Skip to content
Vigil

November 2026

The cost of doing nothing about the deepfake call

The fake executive on the call is no longer the story. The story is how fast it is spreading, how little most workforces can do about it, and what a year of standing still actually costs.

01

It is already operational


In January 2024 a finance employee at the engineering firm Arup, in its Hong Kong office, joined a video conference about a confidential transaction. The company’s UK chief financial officer was on the call, and so were several colleagues he recognised. Over the following week he made fifteen transfers, totalling about US$25.6 million, to five local bank accounts. Every other person on that call had been a deepfake. He had put aside an early doubt, police said, precisely because the people on screen looked and sounded like colleagues he knew, and the fraud came to light only when he checked with head office in the UK.

US$25.6M
stolen from a single company in one week, after a video call on which every colleague the finance employee saw was a deepfake.
Arup, Hong Kong, 2024

One detail is worth keeping straight, because it matters for what you do about it. Hong Kong’s security chief later told the Legislative Council that police believe the conference was pre-recorded, assembled from public clips rather than a live back-and-forth. Arup confirmed only that “fake voices and images were used.” So this was not a chatbot improvising in real time. It was a convincing, populated meeting, built to carry one instruction past a person’s judgment. That is the part that has since become routine.

Because Arup was not a freak event. In March 2025 a finance director at a multinational in Singapore joined a Zoom call where the chief executive and others were deepfaked, and moved over US$499,000 before Hong Kong police managed to withhold the funds. In February 2025 Italian police froze nearly a million euros that a businessman had wired after a phone call in a cloned voice he believed belonged to a government minister. And this month US prosecutors charged three men they allege used AI to impersonate executives and lawyers on calls to steal about US$129 million, a case we wrote about in “Your CEO just called. It wasn’t your CEO.” The attack now arrives as a conversation, in a voice and a face the target has every reason to trust.

02

What the numbers actually say


Start with the scale, from the one source with no product to sell. America’s FBI logged US$20.9 billion in reported cybercrime losses in 2025, up 26 percent on the year before. For the first time it tagged complaints that mentioned AI: more than 22,000 of them, with reported losses over US$893 million, and the Bureau was careful to add that most victims never realise AI was involved, so that figure is a floor, not a ceiling.

US$20.9bn
Reported to the FBI’s crime centre in 2025, up 26% on 2024
FBI IC3, 2026
US$893m
AI-related losses, the first year the FBI counted them separately
FBI IC3, 2026
US$40bn
Projected US generative-AI fraud losses by 2027, from US$12.3bn in 2023
Deloitte projection, 2024

The shape of that total matters more than the total itself. Business email compromise, the classic wire fraud that arrives by email, barely moved: US$2.4 billion in 2021, US$3.0 billion in 2025. Over the same four years total reported losses tripled. As a share of everything Americans lost to cybercrime, email BEC fell from about 35 percent to about 15 percent. The money is not draining out of the problem. The growth is simply happening somewhere other than the inbox.

Email fraud is a shrinking slice of a fast-growing problem

US reported cybercrime losses and business email compromise losses, US$ billions

Total reported lossesEmail BEC losses
$0$5$10$15$20$6.9bn$20.9bn$3.0bn20212022202320242025
FBI IC3, 2021 to 2025. Email BEC’s share of reported losses fell from 34.7% to 14.6% (Vigil calculation from IC3 totals).

Where it is going shows up in how intrusions actually begin. In Mandiant’s 2025 investigations, voice phishing surged to become the second most common way attackers got in, while email phishing fell by more than half, from 14 percent of intrusions to 6. Mandiant drew the line itself: an interactive attack, a live voice on a call, is far more resilient against automated controls than an email a filter can quarantine. The defence that has carried most organisations for a decade is aimed at the channel that is shrinking.

The way in is becoming a phone call

Share of investigated intrusions by initial entry point, where it could be identified

Email phishing, 2024
14%
Email phishing, 2025
6%
Voice phishing, 2025 · now the second most common vector
11%
Mandiant M-Trends 2026, investigations conducted during 2025. Bars scaled to the largest value shown.

And it is landing on real desks. In Gartner’s 2026 survey of nearly 300 chief information security officers, 41 percent said they had dealt with a deepfake on an employee audio call in the past year, and 36 percent on a video call. Be fair to the data: email phishing and BEC are still the most common incident at 79 percent. This is not email’s replacement. It is a fast-growing second front that a plurality of security leaders already met last year, and that almost none of their controls were built to stop.

