A finance worker in Hong Kong once joined a video call with his CFO and several colleagues, and every single one of them was an AI-generated fake. He wired $25 million before anyone realized nothing on that call was real. That is not a warning about where fraud is headed. It already happened, and small businesses are next in line.
Fraud used to have a certain texture to it. A phishing email with clumsy grammar. A caller whose accent did not quite match the story. A voice on the phone that sounded almost, but not quite, like your supplier. Those small tells used to buy a moment to pause and think.
AI is removing that moment. It can write a flawless email in your accountant's tone, clone a colleague's voice from a few seconds of audio pulled off a company video, and generate a face on a video call that looks and moves like someone you have worked with for years. None of this is science fiction anymore, and none of it is reserved for large corporations with deep pockets and dedicated security teams.
This guide looks at how AI is actually changing business fraud right now, what that has already cost real companies, and what a small business without a security department can realistically do about it.
| A NOTE ON SCOPE This article provides general information about a fast-moving threat, not tailored cybersecurity, financial, or legal advice. If your business has experienced a specific fraud incident, contact your bank, local law enforcement, and appropriate professionals directly. |
Fraud itself is not new. What is new is what AI removes from the equation: cost, skill, and time. Three shifts explain most of what businesses are now running into.
• Impersonation now has a face and a voice. Cloning a person's voice used to require real effort. Now it takes seconds of audio that most executives have already put online in a conference talk, a company video, or a podcast appearance.
• Phishing got fluent. The broken English and generic greetings that used to be the easiest warning sign are disappearing, because AI writes convincingly in whatever tone and language the scammer needs.
• Scale got cheap. A single fraudster can now run dozens of tailored attacks a day instead of one, because the writing, the voice, and even the video no longer need to be done by hand.
The numbers back this up. KnowBe4's own research found that 82.6 percent of the phishing emails it analyzed between September 2024 and February 2025 showed some use of AI. On the voice side, McAfee's Beware the Artificial Impostor report found that just three to four seconds of audio was enough for its researchers to produce a voice clone with an estimated 85 percent match to the original speaker.
None of that is abstract. It is exactly what happened to one company that has since become the clearest cautionary tale in this space.
In January 2024, a finance employee at the Hong Kong office of Arup, a London-based engineering firm, received an email that appeared to come from the company's UK-based chief financial officer. It asked him to carry out a confidential transaction. He was suspicious enough to ask for a video call to confirm it.
That call was the trap. The CFO was on it, along with several colleagues the employee recognized. Every one of them was an AI-generated deepfake. Hong Kong police said the fraudsters built them using publicly available video and other footage of the executives. Reassured by seeing and hearing people he knew, the employee carried out fifteen transfers totaling roughly $25.6 million to accounts controlled by the fraudsters, all in a single day.
The fraud only came to light because the employee later called Arup's head office to follow up on the transaction. Nobody there knew anything about it. Arup's chief information officer later described the incident as technology-enhanced social engineering rather than a hack. No firewall was breached and no systems were compromised. The attack went through a person, not through code.
As of the most recent public reporting, the money has not been recovered and no one has been publicly charged, with the case reported as still under investigation by Hong Kong police. What makes this story worth knowing, even for a business nowhere near Arup's size, is how ordinary the setup was. A believable email. A request for urgency and confidentiality. A video call that looked exactly like the real thing. That same setup scales down to a five-person accounting firm just as easily as it scaled up to a multinational.
Arup is the headline case, but it is one data point inside a much bigger, faster-moving trend.
