Thalia Pillay | Co-founder | CEO | Orca Fraud | mail me |
The fines scam. The South African Revenue Service (SARS) refund. The parcel that needs clearing. If you have a South African cellphone number, you have likely encountered all three.
Fraud has become part of the background noise of digital life. Consequently, most people have developed a reflexive response: delete, ignore and move on. However, a more difficult question now confronts banks, fintechs and payment platforms.
Authentic vs manufactured communication
Can consumers still identify fraud using the warning signs they were taught to recognise? This question has become increasingly important as consumer fraud warnings continue appearing across financial platforms and communication channels.
Traditional warning signs such as awkward language, suspicious links and unusual requests no longer provide reliable indicators. Artificial Intelligence (AI) has made fraudulent communication increasingly difficult to distinguish from legitimate contact. As a result, the gap between authentic and manufactured communication is narrowing rapidly.
This Easter, Standard Bank, Absa, Nedbank and GoTyme all issued consumer fraud warnings. The advice they provided remains correct. However, the increasing frequency of these warnings deserves closer attention.
According to SABRIC’s 2024 Annual Crime Statistics, digital banking fraud incidents increased by 86% last year. This rise amounted to nearly 98,000 cases, with losses approaching R1.9 billion. Social engineering remains the primary driver because criminals manipulate customers into surrendering credentials instead of directly breaching banking systems.
Why social engineering is getting harder to stop
Fraudsters now use AI-generated phishing emails that are grammatically flawless, contextually accurate and carefully calibrated to match the tone of the institution they are impersonating. In addition, voice cloning technology can replicate a bank official convincingly enough to survive a real-time phone conversation. Deepfake video has also started appearing in higher-value fraud scenarios.
Furthermore, AI-assisted tools allow criminal syndicates to launch significantly more attacks simultaneously while reducing operational costs. Consequently, consumer fraud warnings now address a far more sophisticated threat landscape than in previous years.
Consumer education programmes have traditionally assumed that alert customers provide a meaningful line of defence against fraud. However, this assumption no longer remains reliable at scale. The growing volume of downstream fraud clearly demonstrates this reality.
Where legacy fraud tools fall short
Rules-based systems operate by identifying known patterns. For example, they flag transaction amounts above certain thresholds, unusual geographies or mismatched device fingerprints. These systems work effectively against fraud that behaves like conventional fraud.
AI-assisted social engineering deliberately avoids these patterns. When criminals manipulate customers into authorising transfers, the credentials remain legitimate, the session appears genuine, and the payment instruction seems valid. Consequently, the transaction passes through rules-based systems because criminals specifically designed it to do so. By the time customers report incidents, the money has already moved.
What real-time transaction monitoring does differently
Real-time transaction monitoring evaluates behaviour within a broader context. It assesses current activity against what appears normal for a specific customer, at a particular time, on a certain device, and within a defined payment corridor.
A transfer that clears the credential layer can still carry behavioural warning signs. For example, the amount may fall outside the customer’s typical transaction range. The destination account may have been registered recently. In some cases, the session may follow an unusual sequence of actions. Alternatively, the transaction timing may differ from the customer’s normal behaviour patterns.
Individually, none of these indicators provides definitive proof of fraud. However, when systems evaluate them collectively and in real time, they can alter the probability assessment enough to trigger a hold, require step-up verification, or flag the transaction for further review.
This represents the critical intervention point that legacy tools often miss. The opportunity does not arise before customers are deceived. Instead, it emerges before the fraud transaction fully completes. Consequently, adaptive models that learn from live behavioural patterns hold a meaningful advantage over static rules-based systems.
In conclusion
Institutions that contain fraud exposure most effectively will be those operating monitoring infrastructure that moves as quickly as the fraud itself. Therefore, successful intervention increasingly depends on acting before transactions settle. Stronger consumer fraud warnings, combined with adaptive monitoring systems, will remain essential as AI-driven fraud techniques continue evolving.
