection running alongside legacy engine.
60-day parallel run. Agentic system identified 847 fraud cases the legacy engine missed. False positive rate tuned from 1.2% to 0.28%. FCA notified.
Blue-green data migration completed with zero customer downtime. Legacy fraud engine decommissioned. Core banking data fully migrated. Team trained and handed over.
"Our fraud losses dropped 72% in the first quarter. The old system required six analysts running reports daily. The new system runs itself and sends us a daily summary. TechProf delivered everything on scope, on time, and our FCA relationship has never been stronger."
How much are you losing
to fraud today?
Tell us your current detection accuracy and false positive rate — we'll show you exactly what an agentic AI system can recover for you.