Fraud losses tied to AI are still a small share of overall cybercrime losses today, but they are growing far faster than almost anything else being tracked.
| $20.9B in total losses reported to U.S. cybercrime victims in 2025, up 26% from the year beforeFBI IC3 2025 Annual Report | $3.05B lost to business email compromise in 2025, from just under 25,000 reported complaintsFBI IC3 2025 Annual Report |
| $893 in losses tied to AI-related complaints, the first year IC3 tracked this as its own categoryFBI IC3 2025 Annual Report | $40B Deloitte's aggressive-scenario projection for annual U.S. bank fraud losses enabled by generative AI by 2027, up from $12.3B in 2023Deloitte Center for Financial Services |
The direction of travel matters as much as the totals. Verizon's 2026 Data Breach Investigations Report found that phishing accounted for 44 percent of AI-assisted initial access into breached organizations, and that the human element, people being persuaded rather than systems being hacked, was present in 62 percent of all breaches it examined. That lines up with what Arup's own leadership concluded about their incident: the technology did not break in anywhere. It talked its way in.
Numbers this size can feel like a big-company problem. They are not, and the reasoning for why is worth spelling out.
It is tempting to assume a small business is beneath a fraudster's notice. The opposite is closer to the truth, for a few plain reasons.
1. Fewer checks stand in the way. Larger companies often require a second sign-off for big payments. Many small businesses do not, because one or two people handle finance and there is no one else to ask.
2. Everyone wears more than one hat. The person answering emails is often also the person approving payments, which is exactly the kind of overloaded, distracted moment a well-timed scam depends on.
3. Public information does the fraudster's work for them. A small business owner's voice, face, and speaking style are often sitting in plain view on social media, a podcast interview, or a local news clip, which is all a fraudster needs to build a convincing clone.
4. Recovery options are thinner. A large company can absorb a loss that would be existential for a smaller one, and a small business is less likely to have cyber insurance or a dedicated response plan in place before something goes wrong.
It helps to see what this actually looks like day to day, rather than only as a headline-grabbing case like Arup's.
A voice that sounds exactly like the owner or a senior manager calls or leaves a voicemail asking for an urgent payment, gift cards, or a change to a vendor's bank details, usually while stressing secrecy and time pressure. That combination, a familiar voice plus urgency plus secrecy, is the same recipe behind the Arup case, just without the video call.
An email that looks exactly like it came from a regular supplier asks for future payments to go to a new account. AI makes it easy to match the supplier's actual writing style, invoice format, and tone closely enough that nothing looks off.
Not every AI-related scam targets a business's customers or staff. Some target the owner directly. Kaspersky's SMB threat research found that fake AI tools have become a common lure among fraudsters going after small businesses and entrepreneurs specifically, often marketed as productivity or marketing tools that instead harvest credentials or payment details once installed.
Most of these examples rely on the same weakness: a person, under time pressure, trusting what they see or hear without a second channel to check it. That is also where the fix lives.
There is no single new law written specifically for AI-enabled fraud, and it is worth being precise about what does and does not exist rather than overstating it. In the United States, the FTC's Government and Business Impersonation Rule is an actual, finalized federal regulation, published in the Federal Register and in effect since April 1, 2024, that makes it unlawful to impersonate a government agency or business. Separately, the FTC has proposed extending similar protections to cover AI-based impersonation of individual people, but as of this writing that extension remains a proposal, not a final rule, and the distinction matters if you are relying on it.
The FTC's own consumer guidance is worth reading directly too, because it applies just as well to a business owner as to a family, though it is U.S. focused guidance rather than a global standard. Its advice on voice cloning scams is simple: if a call asks you to act urgently and in secret, hang up and call the person back using a number you already know is theirs, not one given to you during the call.
Outside of that one finalized rule, most of what exists is older fraud and consumer protection law being applied to a new delivery method, plus a growing set of official warnings rather than AI-specific statutes. Businesses outside the United States should expect a similar pattern: existing fraud law doing most of the work, with AI-specific rules still catching up.
With the regulatory picture reasonably clear, the practical question is what a small business without a security team actually does differently starting this week.
1. Separate the channel that asks from the channel that confirms. If a payment, bank detail change, or urgent request arrives by email, confirm it by phone using a number you already have on file, never a number provided in the message itself. If it arrives by phone or video, confirm it by calling back a known number too. This single habit is the one most likely to have caught the Arup fraud and cases like it.
