Digital businesses increasingly face the challenge of AI-assisted fraud, necessitating tighter controls. The rise of synthetic identities and generated documents complicates traditional verification methods, while automated attacks enable organized groups to exploit vulnerabilities rapidly.
Know Your Customer (KYC) checks are particularly vulnerable, as skilled fraudsters can use AI to impersonate legitimate users. According to Volodymyr Todurov, CEO of the fraud prevention and risk management company Frogo, "Currently, KYC is the most targeted layer, where real-time document generation and deepfake face-swapping are becoming remarkably sophisticated."
However, implementing stringent controls can have its drawbacks. While aggressive measures may prevent fraudulent transactions, they can also lead to legitimate payments being rejected or delayed, prompting customers to seek alternatives. A business might see its fraud rate decrease but simultaneously lose valuable clientele.
Frogo, which combines device intelligence, automated risk scoring, and real-time alerts, emphasizes the need for businesses to rethink their approach to fraud prevention. "Operating with a ‘guilty until proven innocent’ mindset and relying on aggressive, automatic blocks," warns Todurov, "treating legitimate players like suspects completely destroys the customer experience."
This perspective invites a broader evaluation of anti-fraud systems, focusing on profitability by considering not just prevented losses, but also operating costs and the impact on legitimate users.
To accurately gauge commercial impacts, Todurov describes anti-fraud measures as a "revenue optimization engine," using three interconnected metrics: direct loss prevention, operational efficiency, and customer retention.
The first metric, direct loss prevention, measures financial losses due to fraud, including fraudulent withdrawals. While reducing losses is important, the resources involved must also be evaluated. If regular checks require constant manual monitoring, businesses may struggle to handle increasing customer volumes without continually expanding their risk teams.
"Without a nuanced evaluation model, you risk blocking your highest-revenue players," says Todurov.
Operational efficiency assesses how many active users each analyst can support and the duration of investigations. Automating simple decisions permits specialists to focus on more complex threats. The third metric examines the experiences of legitimate customers. Todurov advocates for "automated withdrawals built on the presumption of innocence," asserting that smoother transactions foster greater trust and long-term value. Delays or unnecessary checks can incur costs, even if they do not appear in fraud statistics.
All three metrics should be evaluated collectively. A system can correctly detect fraud while overlooking significant threats or curtail losses through methods that deter legitimate clientele. Understanding both the limitations of the detection system and the costs of interventions is critical to navigating this balance.
The issues with aggressive controls are evident when considering high-value customers, who often display behaviors that diverge from average user profiles. Frequent travel may cause sudden location changes, while significant deposits can create quick surges in activity, and the use of multiple payment cards can further complicate detection. Such patterns may resemble fraudulent actions, including attempts to test stolen cards, leading to potentially costly mistakes for businesses.
"Without a nuanced evaluation model, you risk blocking your highest-revenue players," emphasizes Todurov. Frogo asserts that unusual behaviors should trigger informed evaluations, allowing support or retention teams to respond appropriately instead of initiating immediate blocks.
The expertise needed to discern these distinctions is essential. Specialists must understand both the product and its users to evaluate which signals merit intervention, determining where AI can enhance decision-making versus where traditional methods retain superiority. "AI isn’t a silver bullet," Todurov cautions. "Its success depends on specialists who know precisely where machine learning adds value and where tried-and-true, deterministic logic is still superior."
Equally important is obtaining the right information from the outset. Frogo highlights the importance of context in behavioral assessment, advocating for the collection of user data starting from the initial visit. This begins with device fingerprinting, where hardware and software parameters are analyzed to verify user legitimacy and detect potential bots or spoofed devices.
Frogo's standalone Device Fingerprinting tool allows businesses to integrate this capability within their existing anti-fraud setups independently. This tool is particularly beneficial for enterprises with basic in-house rules but limited abilities to recognize devices. The tool evaluates over 300 client-side signals to facilitate ongoing assessment of user legitimacy.
This information can also help in identifying genuine customer behavior, as well as aiding in fraud detection. According to Todurov, "Anti-fraud tools should be viewed as core data engines. Instead of only scoring bad behavior, you can define triggers that identify high-value, ‘ideal’ player behavior."
Frogo’s real-time alerts manage how different teams within an organization respond to user activity, allowing each department to address the same underlying activity according to their tasks. As businesses expand into new markets, they may find that customer behaviors and methods of payment differ, necessitating adjustments to existing rules.
The goal is to address real threats, minimize unnecessary work, and ensure seamless interactions for legitimate customers as operations grow. Todurov believes that the future of successful fraud management lies in combining a highly skilled fraud detection team with agile technology. "The winners will combine a highly skilled fraud team with agile technology, viewing anti-fraud through a growth mindset," he concludes.
