A UK financial regulator sprint

BankingBank

Authorised payment fraud: designing for speed, signals and supervision

Experimenting with multi-agent fraud detection under tight sprint constraints.

AI & DecisioningEconomic Crime & Fraud

→

Participated in an APP fraud tech sprint, delivering the "Fastest" award-winning prototype

→

Coordinated multiple signals and tools through a single orchestration layer

→

Designed for audited explainability, not just detection performance

→

Demonstrated potential for reduced time-to-market for fraud controls

Request a detailed version of this case study →

Engineering Facets

Domain AlignmentGovernanceAuditabilityEvent DrivenModularity

Expertise

AI-Native Systems Event-Driven Architectures

Accelerators

Event Interaction Runtime Integrity Engineering Agent Orchestration Library Context Security Module

Capabilities

“Multi-agent architectures can improve fraud signal coverage without losing explainability, if the orchestration layer is designed for it from day one.”

Context

Authorised Push Payment fraud is a structurally hard problem: customers approve the payments, attackers move quickly, and regulators are increasing expectations around reimbursement and prevention. The client needed to explore new architectural patterns for fraud detection: multi-agent, multi-signal, and explainable to both internal and external stakeholders.

The Challenge

→
Combine diverse signals: transactional patterns, device data, behavioural features, external intelligence
→
Allow multiple specialised "fraud agents" to run in parallel while maintaining coherent outcomes
→
Ensure that any decision (block, step-up or allow) is explainable post-event
→
Prototype within days, not months, for a regulatory tech sprint setting

Our Approach

Multi-agent orchestration design

Each fraud detection capability (e.g. mule detection, destination risk, behavioural anomaly) modelled as an independent agent. An orchestration layer handled task allocation, result collation and decision aggregation.

Explainability and audit trail

For every payment, the system recorded which agents executed, what they saw, how they voted, and why. Designed to support downstream disputes, reimbursement assessments and regulatory reporting.

Pragmatic implementation

Built using cloud-native components and message-driven patterns. Delivered working prototype within tech-sprint timelines, while still following engineering discipline.

Engagement Timeline

Sprint Preparation

3 days

Designed multi-agent architecture, defined agent interfaces and orchestration protocol, set up cloud infrastructure.

Build & Integration

7 days

Built fraud detection agents (mule detection, destination risk, behavioural anomaly), orchestration layer, and explainability audit trail.

Demo & Evaluation

2 days

Demonstrated prototype to FCA/PSR panel, won "Fastest" award for prototype-to-production-grade delivery.

Outcomes

→
Demonstrated that multi-agent architectures can improve fraud signal coverage without losing explainability.
→
Proved that fraud controls can be iterated quickly when backed by a modular, event-driven architecture.
→
Provided the bank with a blueprint for future fraud platform investments and experiments.

Technologies & Patterns

Multi-agent orchestrationCloud-native componentsMessage-driven patternsExplainable AIReal-time event processing

Sibling case studies

Want the full story?

Client names, detailed architecture diagrams, metrics, and implementation insights are available on request. We're happy to walk you through the engagement.

Discuss a similar challenge