2. Never change a supplier’s bank details on the strength of an email alone. If a supplier says its payment instructions have changed, confirm the new details independently with a known contact at that supplier, using a phone number already on file, not anything included in the email itself.
3. Add a shared code word, as a backup, not a substitute. A short phrase known only to the people who authorize payments is one more thing a cloned voice or a deepfake video cannot easily fake. But it works best layered on top of callback verification and a second-person sign-off, which are the stronger controls, not in place of them.
4. Slow down anything urgent and secret. Urgency and confidentiality together are the two most reliable warning signs across almost every case in this guide. A legitimate request can wait for a callback.
5. Require a second person on anything above a set amount. Even a simple rule, two people must approve any payment over a fixed threshold, closes most of the gap that made Arup's single-employee approval possible.
6. Think about what your own public audio and video reveal. A founder's conference talk, podcast appearance, or promotional video is also, unfortunately, training material for a voice clone. This is not a reason to disappear from public life, but it is a reason to make sure payment approval never rests on voice recognition alone.
7. Treat unfamiliar AI tools with the same caution as unfamiliar email attachments. Before installing an AI tool you have not vetted, especially one that arrived through an ad, a cold email, or a social media post, check who actually built it, what it asks to access, and what happens to your data once it is in use.
This is not a distant, foreign-market problem for the region. The Central Bank of The Bahamas issued a public notice warning that scammers were using deepfake videos, audio, and images to impersonate financial institutions and officials in fraudulent investment schemes. The Bank was explicit that its official communications only ever come through its own verified website and channels, and it directed anyone contacted this way to report it to the Royal Bahamas Police Force.
That warning was not hypothetical. National Security Minister Wayne Munroe publicly confirmed that deepfake videos had circulated impersonating the Prime Minister, the Governor-General, and the Governor of the Central Bank to promote bogus investment products, and confirmed that the cybercrime unit of the Royal Bahamas Police Force was investigating.
The underlying lesson for a Bahamian or Caribbean business owner is the same one running through this entire guide: verify through a channel the fraudster does not control before acting on anything involving money, whichever government agency, bank, or supplier the message claims to be from. If your business has been targeted or has reason to believe it is at risk, reporting it to local law enforcement and your bank directly is the right first step, not a general guide like this one.
Most of what this article recommends costs nothing but a decision. Where CSB Tech Emporium can help is on the technology and operations side: reviewing your business systems and digital workflows, strengthening access and approval processes, and implementing the verification controls this article describes so they are built in rather than left to memory. This is a technology and implementation conversation, not legal, insurance, or law enforcement advice, and if your business has already been targeted, your bank and local police should be your first call. If you would like a hand setting up safer defaults, reach out to CSB Tech Emporium.
Often, partly. Many cyber insurance and crime insurance policies cover business email compromise and social engineering losses, but coverage can vary significantly depending on the policy, the insurer, any endorsements attached, and the specific circumstances of the loss, including whether basic verification steps, like a callback confirmation, were followed. It is worth reviewing your specific policy with your insurer rather than assuming coverage exists. Coverage questions often come down to who was responsible for what, which gets harder to untangle once AI is involved. Our guide to who is responsible when AI gets it wrong covers that question in more depth.
Not reliably. McAfee's own research found that 70 percent of people said they either were not confident, or did not think they would be able to, tell a cloned voice apart from a real one. That is exactly why the safer control is a callback to a known number, not trying to listen harder.
No. The strongest defenses in this guide, callback verification and a two-person rule for larger payments, cost nothing to put in place, and a shared code word is a useful extra layer on top of them. They require a decision, not a purchase.
Not necessarily. The more realistic response is to make sure no payment or sensitive decision in your business ever depends on recognizing a voice alone, regardless of how much or how little audio of you exists publicly.
Callback verification. Before acting on any request involving money, secrecy, or urgency, call the person back on a number you already had before the message arrived. It is the habit most likely to have interrupted almost every case described in this guide.
