When people talk about deepfake scams, they often focus on dramatic examples: a fake CFO joining a video call or a cloned CEO asking an employee to send money urgently. These stories get attention, but they can distract us from a more common and practical threat.

A scam can be much simpler. Imagine your company receives a normal-looking email from a supplier you have worked with for years:

“We’ve changed banks. Please update our bank details for the next payment.”

Your accounts payable (AP) employee follows the company’s usual procedure. They call the supplier to confirm the request before changing the bank details.

That phone call used to be a reliable way to prevent fraud. But AI can now create convincing fake voices, meaning the person on the phone may sound exactly like the real supplier.

The important point is this: hearing a familiar voice on the phone is no longer enough to prove that you are speaking to the real person.

Companies need to rethink how they independently verify supplier payment changes and put stronger controls in place before a simple invoice update turns into a costly fraud.

Follow the money, and the pattern is clear. Spectacular executive-video heists make headlines. Quiet vendor-payment redirections are the industrial-scale problem.

In the US, the FBI’s Internet Crime Complaint Center recorded $2.77 billion in reported business email compromise (BEC) losses across 21,442 complaints in 2024. In Australia, payment redirection fraud rose 9.3% to $166.8 million in 2025, making it one of the few major fraud categories still increasing (National Anti-Scam Centre data, via analysis).

The common thread across jurisdictions is the same: the target is the person who processes the supplier’s bank-detail change and releases the payment.

Now add the accelerant.

Deloitte’s Center for Financial Services projects that generative-AI-enabled fraud in the US could reach $40 billion by 2027, up from $12.3 billion in 2023 (Deloitte Insights). And the FBI reported that in 2025 it received more than 22,000 complaints involving AI-generated voice or video, with losses approaching $893 million (U.S. Bank summary of FBI data).

Attackers aren't spending that capability on theatrics. They're applying it to the highest-yield, lowest-friction step in the finance process.

That step is the callback verification.

“Call the vendor to confirm” was never a magic control. It worked because it quietly depended on two assumptions.

Voice cloning has now broken both.

Assumption 1: You’re calling a number the fraudster doesn’t control.

In practice, rushed AP teams often call the number on the invoice or in the email signature, exactly the information the attacker supplied.

Add caller-ID spoofing, and the fraudster can either answer your call or call you first. The control appears to be working, while the verification channel is already compromised.

Assumption 2: You can recognise the person’s voice.

This is the assumption that has fundamentally changed.

Off-the-shelf tools can now clone a specific person’s voice from seconds of clean audio, a webinar clip, a WhatsApp voice note, or a leaked support call.

The result is uncomfortable but simple: “It sounded like him/her” is no longer meaningful evidence of identity.

For years, we treated a familiar voice as a credential.

A cloned voice means that credential can now be manufactured by the attacker, on demand.

The callback still happens.

It just no longer proves what the organisation thinks it proves.

The fix is not a better phone call. It’s verifying facts the attacker cannot produce on demand, through a channel they don’t sit inside. Apply three principles:

1. Verify data, not senses:
Move the trust signal from “does this sound right?” to “does this attribute match a record I control?” The strongest example is name-to-account matching. The EU’s Instant Payments Regulation (Regulation (EU) 2024/886) made Verification of Payee mandatory across the euro area from 9 October 2025, banks must confirm the payee name matches the IBAN before a transfer completes. The UK’s equivalent, Confirmation of Payee, does the same. Use it as a hard gate: a “no match” on a bank-change request is a stop, not a hesitation.

2. Verify through a pre-established channel, never the one supplied in the request:
Independence means the contact detail was captured before the fraud began. Call the number in your vendor master file, not the invoice and treat any request to “use my new number” as a red flag, not a convenience.

3. Verify the change, not the messenger:
The Australian Signals Directorate’s guidance is blunt and correct: confirm payment-change and large-transfer requests by calling the sender on a known, verified number and route them through a defined approval process (ACSC / ASD, Preventing BEC). Pair that with a second named approver in the system. One cloned call cannot satisfy a two-person control.

Put simply: independent verification checks something the attacker would have to steal from you to fake, a shared secret, a master-data record, a bank’s name-IBAN result, rather than something they can generate - a voice, a face, a plausible email.

Even security-aware teams get tripped up by the same misconceptions.

“We’ll buy a deepfake detector.” Detection is an arms race you don’t need to enter for this problem. You cannot reliably eyeball or ear-ball a modern clone, and vendors’ detection scores lag the generators. Verification sidesteps detection entirely: you don’t have to prove the voice is fake if the process never trusted the voice.

“We called a number, so we verified.” Calling a number proves nothing. Calling the pre-registered number, and reaching a person who can confirm a fact you both already know, is the control.

“The voice sounded exactly like our contact.” In 2026, that sentence is the description of a successful attack, not a failed one.

“MFA on email covers us.” Phishing-resistant MFA is valuable and stops account takeover but the voice-clone attack often doesn’t touch your inbox at all. It targets the human decision downstream.

“Deepfakes mean video.” Voice-only is cheaper, faster, and needs far less source material. It’s the volume play, and volume is where the losses are.

“This is IT’s problem.” It is a finance-process problem that IT supports. The control lives in your AP workflow, your vendor onboarding, and your approval matrix, not in a firewall.

Here is a concrete, low-cost playbook for the single riskiest event, a supplier changing bank details.

For a generation, finance controls quietly outsourced verification to human perception: you knew a colleague’s face, you recognised a supplier’s voice, and that recognition stood in for proof. Generative AI has severed the link between perception and authenticity. What you see and hear is now a rendering, and renderings can be forged at near-zero cost.

The strategic takeaway is not “deepfakes are scary.” It’s narrower and more useful: verification has to move from the senses to the system. The controls that survive this era are the unglamorous, structural ones, name-IBAN matching, pre-registered channels, shared secrets, two-person approval, mandatory cooling-off. None of them require you to detect a fake. All of them make the fake irrelevant.

The callback isn’t dead because phones stopped working. It’s dead because the one thing it verified that the familiar voice on the line is who it claims to be, is exactly the thing an attacker can now manufacture. Rebuild the control around what they can’t.

Action Items To Do following things in next 30 days:

The organisations that lose money to this next year won’t be the ones without a detector. They’ll be the ones who still treated a phone call as proof.

It is a free weekly briefing you can read in under 10 minutes - covering how deepfakes are actually built, real deepfake fraud cases studies, honest detection tool reviews, and prevention protocols that work. Written independently. 1,000+ cybersecurity professionals, founders, and IT managers across SMBs already read it. Feel free to forward it to your friends who might benefit from this newsletter.

Have you seen something that didn't feel right? or Got questions you don't know who to ask? Whether you've encountered a suspected deepfake, want to understand your exposure, or just want to talk through what's happening in this space, I'm setting aside time for free 30-minute calls.