Are Banks Ready for AI Voice Fraud?

AI-generated voice has changed the economics of fraud. What once required a skilled impersonator, a convincing script, or a compromised phone number can now be attempted at scale with tools that sound alarmingly human.

And what’s even more worrying, they’ve got the financial sector squarely in their sights. In early 2024, a finance worker at engineering firm Arup joined what appeared to be a routine video call with the company’s CFO and several colleagues. The voices were familiar, the faces recognizable, and the request appeared plausible enough to pass basic scrutiny. By the end of the call, she had authorized $25 million in transfers. The people she believed she was speaking to, including the company’s CFO, were later reported to have been AI-generated deepfakes. 

For financial institutions, the risk is not limited to internal payment approvals. The same collapse in trust can play out across customer authentication, call center escalation, fraud checks, and account recovery journeys. 

Scammers have exploited the relative anonymity of phone calls to commit fraud for decades. With AI, however, the effort required has now become trivial, and the results far more convincing. For banks and fintechs, that convergence is resulting in an uncharted map of risks.

There’s a more uncomfortable reality underneath this. Even as attackers probe for ways in, the AI infrastructure that financial institutions are deploying to modernize their customer support is making fraud easier to commit in plain sight. 

The call that sounds like it belongs

Before AI, voice fraud was relatively easy to identify. Robocalls sounded flat and followed a script, and vishing attempts relied on human operators following call sheets. In most cases, the scale was limited by the quality of the scammer attempting the social engineering. With adequate training and some common sense, staff could usually spot the signs.

 AI has changed that by nearly collapsing the technical barrier for impersonation. With enough prompting, AI can sound extremely human, complete with inflection, emotion and flaws that you’d expect from a real person. Making things worse, AI tools that can copy voices are readily available on commercial platforms and open-source repositories. 

McAfee Labs found that three seconds of audio is sufficient for an AI tool to produce a voice clone with 85% accuracy. Ten seconds of recording can push that figure above 95%. The source material is everywhere: social media posts, voicemail greetings, conference recordings, earnings calls, or LinkedIn videos. 

Real-time voice conversion tools became widely available in 2024, which means attackers today no longer need a pre-recorded clip. They can simply speak into a microphone, and the output will sound like whoever they’re trying to impersonate.

The Arup case is instructive: The finance worker had doubts, but the presence of familiar voices and faces on the call overrode them, but decided to trust her eyes and ears. That same year, Hong Kong investigators uncovered a separate operation that had cloned a financial manager’s voice to run an $18.5 million cryptocurrency scam. 

A familiar voice, it turns out, is a trusted one, and AI has made that assumption a dangerous one to harbor.

Authentication failures are becoming fraud opportunities

Financial institutions are already responding to external threats. Enterprise spending on governance and compliance tools for AI is forecast to grow from $2.2 billion in 2025 to $9.5 billion by 2035, a signal of how seriously the market is taking this.

 But the more consequential risk isn’t arising outside their walls. Banks and fintechs are layering AI into their customer support stack, building AI-assisted IVR systems, voice biometric authentication, and agentic call flows that handle transactions without human agents. These efforts are aimed at improving customer experience and reducing human labor, but when these systems fail, they leave a weak point that threat actors find all too easy to exploit.

Routing errors, context that’s lost between escalations, and authentication systems that behave unpredictably look like customer experience problems, and are logged as such. 

 But these CX problems can morph into something worse downstream. A customer who’s locked out by voice biometrics will rarely report it. Instead, the chances are high that they’ll call back, find a way to talk to a human agent, and pressure them into bypassing the very protocols those systems were built to enforce. 

The bigger issue, is an attacker impersonating a customer can do that, too. Financial institutions are quite aware of this. In fact, 91% of U.S. banks are reconsidering their voice biometric authentication strategies in light of AI cloning risks.

 Reconsideration will not resolve these problems. An institution whose voice authentication regularly fails legitimate customers has already created the very conditions a determined hacker needs to exploit the system. After all, a hacker doesn’t need to find a technical vulnerability when friction in the customer journey allows them a way in. 

Companies need to change their perspective, not their security stack

Tools that can detect AI-generated synthetic voices are improving, but a threat actor who’s probing a bank’s systems for vulnerabilities isn’t looking for a single point of failure. They’re looking for areas where systems stumble and miss each other, where authentication systems fail to pass signals clearly. Better detection at the perimeter can’t close that gap.

 What can help is reconsidering how leaders treat voice infrastructure. As with any part of your software stack, voice infrastructure requires the same level of scrutiny as perimeter defenses. 

In practice, such a security stance will require both functional and resilience tests across the conditions that IVR, biometric authentication, and other voice systems will face in production: low quality audio, different accents and background noises, edge-case escalations, and authentication boundary scenarios where legitimate callers are just outside the system’s acceptance thresholds. 

 The answer is not a one-time certification exercise. Banks need continuous validation of the full voice journey, from IVR routing and biometric authentication to escalation, transfer, and fallback paths. No matter how frequently voice systems are updated, fraud tactics will naturally evolve to find cracks in whatever defenses they encounter. A system that passes validation at deployment must be tested frequently and repeatedly in real world conditions to ensure it’s still secure even if its performance has evolved.

 It’s worth noting that continuous testing at scale will come with budget considerations. There’s also the inherent tension between moving quickly (deploying new capabilities to stay competitive) and validating systems thoroughly before they interact with real customers and face real fraud attempts. 

While it’s hard to resolve that tension cleanly, it does make clear the cost of skipping validation: A failure in your system that frustrated customers have already found their way around is a weak point, regardless of how your company categorizes it. 

Better customer journeys are becoming a fraud defense

The external and internal risks described above aren’t separate problems with different solutions. The friction produced by unreliable customer journeys is creating behavioral gaps that social engineering is designed to exploit. 

 That’s a solvable problem, but only if institutions realize that the impact of poor customer experience can reach farther than simply retention or revenue. Banks will not solve AI voice fraud by treating it as a detection problem alone. They also need to remove the ambiguity, friction, and failure points inside their own voice journeys. In an era where a familiar voice can no longer be trusted by default, the reliability of the journey itself becomes part of the security model.