Nefarious Use of AI Is Fuelling Fraud, But Human-in-the-Loop Remains Non-Negotiable for AI Application

In the age of Artificial Intelligence (AI), it would be easy to assume that fighting financial crime and fraud should be becoming easier. Attacks may be moving faster, but defensive tools are also more sophisticated and, in theory, risks should be contained. In practice, the opposite is happening. Today’s scams increasingly rely on manipulating people rather than breaching systems, creating a paradox for banks: technology for spotting fraud and compliance risks has never been better, yet losses and regulations continue to rise and pressure on teams is relentless.

In Singapore, reported scam and digital crime cases fell 21.5 per cent in the first half of 2025, to 22,476. On paper that sounds like progress, but total losses were still a hefty SGD $456.4 million—with most of the damage caused by social engineering-led investment (SGD $145.5M) and impersonation (SGD $126.5M) scams rather than technical system breaches. Those funds are often exfiltrated from the banking system through mule accounts and into crypto.

On top of that, while automated bot attacks dropped by 27 per cent, authorised push payment (APP) fraud, often the financial outcome of successful scams, rose 30 per cent last year. All this shows that attackers aren’t “just” trying to break systems at scale anymore. They have become better than ever at manipulating humans to get what they want.

How AI Has Changed the Nature of Financial Crime

Several forces sit behind this shift from high‑volume attacks to smarter, psychologically engineered schemes, and generative AI is near the top of the list

Technology is giving fraud new shapes. Synthetic identities used during account opening, deepfake impersonation used for identity validation, and AI‑powered social engineering used to initiate data compromise and APP fraud attempts are all on the rise. These threats are carried out in real time and across borders, evolving faster than rules, regulatory policies, and manual controls can adapt.

Just as fraudsters are using AI to get ahead, the finance industry is leaning on it too. Roughly nine in ten financial institutions globally have now adopted AI to help detect and respond to fraud, particularly in digital channels. Even as adoption accelerates, banks should maintain a human‑in‑the‑loop approach to ensure AI delivers its full potential safely and responsibly.

The Hidden Failure: Model-First Systems vs. Human Reality

As banks look to insert AI, fraud managers often find themselves stuck between competing anxieties: on one side, technology teams worry that data cannot be integrated or made available quickly enough; on the other, customer‑facing teams fear that overly aggressive controls will drive away customers through false positives. As a result, many fraud systems are still designed around relatively simple models and rule thresholds.

The real‑world consequences are easy to see: customer service teams struggle to explain or justify automated decisions, alerts pile up faster than investigative teams can handle, false positives frustrate customers, and compliance teams risk elevated regulatory scrutiny due to missing signals in a glut of unsubstantiated alerts. The fact is that systems can be technically accurate yet operationally fragile. That is more a failure of design than a gap in technology.

Regulatory pressure is also building towards the responsible use of AI. The Monetary Authority of Singapore (MAS) has been increasingly vocal about the need for explainability, accountability, and human oversight in AI‑driven decision‑making, particularly since issuing its consultation paper on AI risk management guidelines for the financial sector in November 2025. Supervisors are making it clear that “black box” approaches will not be sufficient in high‑stakes domains like fraud and financial crime.

Why Siloed Teams Make the Problem Worse

Even after years of investing in tools and technology, many fraud and risk leaders are still flying blind. Around two-thirds of fraud executives say they lack visibility into data breaches as a result of cyber security weaknesses, and nearly three-quarters say ongoing silos between fraud and cyber teams are their biggest worry. Bain similarly finds compliance remains fragmented, with weak communication across onboarding, due diligence and transaction-monitoring teams, driving delays, errors and false positives.

These silos persist because teams are often stuck with legacy systems built for different purposes, with limited native integration. As a result, they have no choice but to lean on manual data‑sharing or makeshift bridges between systems.

While siloed teams hold deep expertise in their own areas, they tend to crumble under pressure from multi‑faceted risk exposures. It is difficult for specialists working in isolation to thwart complex, fast‑evolving risks, especially as fraudsters quickly learn detection strategies and adapt their tactics to exploit blind spots.

Where Human-First Systems Are Enablers

In this environment, human judgment should not be seen as a bottleneck—especially now that social engineering, rather than system breaches, is causing the biggest losses. AI is astoundingly good at spotting patterns and anomalies, and its insights are most valuable when humans can provide context, interpret intent, and prioritise what really matters during the design and deployment of AI systems.

Fraud prevention is most effective when AI and humans with business context and risk expertise are able to work together throughout the lifecycle—from model design and training data selection to alert triage and case management. That requires systems built to incorporate expert input, not exclude it.

Instead of centering on any single solution, forward‑leaning institutions are experimenting with platforms that: surface clear, explainable insights rather than opaque scores; reduce noise for investigators by prioritising risk; and support workflows that bring fraud, cyber, and compliance teams together rather than keeping them in silos. In practice, this looks like:

• Fewer false positives for heightened customer confidence and less investigator fatigue.
• Processing that allows decisions to stay fast but also explainable.
• Stronger governance in support of audits, incidents, and regulatory reviews.

The ultimate payoff that comes when machine speed meets human accountability, of course, is confidence.  They simply won’t win through automation alone.

Winning the AI Fraud Race Requires Wisdom

Fraudsters are leading but banks are levelling up with AI. But for financial institutions, it’s not just a matter of “fighting fire with fire”—or in this case, AI with AI.  They have the responsibility of explainability and trust through strong control and governance.

In an environment where decisions are made at machine speed, institutions that retain human judgement, accountability, and control will ultimately be the ones that remain resilient.