What is the most important protection against CEO fraud via deepfake?

A fixed verification rule: payments or changes to payment details must always be reconfirmed via a second, independent channel, such as a callback to a stored number. A secret code word adds another layer.

Do we need to buy special detection software?

Not as a first step. Processes such as the four-eyes principle, code words, and training reduce risk significantly at low cost. Detection tools can be useful later as a second line of defense.

How can employees recognize a deepfake call?

Mostly they cannot. Humans correctly identify high-quality deepfake videos only 24.5 percent of the time. Protection therefore focuses on process and verification, not gut feeling.

What role does an AI policy play?

It defines, among other things, what voice and image data employees may share publicly. Less publicly available voice and image material makes it harder for attackers to create a convincing clone.

AI Insights

Deepfakes Hit SMEs: Protecting Your Company from AI Fraud

Chris Jon Graf · AI Strategist & CEOPublished on 23 October 2026

Download PDF Briefing

Deepfakes Hit SMEs: Protecting Your Company from AI Fraud

In short

Deepfake fraud is a concrete but manageable threat for Swiss SMEs in 2026. Just three seconds of audio can produce a voice clone. Effective verification rules, a second-channel callback, a code word, and training against manipulative urgency significantly reduce risk without expensive specialty tools.

What Is Deepfake and Voice Cloning Fraud?

Deepfake fraud uses artificially generated video or audio to impersonate a trusted person. For SMEs, this is no longer a fringe issue. According to Fortune in March 2026, deepfake fraud cost US companies USD 1.1 billion in 2025, a tripling from USD 360 million the previous year. The number of deepfakes rose from 500,000 in 2023 to over 8 million in 2025, and voice cloning fraud grew by 680 percent in one year.

The New Quality: CEO Fraud 2.0

The case of engineering firm Arup shows how far attacks have progressed: a finance employee in Hong Kong paid USD 25 million after speaking with a deepfake version of his CFO in a video call (CNN, February 2024). What used to be forged emails is now a deceptively real video conference. Attackers combine technical deception with psychological pressure, usually urgency and authority.

What This Means for Swiss SMEs

Figures from Surfshark Research in July 2025 show this is not just a large-enterprise problem: deepfake fraud had caused almost USD 900 million in losses by then, with USD 410 million in the first half of 2025 alone, more than the USD 359 million for all of 2024. CEO fraud, the impersonation of a company executive, accounted for USD 217 million. The IRONSCALES Report from October 2025 puts the median loss at over USD 280,000; 61 percent of affected companies lost more than USD 100,000, and 19 percent more than USD 500,000.

When a successful attack leads to financial damage, liability questions arise quickly. Swiss companies should also monitor their dependence on non-European AI platforms; the Swiss AI podcast on AI sovereignty provides useful context.

Why Detection Is So Hard

The technical threshold is now minimal. According to SecurityToday in April 2026, just 3 seconds of audio are enough for a voice clone with 85 percent accuracy. Humans correctly identify high-quality deepfake videos only 24.5 percent of the time. Adaptive Security reported in November 2025 that producing a single deepfake costs on average under USD 2, and free tools can clone a voice in under 60 seconds. That means you cannot rely on the human eye or gut feeling.

24.5%

of deepfake videos are correctly identified by humans (SecurityToday, April 2026)

85%

of companies experienced at least one deepfake incident in the last 12 months (SecurityToday, April 2026)

Protection Measures That Work Immediately

The good news: you do not need a large budget to reduce the risk dramatically. The following measures target the weakest point, the process, not detection technology.

Fixed verification rule: every payment release or change of payment details must be reconfirmed via a second, independent channel, such as a callback to a stored number.
Secret code word for critical transactions, never shared by email.
Four-eyes principle above a defined threshold, so no single person can trigger a transfer alone.
Training against the 'urgency plus authority' pattern: employees learn to pause on unusual requests.
AI policy that defines what voice and image data employees may share publicly.
Access to detection or biometric tools as a second line of defense, not as a substitute for processes.

Second-channel callback: Effort: low | Protection: very high | Cost: none
Secret code word: Effort: low | Protection: high | Cost: none
Four-eyes principle: Effort: low | Protection: high | Cost: none
Urgency training: Effort: medium | Protection: high | Cost: low
AI policy: Effort: medium | Protection: medium | Cost: low
Detection tools: Effort: medium | Protection: medium | Cost: variable

Do this now

Set a rule today: no payment or change of bank details may be approved based solely on a video or voice request. This is the single most effective measure.

A Sober Approach, Not Alarmism

The threat is real but manageable. Many SMEs underestimate the risks of shadow AI, uncontrolled use of AI tools in daily work, which often supplies the voice material for cloning attacks. Identifying where AI is already used is a necessary first step.

Your Next Step

Start with a simple one-hour workshop: define thresholds, choose a code word, and adjust your payment processes. Then codify the rules in a short policy and train relevant employees. Full technical monitoring can come later, but it is not the first step.

Frequently asked questions

Do 3 seconds of audio really enable a voice clone?: Yes, according to SecurityToday in April 2026, 3 seconds of audio are enough for a voice clone with 85 percent accuracy. Free tools can do this in under 60 seconds.
What is the most important protection against CEO fraud via deepfake?: A fixed verification rule: payments or changes to payment details must always be reconfirmed via a second, independent channel, such as a callback to a stored number. A secret code word adds another layer.
Do we need to buy special detection software?: Not as a first step. Processes such as the four-eyes principle, code words, and training reduce risk significantly at low cost. Detection tools can be useful later as a second line of defense.
How can employees recognize a deepfake call?: Mostly they cannot. Humans correctly identify high-quality deepfake videos only 24.5 percent of the time. Protection therefore focuses on process and verification, not gut feeling.
What role does an AI policy play?: It defines, among other things, what voice and image data employees may share publicly. Less publicly available voice and image material makes it harder for attackers to create a convincing clone.

Sources

About the Author & Your Direct Contact

Chris Jon Graf · AI Strategist & CEO

For over 15 years, Chris Jon Graf has built productive Swiss software platforms and advised executive teams on the strategic deployment of artificial intelligence. Would you like to evaluate this topic for your company?

Check AvailabilityWould you like to explore this topic for your company?

AI Agents Have Arrived, Processes Haven't: Three Gaps Swiss Firms Must Close Introducing AI Agents: Why Day 14, Not Launch, Decides Your Success Connecting AI Agents to Swiss ERP: What Works with Abacus, Bexio & Co.