TL;DR

• Feedzai has released RiskFM, which it describes as the first tabular foundation model purpose-built for financial crime detection across fraud, anti-money laundering, and broader risk decisions
• The model is trained on data from over £7.8 trillion ($9 trillion) in annual payment assessments spanning 120 billion events worldwide
• Lloyds Banking Group, one of the UK’s largest retail banks, is an early collaborator on the technology

A new approach to financial crime AI

Feedzai, a Portugal-headquartered fintech specialising in AI-driven financial crime prevention, has launched RiskFM (Risk Foundation Model) — a foundation model trained specifically on tabular financial data rather than text or images. The company says RiskFM can work across the full financial crime lifecycle, covering fraud detection, anti-money laundering, and risk decisioning in a single model.

The announcement represents a shift from how most financial institutions currently tackle crime prevention. Banks have traditionally relied on rules-based systems and bespoke machine learning models that are built individually for each client. RiskFM aims to replace that fragmented approach with a pre-trained model that can perform well from day one, without the manual feature engineering that typically delays deployment.

How it works

RiskFM draws on Feedzai’s position as a processor of large-scale financial data. The company annually assesses more than £7.8 trillion ($9 trillion) in payments across 120 billion events globally, covering onboarding checks, digital activity monitoring, payment processing, transfers, and AML workflows.

Pedro Bizarro, Feedzai’s Chief Science Officer, noted that financial risk presents a fundamentally different AI challenge to language processing. Transaction patterns shift constantly as consumer spending habits change, new payment methods emerge, and — most critically — fraudsters actively adapt their tactics to evade detection systems in real time.

Feedzai says early results show RiskFM can match the performance of highly tuned, client-specific supervised models even when working with a single customer’s data. When trained across multiple institutions and geographies, it outperforms them.

The UK connection

Lloyds Banking Group, which serves around 26 million customers across its Lloyds, Halifax, and Bank of Scotland brands, is named as a collaboration partner. Tom Martin, Business Platform Lead for Economic Crime Prevention at Lloyds, described RiskFM as “an exciting milestone” in the bank’s ongoing work with Feedzai to give fraud teams an advantage over criminals.

The partnership is significant in a UK context. Authorised push payment (APP) fraud remains a persistent problem for British banks, with UK Finance reporting hundreds of millions of pounds lost annually. The Payment Systems Regulator introduced mandatory reimbursement rules for APP fraud victims in 2024, increasing pressure on banks to improve detection. A foundation model that can identify suspicious patterns across institutions — rather than within a single bank’s data — could help address the cross-institutional blind spots that fraudsters exploit.

Foundation models have reshaped language and vision tasks, but their application to structured financial data remains relatively new territory. For UK banks under regulatory pressure to reduce fraud losses, a model that improves with cross-institutional data could address a persistent gap in the industry’s defences.