A plurality of CISOs met a deepfake on a call last year

Share reporting at least one such incident in the previous 12 months

Email phishing, spear-phishing or BEC
79%
Voice or text phishing
58%
Deepfake during an employee audio call
41%
Deepfake during an employee video call
36%
Gartner survey of 297 senior cybersecurity leaders, fielded March to May 2026
03

Why the obvious defence misses it


The instinct is to teach people to spot the fake. The evidence says that is a losing game, and getting worse. When researchers at UC Berkeley played listeners voices cloned with an off-the-shelf tool, people judged the clone to be the real speaker about 80 percent of the time. In a University College London study, even after being warned that some clips were synthetic, listeners missed roughly one in four. The largest review to date, 56 studies and more than 86,000 people, puts human accuracy close to chance, and the gap widens as the generators improve. The detectors do not rescue this either: they fall apart on deepfakes they have not been trained on, and a passive detector cannot sit inside a live call in any case. We went through that evidence in “The live call can now be faked” and “The fake you were trained to spot is gone.”

Now look at how the real attacks actually ended. The breaches went through a person: Retool lost access after a cloned voice on a phone call talked an employee into reading back a one-time passcode, and Arup’s “I can see them on the video” was the very control the deepfake defeated. The saves were not sharper eyes, they were behaviour. Ferrari’s executive asked the caller to name a book he had recommended days earlier, and the call ended. WPP’s people challenged the request, and it collapsed. What worked, every time, was a rehearsed reflex to verify through a second channel before acting. Not perception. Procedure.

04

The cost of doing nothing


Put those together and you can see the shape of the next few years without having to invent a number. The trajectory is not seriously in dispute: Deloitte projected in 2024 that generative-AI-enabled fraud losses in the United States could reach US$40 billion by 2027, from US$12.3 billion in 2023. Europol calls fraud the fastest-growing area of organised crime and says AI-enabled fraud is “set to further escalate.” We will not print a 2030 figure, because no credible institution has one, and the honest claim is narrower and harder to argue with. The vector is growing, the bodies that track it expect it to keep growing, and it is arriving faster than most workforces are being prepared for.

And the preparedness is not there. When the Association of Certified Fraud Examiners surveyed more than 700 anti-fraud professionals this year, only 7 percent said their organisation was more than moderately prepared to detect or prevent AI-driven fraud. A separate survey of corporate-affairs leaders found 18 percent who felt ready for a deepfake or AI-misinformation incident and 43 percent who did not. On every independent measure we could find, most organisations are not ready for the attack that is already reaching them.

7%
of anti-fraud professionals call their organisation more than moderately prepared for AI-driven fraud
ACFE & SAS, 2026
18%
of corporate-affairs leaders feel prepared for a deepfake or AI-misinformation incident
GlobeScan, 2026
2%
the measured effect of embedded phishing training on whether staff clicked, across 19,500 employees
UC San Diego, 2025

The uncomfortable part is that the training most firms do run does not close the gap. In the two largest field studies ever done on security-awareness training, each randomised across tens of thousands of staff, the annual module made no measurable difference to whether people fell for a simulated attack, and the embedded “gotcha” lesson moved it by about two percentage points. Knowledge from a once-a-year course decays within months. The one thing that did change behaviour in the field was continuous, realistic practice: repeated, just-in-time exercises that halved successful compromises within six months. None of those studies tested a live deepfake call directly, so we hold the rehearsal case as a reasoned extension rather than a proven result. But the direction is unmistakable, and Gartner, independently, now tells its clients to stop teaching employees to “spot the fake” and instead make verification the expected behaviour for any risky request.

Move from teaching employees to spot the fake, toward making secure verification the expected behaviour.

Gartner, September 2026

Then there is the asymmetry, which is the whole argument in one line. One successful call is not a rounding error. On the FBI’s own totals, the average reported business email compromise cost about US$123,000 in 2025. Documented single incidents run to US$6 million in one FBI case file and US$25.6 million at Arup. Among firms actually hit by a deepfake, one industry survey put about a quarter at a million dollars or more once remediation and staff time are counted. Set any of that against the cost of rehearsing a few hundred people a few times a year, and the maths of standing still stops being close.

