For years, the advice for spotting business email compromise included a reassuring fallback: if in doubt, call the person.A phone call was proof. You'd recognise the voice.
That fallback is breaking. AI voice cloning has become good enough, cheap enough, and easy enough that an attacker can place a call in a near-perfect imitation of your CEO's voice — or join a video meeting wearing their AI-generated face. The trusted verification channel has become another attack surface.
What Deepfake BEC Actually Is
Deepfake BEC isn't a brand-new category of attack. It's ordinary business email compromise — deceiving someone into moving money or data — with an AI upgrade. Where traditional BEC relied on a convincing email, deepfake BEC adds a convincing voice or videoto overcome the victim's natural urge to double-check.
The economics are alarming. Cloning a voice convincingly can take under a minute of source audio — and that audio is everywhere: a voicemail greeting, a conference talk, a podcast appearance, a webinar recording, a social-media video. Your executives' voices are, for practical purposes, public.
Publicly reported incidents have included multi-million-dollar losses where an employee joined what appeared to be a routine video call with senior leadership and authorised transfers — only every "executive" on the call was AI-generated. This is not a future threat. It is in active use now.
How a Deepfake BEC Attack Unfolds
Harvest
Attackers collect voice and video samples from public sources — voicemail greetings, conference talks, webinars, podcasts, social media. A minute of audio is often enough to clone a voice.
Research
They study the target organisation: who approves payments, who reports to whom, what a normal request looks like. Often a mailbox is already compromised, providing the inside view.
Stage
They build the scenario — a spoofed email to set context, then a follow-up call in a cloned executive voice, or an invitation to a video meeting populated with AI-generated likenesses.
Pressure
The request comes with urgency and authority: a confidential acquisition, an urgent supplier payment, a wire that 'must go out before the bank closes'. Pressure short-circuits scrutiny.
Extract
The employee, convinced by a familiar voice or face, approves the payment or shares the data. By the time the real executive is reached, the money has moved.
Why You Can't Defend This by "Spotting the Fake"
The instinct is to look for tells — unnatural blinking, audio glitches, lip-sync errors. Detection tools that hunt for these exist, and they have a role. But betting your defence on spotting the fake is a losing strategy: the technology improves every month, and a stressed employee on a quick call is in no position to run forensic analysis.
The durable defence doesn't try to detect the deepfake at all. It removes the deepfake's power by ensuring that a voice or a face is never sufficient authorisation for anything that matters.
The Controls That Still Work
Deepfakes defeat your senses. Process controls don't rely on your senses — so they keep working no matter how good the fake gets:
- Out-of-band verification: confirm any sensitive request via a separate, known channel — never the one the request arrived on
- Mandatory dual authorisation for payments and payment-detail changes above a set threshold
- A pre-agreed verification phrase or challenge question for urgent financial requests
- A firm policy that voice or video alone never authorises a payment or data change
- Advanced email security — most deepfake BEC still starts with a spoofed or compromised email
- MFA on every account to make the initiating mailbox compromise harder
- Staff training that specifically covers voice and video deepfakes, not just email phishing
- A no-blame culture so employees feel safe pausing to verify a 'senior' request
The cultural control matters as much as the technical ones. A deepfake attack works partly by exploiting an employee's reluctance to question a senior leader. If your team knows — because leadership has told them clearly — that pausing to verify a CEO's urgent request is not just allowed but expected, the attack's core leverage disappears.
The Bottom Line
Deepfake BEC turns the old advice — "just call to verify" — into a liability, because the call itself can be the attack. The answer isn't paranoia about every conversation. It's a simple, firm rule: money and sensitive data move on process, never on a voice or a face alone. Build that rule, train your team on it, and the most convincing deepfake in the world still can't get your money out the door.
Related reading: the full BEC guide, AI phishing and why scam emails got convincing, and social engineering tactics.
Frequently Asked Questions
What is deepfake business email compromise?
It's traditional BEC fraud — tricking someone into sending money or data — upgraded with AI-generated voice or video. Instead of just a spoofed email, the attacker adds a phone call in the cloned voice of an executive, or a video meeting with an AI-generated likeness, to make the fraudulent request far more convincing.
How do attackers clone a voice or face?
Voice cloning needs only a short sample — often under a minute — which attackers harvest from voicemail greetings, webinars, podcasts, or social video. Video deepfakes need more source material but are increasingly feasible from public footage. The tools are widely available and require little technical skill.
Has deepfake fraud actually cost businesses money?
Yes. Publicly reported cases include multi-million-dollar losses where employees joined what looked like a legitimate video call with executives and approved transfers — only the 'executives' were AI-generated. The technique is real and actively used, not hypothetical.
Can technology detect a deepfake call?
Detection tools exist but are imperfect and the deepfakes keep improving — it's an arms race. You cannot rely on spotting the fake. The durable defence is process: verification steps that don't depend on recognising a voice or face at all.
What's the best defence against deepfake BEC?
Process controls that remove trust in voice and video for sensitive actions. Out-of-band verification through a separate known channel, mandatory dual authorisation for payments and data changes, and a pre-agreed verification phrase for urgent requests. If the action can't proceed on a voice or video request alone, the deepfake fails.
Is Your Team Ready for Deepfake Fraud?
Free 30-minute assessment. We'll review your payment-verification process and staff training against AI-enabled social engineering.
Get Free Assessment