A cloned voice asks for a wire: the payment checks a small business adds against deepfake fraud
Real cases from the EUR 220,000 voice clone of 2019 to Arup's $25.6m video call, the 2025 loss figures (FBI: $3.05bn to business email compromise; UK CEO fraud averaging GBP 28,000 a case), and five checks that cost a 12-person company about $300 a year: call back on a known number, two approvers, a code word, the bank's name check and a one-day hold on changed details. Plus the first hour after a payment goes wrong.
The call comes on a Thursday afternoon. It is the owner's voice, a little rushed: a supplier deal is closing today, the lawyer will email the bank details, please send $48,000 before the bank's cut-off and keep it quiet until it's announced. The bookkeeper has heard that voice for six years. Nothing about it sounds wrong.
A clone of someone's voice now takes a few seconds of audio and a $5 subscription, and a fake video call takes not much more. You can't train staff to hear the difference, and you don't have to. The scams that work all end the same way, with a payment to an account you have never paid before, and the checks that stop them are about the payment, not the voice. Below are the real cases and their sums, the 2025 loss figures, five controls a small business can set up in an afternoon, and what to do in the first hour if the money has already gone.
What has actually happened
Two things stand out. First, every attempt that failed failed on a check, not on the quality of the fake: an unusual channel (LastPass), a question only the real person could answer (Ferrari), a call from a number that didn't fit (the 2019 case, too late). Second, the losses are not only at large companies. The Singapore victim was a finance director paying from an ordinary business account, and the attack that hits small firms most often doesn't need a voice at all: an email from a real supplier's hacked mailbox saying the bank details have changed.
How much money moves this way
The UK's CEO fraud line is the one to look at: few cases, the highest loss per case of any scam type, and only one pound in five comes back. That is what an impersonated boss asking for an urgent payment costs.
Why you can't rely on spotting the fake
Microsoft's VALL-E research model cloned a voice from a three-second sample in January 2023. Commercial tools need a bit more audio and are cheap: ElevenLabs offers instant voice cloning from its starter plan, about $5 a month. When Consumer Reports tested six voice-cloning products in March 2025, four of them (ElevenLabs, Speechify, PlayHT and Lovo) only asked the user to tick a box saying they had the right to clone the voice. A business owner's voice is on podcasts, webinars, voicemail greetings and conference videos.
So the rule has to work even when the voice, the face and the email address are perfect. The five controls below don't ask anyone to judge whether a call is real. They make sure money only goes to an account that has been confirmed through a channel the attacker doesn't control.
Five controls, and what each costs
1. Call back on a number you already had. Any request to pay a new account, change a supplier's bank details or make an urgent transfer is confirmed by calling the person on a number from your own records: the contract, the supplier file, your phone's contacts from before the request. Never the number in the email, the signature or the voicemail. If the boss is the one asking, call the boss's own mobile. It costs nothing and takes five minutes. Insurers ask about it by name: Coalition's application asks whether you use "a secondary means of communication to validate the authenticity of" transfers over $5,000 or $25,000 and "any request to change banking details", and social engineering claims are denied when the call-back wasn't done.
2. Two people for new payees and large payments. One person sets up a payment, another approves it, and the second person sees the bank details, not just the amount. Most business banking does this already: Chase's Dual Control makes a second user approve scheduled payments; Wise Business lets you require one or two approvers for all payments or for those above an amount; Mercury has payment approval rules and a dual-admin policy; QuickBooks needs Bill Pay Elite or QuickBooks Online Advanced for approval workflows. In a three-person office the second approver can be the owner from a phone; what matters is that the person who took the call is not the only one who can send the money.
3. A code word, or a question only the real person can answer. The FBI's December 2024 alert (I-120324-PSA) recommends a secret word or phrase for exactly this. Pick one for payment requests, keep it out of email and chat, and change it when someone leaves. Ferrari's executive did the improvised version: he asked which book the chief executive had recommended to him a few days earlier.
