\
Insight
\
Gartner® names Reality Defender a Market Shaper in deepfake detection, June 2026
Get the report\
Insight
\
Gabe Regan
VP of Human Engagement
Published January 17, 2025 Updated November 5, 2026
To detect a deepfake voice call, run deepfake detection on the live call audio while the conversation is happening. Detection listens for synthesis artifacts. The result comes back within seconds, so the agent can step up verification before approving a transfer, a password reset, or an account change.
Vishing used to depend on a convincing script. Now the voice is convincing too. Microsoft Research demonstrated voice cloning from a three-second audio sample. In a Gartner survey of 302 cybersecurity leaders published September 2, 2025, 43% of organizations reported a deepfake audio-call incident in the past 12 months, and 37% reported an incident involving video.
Banks, insurers, and their contact centers sit in the middle of this shift, because so many high-value decisions still happen on a call.
Voice phishing, or vishing, moves the phishing con from the inbox to the phone. An attacker calls, impersonates someone the target trusts, and asks for account access, data, or a payment.
Deepfake technology changes who the attacker can sound like. A cloned voice lets a caller pass as a customer, an executive, or a colleague, and it removes the accent, hesitation, or mismatch that used to give a fraudster away.
Regulators have taken notice, and in the U.S. the Federal Communications Commission (FCC) enforces the Telephone Consumer Protection Act (TCPA). The FCC ruled in February 2024 that AI-generated voices in robocalls are illegal under the TCPA.
Impersonation works because people trust a familiar voice or face. Engineering firm Arup lost $25M when an employee joined a video call with deepfaked versions of the CFO and colleagues and made 15 transfers. A phone call leaves less to check.
Attackers borrow authority too. In 2025, an impostor used AI-generated voice and text messages to impersonate Marco Rubio and contact senior foreign and U.S. officials. If a cloned voice can reach senior government officials, it can reach a branch manager or a wire desk.
Listening carefully is a weak defense. In a 2025 spearphishing experiment, 66% of participants failed to identify AI-created audio and 43% failed to identify AI-created video as fake. A 2026 vishing study found participants averaged 37.5% accuracy, with 75% of AI-generated clips majority-labeled as human. Neither study tested contact center agents, but neither gives a reason to trust the ear on a high-risk call.
Voice authentication has the same gap. A 2025 study found that face-to-voice attacks can generate synthetic voices from a single facial image and bypass industry-standard voice-authentication systems.
Financial services still run high-value decisions over the phone, including account recovery, payee changes, wire approvals, and limit increases. Each one is a moment where a convincing voice can move money.
The U.S. Financial Crimes Enforcement Network (FinCEN) has put banks on notice. FinCEN warned financial institutions in November 2024 about fraud using deepfake media and AI-altered identity documents. That alert covers more than voice. For what it means for bank fraud teams and the contact centers they rely on, see our breakdown of the FinCEN deepfake fraud alert.
To put a number on a successful attack, see the real cost of voice fraud in banks.
No single control stops deepfake vishing. Four work together.
Deepfake detection scores call audio in parallel with the conversation and flags likely synthetic speech within seconds, early enough to pause a high-risk action. RealCall does this for contact centers. A global bank analyzed over 2.4M customer calls in 2025 with Reality Defender and detected ~1,010 synthetic voice calls. For the architecture, see how real-time voice deepfake detection works in call centers.
A voiceprint match no longer proves a human is speaking. For high-value requests, add a callback to a number on file, a second approver, or a step-up check in another channel.
Since the ear is a weak control, train staff to follow the verification process every time a request is high-risk, however familiar the caller sounds. A clear rule beats a good ear.
Decide in advance who an agent escalates a flagged call to, how the team protects the account, and how the institution warns the customer and peers.
Run deepfake detection on the live call audio while the conversation is happening. It looks for signs of voice synthesis and returns a result within seconds, so the agent can step up verification before approving anything.
Not reliably in testing. A 2026 vishing study found participants averaged 37.5% accuracy, with 75% of AI-generated clips majority-labeled as human. That’s why training should teach staff to verify every high-risk request rather than judge the voice.
Not on their own. Voice biometrics confirm that a voice matches an enrolled customer, and a good clone can match. Deepfake detection checks something else, whether the voice is synthetic. Run both.
Pause the request and move to a verification step that doesn’t depend on the voice, such as a callback to the number on file. Then route the call to the fraud team under the institution’s response plan.
Voice will stay a core banking channel, so cloned voices will keep reaching the line. Detection after the money moves is a post-mortem. Verification has to come first. Treat every high-risk call as a decision to verify, and give agents a signal they can act on before anything is approved. For the wider picture, read 5 ways voice AI detection is reshaping financial security and Reality Defender’s approach to contact center security
See how deepfake detection runs on your own contact center calls.
Request a RealCall demo