How AI Is Transforming Fraud Detection: Techniques, Tools, and Use Cases

Fraud is a technological arms race, and the teams building with AI right now are the ones setting the terms.

Across Asia in 2024, consumers lost an estimated $688 billion to scams (Global Anti-Scam Alliance, 2024). Singapore alone recorded a record S$1.1 billion in scam losses across 51,501 cases, a 70.6% surge from the year prior (Singapore Police Force, 2025). Globally, illicit financial flows have now crossed $4.4 trillion annually (Nasdaq Verafin, 2026 Global Financial Crime Report). That is a number so large it almost stops meaning anything, until you think about what is on the other side of it: real businesses, real customers, and real product teams scrambling to keep up.

The good news is that AI-powered fraud detection is delivering measurable results. HSBC’s system, built in partnership with Google Cloud, now catches two to four times more suspicious activity than its old rule-based approach, while simultaneously cutting false positives by 60% (Google Cloud Blog). Mastercard’s generative AI model, trained on roughly 125 billion transactions annually, improved average fraud detection rates by 20%, with some categories seeing gains of up to 300% (CNBC, 2024).

By the Numbers: The Global Fraud Landscape

Why Rule-Based Systems Are Losing the Fight

For years, fraud detection looked like a spreadsheet problem. Analysts would write rules: flag any transaction over a certain amount, from a new device, in an unfamiliar country. And for a while, it worked reasonably well. The problem is that fraudsters read the playbook too. Once they understand the rules, they learn to work around them, staying just below thresholds, using local networks, and mimicking legitimate behaviour. Traditional rule-based systems now generate false positive rates of 90 to 95% in practice, meaning that for every genuinely suspicious transaction flagged, many perfectly legitimate ones are blocked alongside it. Every one of those is a customer who cannot complete their payment and a business relationship placed under unnecessary friction.

AI changes the dynamic in a fundamental way. Instead of being told what fraud looks like, a model learns from the data what normal looks like, and then surfaces the deviations. It does this at a scale and speed no analyst team can match. More importantly, it keeps learning.

The Core Techniques

Understanding the AI landscape in fraud detection does not require a data science background. What it does require is knowing which tools address which problems, so you can make informed decisions about what to build or buy.

• Supervised Learning Models like XGBoost are trained on labelled transaction history. They excel at known fraud patterns and are the most common production starting point, achieving precision above 91% on standard benchmarks (ScienceDirect, 2025).

• Anomaly Detection Unsupervised methods like autoencoders and isolation forests need no labelled fraud data. They learn what normal looks like and flag deviations, making them essential for catching novel attack types before enough cases exist to train on.

• Graph Neural Networks GNNs map relationships between accounts, devices, and transactions. Even clean-looking accounts get flagged when connected to known bad actors. Alibaba’s eFraudCom used this to expose fraud rings that transaction-level models missed entirely.

• NLP and Generative AI NLP analyses unstructured text like claim descriptions and chat transcripts for suspicious patterns. LLMs now help analysts generate investigation summaries and automate Suspicious Activity Reports, cutting compliance workloads by up to 40%.

Running all of this in real time requires streaming infrastructure most people outside engineering never see: Apache Kafka for event streaming, Apache Flink for stateful processing, and Redis for low-latency feature retrieval. GrabDefence, Grab’s internally-built fraud platform, runs on exactly this architecture and brought its fraud rate down to 0.2%, well below the regional average (Grab, 2022).

Tools That Teams Are Using

The tooling landscape has matured significantly. There is now a meaningful difference between what early-stage startups can access and what enterprise-grade teams are building.

• Stripe Radar: The easiest starting point for startups already on Stripe. ML trained on billions of transactions, zero additional integration, with custom rule layers for growing teams.

• Sardine: Built for fintechs. Layers device intelligence with behavioural biometrics across 4,800 risk attributes. Raised a $70M Series C in 2025 backed by Andreessen Horowitz.

• Unit21: Agentic platform handling the full fraud lifecycle from detection to SAR filing. Claims a 15% false positive rate versus the industry norm of 90 to 95%.

• Feedzai: Enterprise-grade RiskOps with real-time scoring and network analysis. Trusted by Citi, Visa, and PayPal.

• SEON: API-first, pay-per-check pricing suited to early-stage teams. Analyses 900 digital footprint signals per user.

What Is Happening in Asia and Singapore

Southeast Asia’s rapid mobile-first financial adoption has created a large and relatively young attack surface. A 2022 Forrester survey found seven in ten businesses in the region had experienced online fraud in the preceding year. The regulatory response has been fast. MAS’s FEAT Principles set the framework for responsible AI use in financial services and allow fraud models to withhold decision logic, precisely to prevent adversarial gaming. In November 2025, MAS issued draft Guidelines on AI Risk Management requiring institutions to document why they chose complex models over simpler alternatives. Singapore also enacted the Protection from Scams Act in July 2025.

On the ground, SPF’s Anti-Scam Centre recovered S$182 million and averted S$483 million in potential losses in 2024 alone (SPF, 2025). The 2025 data marked the first-ever annual decline in scam cases, meaningful evidence that AI tooling and inter-agency coordination are starting to shift the curve. At the same time, the threat is escalating in new ways. Singapore registered a 207% year-on-year spike in identity fraud in 2024. A finance director was nearly defrauded of US$499,000 after joining a Zoom call where executives had been replaced by deepfakes. SWIFT and Google Cloud also launched a federated learning initiative with major global banks to detect cross-border fraud without sharing raw customer data, a structural breakthrough Singapore is well-positioned to participate in (SWIFT, 2025).

The Arms Race

The same technology powering fraud detection is now being used to commit it. Deepfake production costs have fallen to under US$10 per clip. GenAI-enabled scam reports rose 456% year-on-year between May 2024 and April 2025 (TRM Labs), and 87 deepfake scam rings were dismantled across Asia in Q1 2025 alone. The World Economic Forum’s Global Cybersecurity Outlook 2026 found 72% of organisations reported increased cyber risk attributable to generative AI. What this means for product teams is straightforward: a fraud model deployed today will degrade. One European insurer saw detection precision fall from 78% to 51% within 18 months without retraining. The teams winning this fight treat fraud detection not as a feature to ship once, but as an ongoing capability requiring continuous retraining and investment in explainability, because regulators and customers now both demand it.

How DevLabs by AngelHack Can Help You Build It

Whether you are a fintech startup designing your first fraud pipeline or a CTO weighing build versus buy, DevLabs by AngelHack brings the technical depth to help you move faster. We build end-to-end AI solutions that are practical, scalable, and built around your specific context.

• AI Development: ML models, data pipelines, and AI-driven fraud solutions built into your product.
• Full Stack & Mobile: Web and mobile applications with modern back-end systems and built-in fraud signals.
• Blockchain & Web: Decentralised and scalable web systems enhancing transaction transparency and auditability.

Conclusion

Fraud is not going away, and the nature of it is changing faster than any single team can manually track. AI brings a meaningful shift in leverage: processing millions of signals simultaneously, adapting to new patterns through retraining, and surfacing connections invisible to human analysts. Singapore’s first recorded annual decline in scam cases in 2025 did not happen because fraudsters ran out of ideas. It happened because AI-powered systems, cross-agency coordination, and regulatory clarity started moving faster than the attack surface was expanding. For startups and product teams entering this space now, the tooling is mature and the case is clear. The time to start is before fraud becomes a problem, not after.

Frequently Asked Questions