4. Let the bank check the name. In the UK, Confirmation of Payee checks that the name you typed matches the account; it has been compulsory for about 400 payment providers since 31 October 2024, runs more than 70 million checks a month, and Pay.UK reports 59% fewer "wrong account" recovery claims since launch. In the EU, Verification of Payee became compulsory for all euro credit transfers on 9 October 2025, with match, close match and no match results (businesses paying in bulk can opt out; don't). The US has no national name check yet. Ask your bank about account validation (Early Warning's Verify Account is the service behind many of them), and note that Nacha's new fraud-monitoring rules, which name business email compromise and vendor impersonation, applied to all originators from 22 June 2026.
5. A waiting period on changed bank details. New or changed details go on hold for a business day before the first payment, and the supplier is told so in your onboarding letter. Nobody legitimate is hurt by a one-day wait; every scam depends on today. This one is our recommendation rather than a regulation, and it costs one day's delay on a payment that changes perhaps a few times a year.
Worked through: a 12-person company
A design studio with 12 people pays about 40 supplier invoices a month and makes two or three transfers above $10,000. Here is what the five controls cost it in a year, against what one successful fraud costs.
The arithmetic is lopsided enough that the real question is not whether to do it but whether people will follow the rule when the boss sounds stressed. That's why the rule has to be written down, agreed by the boss in advance, and include the sentence "I will never be annoyed that you called me back".
If the money has already gone
Speed decides how much comes back. At-Bay's 2026 report says clients who reported funds transfer fraud within three days got at least some money back 70% of the time, and the odds dropped sharply after that. The FBI's Recovery Asset Team handled 3,900 incidents in 2025, tried to stop $1.164bn and froze $679.0m, a 58% success rate. In the first hour:
Who carries the loss
Usually the payer. In Mobius Group v Inoteq, an Australian court held that a supplier whose hacked mailbox sent fake bank details owed its customer no duty of care, and the customer, whose phone check had failed, had to pay the invoice of more than $190,000 again. A German court (LG Rostock, 20 November 2024) ruled the same way where the forgery was recognisable. In the US, courts in cases such as Beau Townsend Ford v Don Hinds Ford have split the loss according to which side failed to take ordinary care, which is another reason to have a written check you can show you followed.
Two partial safety nets. In the UK, payment service providers have had to reimburse victims of authorised push payment fraud since 7 October 2024, up to GBP 85,000 a claim, and this covers micro-enterprises (fewer than ten staff and turnover or balance sheet up to EUR 2m) and small charities, not larger firms; 88% of claimants were reimbursed in the first year. And cyber or crime insurance with social engineering cover, typically sublimited to $100,000 to $250,000, usually on condition that you did the call-back. What the rest of a cyber policy pays, and the exclusions, are in our piece on cyber insurance for a small software company.
The one-page payment rule
Keep it next to whatever your team already has on AI use; if you have nothing yet, our one-page AI policy is a short start, and the payment rule can sit at the bottom of it.
Your insurer will ask about payment callbacks, MFA and backups. See which of your answers would stop a quote, shrink the cover or raise the price, with the fixes in order. Runs in your browser.
Open the cyber insurance application checkSources
Cases: PaymentsJournal on the 2019 energy firm and Sophos; CNN on Arup (May 2024); LastPass (Apr 2024); The Guardian on WPP (11 May 2024); MIT Sloan on Ferrari; Reuters via Yahoo on the Crosetto voice clone (2025); Singapore Police Force (7 Apr 2025); Mondaq on Mobius v Inoteq; heise on LG Rostock 2 O 450/24.
Figures: FBI IC3 2025 report and 2024 report; FBI alert I-091124-PSA (Sep 2024); UK Finance Annual Fraud Report 2026 (Jun 2026); AFP 2026 Payments Fraud and Control Survey (Apr 2026); Coalition 2026 Cyber Claims Report; At-Bay 2026 InsurSec Report.
Voice cloning: VALL-E, three-second samples (Jan 2023); Consumer Reports on six cloning tools (10 Mar 2025).
Controls: Coalition US cyber application; Amwins on social engineering claims; Ward and Smith on sublimits; Chase, Wise, Mercury and QuickBooks approval settings; Pay.UK on Confirmation of Payee; PwC on EU Verification of Payee; Nacha fraud monitoring phase 2; Early Warning Verify Account; PSR reimbursement dashboard. Read 24 October 2026. Not legal advice; your bank's and insurer's terms decide what applies to you.