Blog · All industries
Deepfake executives and AI phishing: updating your payment controls
In early 2024, a finance employee at the global engineering firm Arup joined a video call with the company's CFO and several colleagues. They discussed a confidential transaction. By the end of the day, the employee had sent 15 transfers totaling about US$25 million. Every other person on that call was a deepfake.
That case made headlines because of its size. The technique behind it is now cheap enough to use against much smaller businesses.
What changed
For years, the advice was to look for the tells: bad grammar, odd phrasing, a strange greeting. AI has removed most of them.
- Phishing emails are now well written, in your industry's vocabulary, and can be tailored to each recipient from what's public about them.
- Voices can be cloned from a short clip, such as a conference talk, a podcast or a voicemail greeting.
- Video can be faked convincingly enough to pass on a call, as Arup's employee found out.
The FBI's 2025 internet crime figures include more than $30 million in business email compromise losses with a confirmed AI connection. That's only the cases where AI was identified; the real number is higher.
Why “train people to spot it” isn't enough
Training still matters, but you can't train people to detect a fake that is indistinguishable from the real thing. The defense has to shift from spotting the fraud to making it not work, even when it's convincing.
Payment controls that still work
- Verify through a second channel you start. Any request to send money, change bank details or buy gift cards is confirmed by calling a number already on file. Never use the number, link or meeting invite in the request itself.
- Two people for every payment above a set amount. The second approver checks the request against the vendor file or contract, not against the email.
- Agree on a verbal passphrase between executives and finance staff for urgent requests. A clone can copy a voice; it can't know a phrase that was never written down.
- Treat urgency and secrecy as red flags. “Keep this confidential” and “it has to go out today” are the two pressures nearly every one of these frauds relies on. Make it policy that a confidential payment still gets verified.
- Lock down vendor changes. New or changed bank details for a vendor require a call-back and a waiting period before the first payment.
- Cut the attacker's research. Multi-factor sign-in and alerts on mailbox rules stop most attackers from reading your email, which is where they learn whom to impersonate and when.
The other side of AI risk: what your staff paste in
AI tools create a second, quieter problem. Staff paste client documents, financials or code into free public AI tools to save time, and that data may be stored or used outside your control. A short policy helps: which AI tools are approved, what data must never go into them, and a business-grade option (such as Microsoft Copilot under your company's Microsoft 365 account) so people aren't tempted to use personal accounts.
A five-minute check
Ask your finance team one question: “If the owner called you right now and asked you to wire $40,000 today, quietly, what would you do?” If the answer isn't “hang up and call them back at the number I have,” your controls depend on nobody ever being fooled.