Three seconds of audio — a voicemail greeting, a snippet from a webinar, a clip pulled off a company's own YouTube channel — is enough to clone a voice convincingly today. The tooling to do it costs under $20 on dark web markets. On the other side of that transaction: business email compromise attacks that incorporate an AI-generated deepfake now cost victims an average of $4.1 million per incident, roughly three times the $1.3 million average for a traditional, non-AI BEC attempt. The attacker's cost went down. The victim's cost, when it works, went up — by a lot.
From a rare tactic to a large minority of BEC in under three years
As of early 2026, an estimated 40% of business email compromise attacks now involve some form of AI-generated deepfake — audio, video, or both — up from under 5% in 2023. That's a widely cited industry estimate rather than a single official count, and it's worth treating as directionally right rather than precise to the decimal. Directionally, it lines up with everything else in this space: US BEC losses reported to the FBI's Internet Crime Complaint Center hit $2.9 billion in 2024 alone, widely acknowledged as an undercount, and the Bureau's most recent annual report tracks AI-assisted fraud as its own category for the first time — a tacit admission that "is this AI-generated" is now a routine question in a fraud investigation, not an edge case.
Why the AI version costs three times as much per incident
Scale and confidence, not just novelty. A cloned voice or a synthesized video call lets an attacker credibly impersonate a specific, senior person rather than send a generic request — which raises both the plausible transfer amount and the victim's willingness to skip a second look. It also buys time: the average window between an attacker first triggering suspicion and funds actually leaving the account is around 18 hours. That's not "instant." It's a real, if narrow, window where a verification step that doesn't rely on how convincing the request sounded can still catch it before the money moves.
Detection has a ceiling. Verification doesn't need one.
This is the part worth sitting with: even purpose-built deepfake detection tools top out around 80–85% accuracy under clean lab conditions — and drop to somewhere between 50% and 65% against real phone call audio, with background noise, compression, and a nervous employee on the other end. Betting your payment security on "we'll be able to tell it's fake" is betting against a coin flip that's only getting harder to win as the tooling improves. That's exactly why detection was never the right layer to defend this at.
The fix doesn't care how good the fake is
A verification check that confirms payee name, account number, and routing details against a previously verified record — through a channel independent of wherever the request arrived — doesn't need to judge whether a voice or video was real. It only needs to know whether the destination account matches what was verified before. A perfect deepfake, played to a distracted employee on a completely legitimate-looking call, still fails that check the moment the payee details don't match — because the check was never evaluating how convincing the call was in the first place.
CyberFence's Fraud & Payment Verification engine works on exactly that principle and remains in private pilot as we test it against real cases before wider release. If AI-augmented payment fraud is on your risk list, ask about early access.
Sources: industry reporting on AI-augmented BEC prevalence and loss multipliers, 2026; FBI Internet Crime Complaint Center (IC3) 2024 BEC loss data; published deepfake-detection accuracy benchmarks, 2026.