The Deepfake Problem Nobody’s Pricing In Yet
Deepfake fraud has moved from a technology story into a balance sheet story faster than most boards expected. In the first half of 2026 alone, reported deepfake fraud losses reached US$764 million worldwide, a businessman in Singapore transferred 4.9 million after a Zoom call with an AI-generated version of Prime Minister Lawrence Wong, and the chairman of an Italian private bank authorised around €95 million in transfers to accounts in China and Hong Kong after a message from what appeared to be his group chief executive and a call from what sounded like a senior lawyer. The losses are real and rising, but the insurance market that would normally absorb them has almost no claims data to price from, and much of it is moving to exclude the risk rather than cover it. That leaves the exposure sitting quietly with the companies and individuals being impersonated, many of whom do not yet know it is there. This is a briefing on that gap.
The Numbers Are Already Here
For the first time, the FBI’s 2025 Internet Crime Report gave artificial intelligence its own section, recording 22,364 complaints that referenced AI and US$893 million in associated losses, with the FBI itself cautioning that underreporting means the true figure is likely higher. US $3.7 billion since 2020, roughly 89% of it in 2025 and the first half of 2026, and because it only counts cases that became public, it is best read as a floor rather than an estimate. Looking forward, Deloitte projects that losses from AI-enabled fraud in the United States alone could reach US$40 billion a year by 2027.
The pattern inside those figures matters as much as their size. Ponemon Institute’s 2025 survey of security professionals found that 41% of organisations had experienced deepfake attacks targeting their executives, up from 34% in 2023, and a RAND study published in September 2026 found that deepfakes accounted for 84% of the 713 generative AI incidents it reviewed. This is no longer a niche technique used against a handful of unlucky firms. It has become the dominant way generative AI is being turned against people.
Why the Insurance Market Has Not Caught Up
Insurers price risk from loss history, and for deepfake fraud that history barely exists yet. Industry reporting this year describes public AI claims data as effectively nonexistent, with no reliable way yet to estimate either the size of the protection gap or how losses might accumulate across policies. Without that data, the market has largely chosen caution over cover. In the US admitted market, 60 insurers have filed 112 AI-related exclusion forms against just 7 endorsements, and at the time of reporting no stand-alone AI policy had been filed at all.
The result is a split market. At the January 2026 renewals, some carriers removed deepfake fraud from their standard social engineering cover, while others began marketing it as an affirmative extension, usually at additional premium. Even where deepfake fraud is covered, social engineering losses are commonly subject to a sublimit set well below the headline policy limit, so a business that believes it is insured for a seven-figure loss may find the relevant cover is a fraction of that. Many companies renewed their policies this year without noticing which side of that split they ended up on.
What That Means When It Happens
When cover is excluded, sublimited or disputed, the loss stays with the business. The engineering firm Arup lost HK$200 million in Hong Kong in 2024 after an employee joined a video call in which every other participant was a deepfake. In the Fideuram case reported this year, around €53 million had been recovered at the time of reporting, leaving roughly €36 million still missing after it was routed through overseas accounts and converted into cryptocurrency. Recovery in these cases is partial at best and slow at worst, and the reputational cost of a senior figure being successfully impersonated does not appear on any claim form.
Where Every Deepfake Starts
What these cases have in common is not sophisticated hacking but source material. A convincing deepfake of an executive needs their face, their voice, their manner of speaking and enough context about their relationships to make a request sound plausible, and for most senior people all of that is already public, in conference recordings, media interviews, earnings calls, podcasts and social media. In the Fideuram case, the voice reportedly cloned was not even the executive’s, but a well-known lawyer’s, chosen because his confirmation would make the request believable. The people whose likeness is used rarely know how much usable material exists, or where it sits.
What Can Actually Be Reduced
Insurance may eventually catch up, but the one input a company or individual can control today is the raw material itself. An Advanced Digital Vulnerability Assessment includes a specific review of AI-generated impersonation, looking for deepfake images, video or audio depicting the individual, alongside an assessment of how much of their voice, image and personal context is publicly available to build one. It also identifies impersonation accounts and fake profiles already using a person’s name or likeness, which can then be taken down through Digital Footprint Reduction.
The second control is procedural rather than technical. Every one of the cases above succeeded because a request that appeared to come from a trusted senior person was acted on through the same channel it arrived in. A verification step that requires confirmation through a separately established contact, whatever the seniority of the person asking, would have interrupted each of them.
Why This Matters Now
The organisations most exposed to deepfake fraud over the next few years are unlikely to be the ones with the weakest technology. They are more likely to be the ones whose senior people have the largest public footprint, the least awareness of it, and an insurance policy that quietly stopped covering this risk at the last renewal. The market will price this problem eventually, once enough losses have been paid or refused. Until then, the cost of it sits with whoever has done the least to reduce their exposure.
