AI voice clones and deepfake calls: the new CEO fraud, and how to protect your team

The phone rings and it is your boss: same voice, a little rushed. There is a confidential acquisition, the lawyers need a transfer before five o’clock, and nobody else can know until it is announced. Except your boss is on a train with no signal, and the voice was generated by software from a video on your website.

That is deepfake CEO fraud: the old fake-boss payment scam, now run with an AI copy of a real person’s voice, sometimes with a fake face on a video call. What has changed is that the impersonation now sounds right. AI has made a lot of work faster, mine included (my guide to AI for small teams covers the useful side), and it has made fraud faster too.

The defence costs nothing. For the wider picture, see my cybersecurity guide for small teams.

How a voice clone is made, and what it needs from you

A voice clone is an AI model that reads any text, or converts a fraudster’s own speech, in a specific person’s voice. A deepfake video does the same with a face, live on a video call if needed.

The raw material is small. In March 2024, OpenAI wrote that its Voice Engine model needed “a single 15-second audio sample” to imitate a speaker closely, and it limited access to a few partners because of the risk of misuse (OpenAI’s post). Fifteen seconds is one answer in a podcast, a LinkedIn clip or a voicemail greeting. In December 2024, the FBI warned that criminals already use AI-generated voices to impersonate relatives in distress and to get into bank accounts.

The script comes from public information: your website, LinkedIn, company registers and press releases. A typical attack opens with a message on a plausible channel (email, WhatsApp, a meeting invitation) and a reason for speed and secrecy: a confidential acquisition, a non-disclosure agreement to sign, the market regulator “already informed”. Then a call, a voice note or a video meeting removes the last doubt, just before the request: a payment, new bank details, passwords or personal data.

Three real cases: Arup, WPP and Ferrari

These are three of the best documented cases. They involve large companies, but nothing in the method needs a large target.

Arup said none of its internal systems were compromised (report of Arup’s confirmation). The Guardian reported the WPP attempt from Mark Read’s email to staff, which listed warning signs such as requests for passports or money transfers and any mention of a “secret acquisition, transaction or payment that no one else knows about”. Bloomberg reported the Ferrari attempt; Ferrari declined to comment. In France, the DGSI, the domestic intelligence service, describes the same pattern: a site manager received a call, then a video conference, from someone with the chief executive’s face and voice, asking for a transfer linked to an acquisition (DGSI page, in French).

At Ferrari, one question ended the call, and WPP credited “the vigilance of our people, including the executive concerned”; at Arup, the check came after the money had gone. In all three, the clearest warning signs were in the request: a secret deal or venture, an unusual channel, and money or personal details at the end.

Why you cannot rely on spotting the fake

The Ferrari executive heard slight mechanical intonations, and France’s Interior Ministry suggests listening for missing hesitations or a tone without natural variation, and watching for lips out of sync on video. Useful cues, but not a defence: a bad line, a “camera problem” or a short voice note hides them, and you are listening to your boss, not hunting for errors.

Detection software exists, but a small team rarely has it running during a live call, and “probably genuine” is not a payment approval. Voice recognition is weaker too: in the same 2024 post, OpenAI encouraged phasing out voice-based authentication, for example to access bank accounts.

Treat these signs as reasons to verify, not as a test the call has to fail:

• a new number, account or channel
• urgency tied to a deadline you cannot check
• secrecy: “don’t tell anyone yet”, “this is under NDA”
• a reason why you cannot call back or ask someone else
• a request to skip the usual procedure “just this once”

Process beats technology: a rule applied to every request stops a perfect fake as surely as a clumsy one.

The rules that stop deepfake CEO fraud

None of these rules needs software or an IT department, and you do not need me to set them up. Write them on one page and apply them to every request, the owner’s included.

1. Hang up and call back. Any request for money, bank details, passwords or confidential files that arrives by phone, voice note, video call or chat is confirmed by calling back on the number saved in your contacts, never the number that called or one given in the message. A WhatsApp account with the boss’s name and photo proves nothing: the fraudsters who targeted WPP built one from a public picture. If you cannot reach the person, the request waits.
2. Ask a question only the real person can answer. The Ferrari question is the model: a detail from a recent conversation that appears nowhere online. A code word agreed in person works too; never agree it by email or chat, and change it once it has been used on a call.
3. Two people for any unusual payment. One prepares it, a second approves it after their own check. Your bank’s name check helps (Confirmation of Payee in the UK, Verification of Payee for euro transfers in the euro area since 9 October 2025), but a matching name does not make a request legitimate. My guide to invoice fraud and fake bank details covers the payment procedure.
4. Urgency plus secrecy means stop. The owner tells everyone, in writing, that they will never ask for a secret or urgent payment by phone, voice note or video, and that nobody will be blamed for checking, even when the request is genuine. That sentence removes the pressure the script relies on.
5. Brief everyone who can pay or open doors: assistants, your bookkeeper or outside accountant, anyone who can change bank details, reset passwords or send payroll or client files. Ten minutes at a team meeting is enough.

Voice clones are used on families too, with a “grandchild” in trouble asking for money: the FBI recommends that families agree a secret word or phrase in advance.

Make your team a harder target

You cannot stop people hearing your voice, but you can starve the script.

• Say less about who pays. Keep the name and direct line of whoever approves payments off public pages: French official guidance recommends it, because fraudsters study social networks and organisation charts first.
• Watch your out-of-office replies. “In Singapore until Friday; for urgent payments, contact Sam” tells a fraudster when the boss is unreachable and whom to call. The same guidance notes that these calls often come in the evening, during holidays or over a long weekend, when the accounts team is thin.
• Lock your mailboxes. A hacked mailbox gives a fraudster the whole script: who pays whom, the boss’s tone, the deals in progress. Switch on two-factor authentication or passkeys everywhere.
• Run one drill. The owner calls a colleague from an unknown number with a harmless urgent request and sees whether the call-back happens. Once, without blame.

If the call comes, or the money has already gone

On the call, stay calm and do not argue: ask your check question from rule 2, or simply say you will call back, then hang up and follow rule 1. Note the number, the time and the request, and screenshot the messages and the profile that sent them.

If money has left:

1. Call your bank’s fraud line at once, on the number from its website or your card, and ask it to stop or recall the payment. Minutes matter.
2. Report it. In England, Wales and Northern Ireland, use Report Fraud, online or on 0300 123 2040; in Scotland, call Police Scotland on 101. In France, file a complaint with the police or the gendarmerie; the 17Cyber service guides you and, for serious cases, connects you with an officer online.
3. Keep the evidence: call logs, numbers, messages, voice notes, invitations and bank confirmations.
4. Tell your insurer or broker the same day; my article on what cyber insurance covers explains what to check.
5. Check the accounts involved. If the fraudster knew things that only appear in your emails, treat the mailbox as hacked: my first-hour plan after a cyber attack covers the steps.

This is general information, not legal advice: when a large sum is lost or your bank will not help, a solicitor who handles banking disputes is worth the fee.

Set it up this week

About an hour of work:

• Write the call-back rule and the “urgency plus secrecy means stop” sentence on one page, signed by the owner.
• Agree a challenge question or code word, in person, with everyone who can approve payments.
• Switch on dual approval for new payees in your online banking, or make it a written rule.
• Brief everyone who can pay or reset access, new starters included.

When to hand it over: if you think a mailbox or account was used to prepare an attack, have it checked before you trust it again; my guide to free cybersecurity help in the UK and France lists good first stops. The rest stays yours, and no tool beats a team that calls back.