4 September 2025
by Payments Intelligence
What is this article about?
How financial firms are using AI for fraud detection while grappling with false positives, compliance pressures, and the need for transparency.
Why is it important?
Poor explainability and excessive false positives can damage trust, trigger regulatory action, and harm both customers and institutions.
What’s next?
Frms must embed explainability and compliance into AI model design to balance accuracy with customer confidence and regulatory requirements.
Fraud is both pervasive and evolving, costing the UK economy over £200 billion each year. Financial services firms have turned to artificial intelligence (AI) to keep pace with increasingly sophisticated scams, deploying machine learning (ML) models that can process transactions at scale and spot anomalies in real time. These tools promise faster detection, fewer losses, and greater resilience against cybercrime.
Yet the very strengths that make AI attractive also introduce new risks. False positives can erode customer confidence, opaque decision-making can leave institutions exposed to regulatory challenge, and overly rigid systems can damage relationships with merchants and individuals alike. As fraud prevention becomes a test case for high-stakes AI adoption, firms must balance detection accuracy with compliance obligations and the imperative of maintaining trust.
ML tools have become foundational to fraud detection across the financial sector. According to Thomson Reuters Institute, in 2025, 71% of financial services firms used AI for risk assessment and reporting, second only to document summarisation at 82%.
The movement towards this technology is driven by a host of factors: the volume of transactions; the speed of detection; the evolving nature of cybersecurity threats; cost effectiveness; and enhanced customer trust. It is necessitated by the ubiquity of fraud: the Global Anti-Scam Alliance estimated £1.03 trillion in losses to fraud in 2024, with close to half of the world encountering at least one scam per week. Payments Intelligence’s 2025 consumer behaviour survey found 37% of adults in the UK have been a victim of fraud.
Financial services firms are increasingly deploying ML models to address these threats. Yet, as models scale in complexity and reach, the consequences of false positives, opaque scores, and biases are both increasingly widespread and visible, putting payment providers at risk of reputational and regulatory fallout.
The data from early-movers in AI fraud detection appears positive. Companies are simultaneously reporting higher detection accuracy and a reduction in false positives; typically, an improvement in one of these factors leads to deterioration in the other. And there is interesting data on public perception of its deployment.
An EY survey of more than 1,000 people found that 63% were comfortable with AI being used for fraud protection and the detection of fraudulent activities. At the same time, just 31% were comfortable with the technology being used for the evaluation of claims and automation of decisions, such as insurance and fraud.
There is a legitimate difference between these two use cases, but they are similar enough that a 30 percentage-point difference is surprising. Undoubtedly, public understanding of AI provides some explanation for this discrepancy, offering insight into the importance of branding. Specifically, emphasising protection and detection of fraud is supported, whereas decision automation is not.
More common is the use of generative AI (GenAI), which are models that produce new content based on patterns learned from a large dataset of similar content.
While this category comprises the most famous models like ChatGPT and Gemini, it is important to note that the infrastructure that produces this technology is substantially different to that of an AI model used for fraud detection.
Fraud-detection models are far smaller and more specialised than the vast, general-purpose LLMs used for natural-language tasks. They are built for speed and efficiency rather than breadth. LLMs are trained on trillions of words of unstructured text, require enormous GPU clusters (i.e. vastly greater compute) as well as thousands of terabytes of storage, and are updated relatively infrequently.
Fraud-detection models, by contrast, are trained on structured transactional data — amounts, merchants, device IDs, and timestamps — with labelled outcomes, often retrained daily or weekly to catch emerging scams. They utilise lighter architectures, such as gradient-boosted trees or compact neural networks, which run in real-time payment systems with millisecond-level latency. These models are designed to be explainable for regulatory compliance purposes.
This is to say that AI is an immensely broad field and that progress in one area does not mean progress in another: the sophistication and wide use of GenAI does not mean the same as AI fraud risk models.
A striking example of the dangers of poorly designed automated systems, which carries lessons for AI fraud detection, comes not from financial services firms but the UK’s Department for Work and Pensions (DWP). The algorithm deployed by DWP, which was a rules-based review system, not AI, wrongly identified 200,000 people as being high-risk for housing benefit fraud over three years. Approximately two-thirds of claims flagged as high-risk by the algorithm were, in fact, legitimate, incurring a cost of around £4.4 million in unnecessary checks.
Closer to the financial services sector, the February 2024 Debanking Report by the All-Party Parliamentary Group on Fair Business Banking stated that banks are willing to offload customers who offer little or no profit, citing ‘economic crime’ for their decision, while tolerating more profitable customers who pose a higher risk. Consequently, certain ethnic, industrial, or political groups are being “frozen out” of the UK’s banking system, the group said.
Opaque “black‑box” AI models present material risks. With decisions inscrutable, institutions face challenges in accountability and governance. Regulators demand transparency. To give a specific example, the EU’s AI Act requires “high-risk” AI systems, such as those in credit scoring, hiring, and healthcare, to be transparent and understandable, requiring information on how they work, their capabilities, and their limitations.
In May of this year, the UK announced similar legislation for payment service providers (PSPs) regarding consumer rights and debanking. If passed by parliament: the minimum written notice before terminating an account will be extended from 60 to 90 days; PSPs will have to provide specific reasons for the closure (i.e. “commercial decision” or “risk appetite” will no longer be adequate); PSPs will have to inform customers of their right to complain to the Financial Ombudsman Service (FOS).
Explainability is becoming an increasingly necessary requirement for regulatory compliance. It can also be a competitive edge. Offering customers detailed explanations as to why a payment was flagged or declined not only fosters trust in the institution’s processes but also enables both the customer and PSP to avoid further disrupted transactions. Clear rationales also reduce call-centre demand by resolving queries at first contact, lowering operational costs.
For PSPs, transparency can shorten onboarding times, streamline dispute resolution with partners, and help secure strategic relationships with merchants who need assurance that legitimate transactions will not be wrongly blocked. Explainability can not only be a regulatory safeguard but also a driver of efficiency, customer retention, and reputational strength.
The tension between technological development and regulatory oversight plays out both between companies and governments, as well as within companies themselves, between the teams developing the ML systems and the compliance teams ensuring regulatory alignment.
Fraud detection models are the product of ML teams focusing on accuracy, throughput, and innovation, whilst compliance and governance units emphasise transparency, auditability, and fairness. This is, of course, a simplification, but the principle of these competing tenets – model improvement and regulatory compliance – holds true.
Leading PSPs are addressing this by embedding “compliance by design” into model development. This means integrating governance controls, fairness testing, and explainability tools (such as SHAP or LIME) into the build process rather than applying them retrospectively. Formal joint working groups between ML engineers, compliance officers, and product managers help align objectives and avoid late-stage rework.
Regulators are also signalling that such cross-functional governance will become an expectation, not an option. For senior leaders, these shifts elevate explainability from a technical afterthought to a core element of operational risk management – one that directly affects market share, regulatory relationships, and customer trust. Ultimately, the firms that succeed will be those that balance detection accuracy with transparency and fairness, protecting both compliance and customer confidence.
New research shows vulnerable customers are strong adopters of AI and digital banking, but are far more likely to experience failed payment journeys and poorer outcomes.
UK merchants expect agentic commerce to grow rapidly, but uncertainty around liability, fraud, and standards is slowing readiness.
Stablecoins are moving into mainstream finance, reshaping payments, trade, and regulation as institutions explore faster, programmable settlement.
You need to be logged in to do this!