Doing nothing is not a neutral choice. It is a decision to let the first time your finance team hears a flawless fake of your chief executive’s voice be the real attempt, with the wire instructions already attached.

Where Vigil fits

Rehearse the call before an attacker makes it

This is the work we do, and we want to be precise about what it is and is not. We do not detect deepfakes, sit in your call platform, or stand in your payment flow. Detection vendors and identity providers own that ground, and the evidence above is clear that passive detection will not catch a good live fake on its own. What we do is the human layer every one of these attacks actually targets.

That means measuring, person by person, who is most likely to act on a convincing fake, and watching that risk move over time rather than guessing at it. It means running the attack for real, as a programme built around the scenarios your people will actually see: credential-request messages, and live AI phone calls in the consented, revocable cloned voice of a real colleague, so that the first time someone hears a perfect fake of a voice they trust, it happens in a rehearsal and not on a wire request. It means a one-click way for anyone to report a suspicious message and get an instant answer, with the people who need help routed straight into short, targeted practice. And it means the evidence, mapped to the frameworks an auditor or an insurer asks about, the day after something goes wrong.

None of that detects a deepfake. It does the thing detection cannot. It builds the verification reflex that stopped the attacks that were stopped, and it tells you, in numbers, how much stronger your people are getting.

To be clear about what Vigil is not

  • It does not detect, block or authenticate deepfakes.
  • It does not sit inside your call platform, your infrastructure or the payment path.
  • It is not a detection engine, an identity provider or a fraud filter.
  • It measures and changes human behaviour, and gives you the evidence for it.

The attack is already here, it is growing fastest in the one channel most defences ignore, and the training most organisations run does not change what their people do on a call. That gap is the exposure. It is also the one thing on this page you can actually close.

If you want to see what rehearsing and measuring that gap looks like for your team, we would be glad to show you.

The Vigil teamHuman risk management

Sources

  1. Arup deepfake video-conference fraud, January 2024: Kathleen Magramo, CNN, 17 May 2024; Hong Kong Secretary for Security to the Legislative Council (pre-recorded conference), via The Standard, 26 June 2024.
  2. Singapore Police Force, deepfake Zoom CFO scam, media release, 7 April 2025. Crosetto voice-clone case: Reuters, 12 February 2025.
  3. US Department of Justice (SDNY) indictment, three men charged over an alleged US$129 million AI impersonation scheme, 5 November 2026 (Associated Press). See also our note, “Your CEO just called. It wasn’t your CEO.”
  4. FBI IC3 Annual Report 2025. BEC and total-loss series, 2021 to 2025, from the IC3 2023, 2024 and 2025 reports; BEC loss-share is a Vigil calculation from those totals.
  5. Mandiant M-Trends 2026 (Google Cloud), executive edition, March 2026.
  6. Gartner, “Survey Finds 41% of CISOs Reported at Least One Social Engineering Incident Involving a Deepfake in the Past 12 Months,” 22 September 2026. Figures and the quoted guidance are press-verified (Help Net Security, CXOToday); the Gartner release page blocks automated retrieval.
  7. Deloitte Center for Financial Services, “Deepfake banking fraud risk on the rise,” 29 May 2024 (US projection, 2023 to 2027). Europol, IOCTA 2026.
  8. Preparedness: ACFE and SAS 2026 Anti-Fraud Technology Benchmarking Report, 25 March 2026; GlobeScan Global Business Risk Outlook, 13 May 2026.
  9. Training efficacy: Ho et al., UC San Diego, 46th IEEE Symposium on Security and Privacy, 2025; Lain, Kostiainen and Capkun, ETH Zurich, IEEE S&P 2022; continuous just-in-time training, Toth et al., 2025.
  10. Detection: Barrington, Cooper and Farid, Scientific Reports, 2025 (voice clones); Mai et al., PLOS ONE, 2023 (UCL); Diel et al., Computers in Human Behavior Reports, 2024 (meta-analysis); Deepfake-Eval-2024 and CSIRO, 2025 (detectors).
  11. Controls bypassed and saved: Retool, company disclosure, 27 August 2023; Ferrari, Bloomberg, 26 July 2024; WPP, The Guardian, 10 May 2024. Per-incident cost among affected firms: Pindrop Deepfake Readiness Index, 2026 (vendor survey).
  12. Verizon 2026 Data Breach Investigations Report (human element and mobile-vector findings).