For years, the advice on CEO fraud was to look closely at the email: check the sender address, notice the odd phrasing, question the urgency. That advice still matters, but it no longer covers the whole problem. In January 2026 the Swiss National Cyber Security Centre (NCSC) reported that a company in the canton of Schwyz had lost several million francs after fraudsters used deepfake audio to imitate a manager or business partner. The voice on the phone sounded right. The request was still fraudulent.

This article is about what that changes for a small or medium-sized business, and what it does not.

What has actually changed

The NCSC describes CEO fraud as one of the fraud types most often reported to it, with reported cases rising from 719 to 971 in its latest annual comparison. Three developments stand out in its account.

First, the research behind these attacks is careful. Fraudsters study LinkedIn profiles, company websites and the commercial register to learn who approves payments, who reports to whom, and who is on holiday. They register lookalike domains that differ from the real one by a letter.

Second, the channel has moved. Email is still common, but the NCSC notes that attempts increasingly arrive by WhatsApp and phone, and that AI tools are used to copy a real person's writing style, down to their usual greeting.

Third, the voice itself can now be faked. Euronews, citing researchers at Charles Sturt University, reported that a usable voice clone can be built from roughly three seconds of audio, the kind of clip many managers have already posted in a conference video or social media post. The same piece notes that caller ID can be spoofed as well.

The NCSC is measured about the more cinematic version of this threat. It has seen deepfake video conferences, but considers them too complex for widespread use at present. The practical risk today is the phone call and the voice message.

Why "learn to spot the fake" is the wrong goal

It is tempting to respond with awareness training focused on detecting synthetic audio. That is unlikely to be enough. In an EU-wide Eurobarometer survey reported by Euronews in October 2026, fewer than half of workers said they could recognise a deepfake video, and that is people's own estimate of their ability. Audio quality will keep improving. A control that depends on an employee's ear is a control that will eventually fail.

There is good news on the network side. Since July 2026, Swiss providers have had to identify calls from abroad that spoof Swiss numbers, and reports of fake-authority calls fell by more than 75 per cent that month, according to Swissinfo. That helps, but it does not reach a WhatsApp voice note or a call that arrives on a legitimate line. The defence that holds up is a process that does not care how convincing the voice is.

The payment controls that hold up

The NCSC's own recommendations are deliberately unglamorous, and that is their strength. Translated into a checklist for an SME owner:

  • Two people for every outgoing payment above a threshold you set, and for every change to supplier master data, especially a new IBAN. No single phone call, however senior the voice, can move money on its own.
  • Call back on a number you already hold. If a request is urgent or confidential, end the conversation and phone the person on the number in your own directory, never the number supplied in the message or call.
  • No exceptions for management. Write down that the process applies to the owner and the board as well. Fraudsters rely on staff being afraid to slow down a senior person.
  • Treat "confidential" as a warning sign. Requests to keep a payment secret from colleagues, or to deal with an unknown lawyer about a sensitive deal, deserve more scrutiny, not less. The NCSC reports that fraudsters use the names of real Swiss law firms in exactly this way.
  • Flag external email. Configure your mail system to mark messages from outside the organisation, so a spoofed internal address stands out.
  • Agree a verification phrase for the few people who authorise payments. It costs nothing, and a cloned voice cannot produce it.
  • Review what you publish. You do not need to hide your team, but consider whether holiday announcements and detailed reporting lines need to be public.

Culture is part of the control

The hardest part is not technical. Every one of these controls asks a junior employee to slow down a request that appears to come from someone senior. The NCSC puts it well: "Healthy scepticism should be recognised as a strength." In practice, that means the owner saying, explicitly and more than once, that a staff member who delays a payment to verify it will be thanked, even if the request turns out to be genuine.

It is also worth keeping perspective. Research reported by Euronews in May 2026 found that criminals' use of AI has so far had most impact on easy-to-automate fraud rather than sophisticated hacking. Voice-cloned CEO fraud fits that pattern: it is old social engineering with a new tool. The answer is the same as it has always been, applied consistently: verify through a second channel, and never let one person move money alone.

A practical next step

Pick a quiet afternoon this month and walk through your last ten outgoing payments. For each, ask who approved it, how a bank detail change would have been verified, and whether a convincing phone call could have skipped a step. Where the answer is yes, that is your first fix.

If you are adding automation to finance or approval workflows, the free AI Readiness Audit includes a review of where human checkpoints belong.

Sources

  • Swiss National Cyber Security Centre (NCSC), "Week 4: "Urgent money transfer" – CEO fraud remains a persistent threat to SMEs", 27 January 2026. link
  • Euronews, "AI scams can now impersonate your voice. Here's how to avoid them", 13 July 2025. link
  • Euronews, "Three in four EU workers face suspicious emails and links, report finds", 9 October 2026. link
  • Swissinfo (Keystone-SDA), "Fake authority scams fall 75% in Switzerland", 24 August 2026. link
  • Euronews, "Cybercriminals gave AI a go — and came away disappointed, study finds", 5 May 2026. link