When the voice sounds right: Stopping AI impersonation fraud
A call that sounds exactly like your CFO used to be reassuring. Now it is just a call.
It is late on a Friday before a long weekend. The CFO calls, sounding rushed but unmistakably like herself, and asks you to push out an urgent payment to a new supplier before the bank cutoff. She mentions the deal by name. There is even a short video meeting where a familiar face nods along. Everything checks out, except that none of it was real.
Generative AI has made voices and faces cheap to fake, and finance teams are an obvious target because they are the people who can move money. If you read our guide to business email compromise, this is the same attack in a new disguise. The defense is the same too: a process that does not depend on anyone recognizing a fake.
Key takeaways
- A familiar voice or face is no longer proof of identity. AI tools can imitate both well enough to pass a busy moment.
- Spotting fakes by ear is a losing game. The quality keeps improving, so the defense has to be a process, not a talent.
- Verify through a channel you already trust, such as a callback to a number you have on file, never one the requester gives you.
- Never change payment details because of a call, a message, or a meeting. Bank detail changes are where many of these schemes end.
- Dual approval holds even when one person is fooled, because a convinced employee still cannot release funds alone.
What deepfake payment fraud looks like
The term covers any scheme that uses AI-generated audio, video, images, or text to impersonate someone you trust and get money moved. In practice, it shows up in a few forms.
- Cloned voice calls. A call or voicemail that sounds like an executive, a colleague, or a long-standing vendor, asking for a payment or a change to payment details.
- Video meetings. A call where one or more participants are synthetic, used to make an unusual request feel approved by people you recognize.
- Polished written requests. AI-written emails and messages that no longer carry the typos and awkward phrasing people were trained to look for.
- Combinations. An email sets up the story, then a call "confirms" it, so each channel seems to vouch for the other.
What federal agencies have said
In December 2024, the FBI issued a public service announcement warning that criminals are using generative AI to commit fraud on a larger scale and make their schemes more believable. It describes AI-generated audio used to impersonate public figures and personal contacts to get payments, AI-generated video used in real-time video chats, and AI-written text that removes the usual warning signs of a scam.
A month earlier, FinCEN, the Treasury bureau that combats financial crime, issued an alert on fraud schemes involving deepfake media. Its main focus is fake identity documents used to open accounts at financial institutions, and it reports an increase in suspicious activity reports describing deepfakes since 2023. It also notes that deepfakes can be used to impersonate trusted people in scams aimed at businesses. The message from both is the same: realistic synthetic media is now a routine part of fraud, not a novelty.
Why finance teams are the target
These schemes lean on the same pressure points as any payment fraud: authority, urgency, and secrecy. A request appears to come from someone senior, it has to happen now, and it should stay quiet because of a confidential deal or a sensitive supplier issue. Timing helps the attacker too. End of quarter, year-end, and holiday weeks are when approvers are out and people are trying to clear their queues.
Here is the uncomfortable part. A lot of awareness training still tells people to listen for a robotic tone or watch for odd lip movement on video. The FBI does mention tone and word choice as possible clues, and they are worth noticing. But that advice has a short shelf life, because the tools keep improving. A team that relies on spotting the fake will eventually meet one it cannot spot.
Warning signs worth pausing for
None of these prove a request is fake. Each is a reason to slow down and verify.
- Urgency combined with secrecy, especially a request to skip the usual approvals or keep others out of the loop.
- New or changed bank details, for a payee you have paid before or a supplier you have never heard of.
- A switch of channel, such as a request that moves from email to a phone call or a messaging app.
- Reluctance to be called back, or an excuse for why the normal number will not work today.
- Small oddities in tone, timing, or wording that do not quite fit the person.
A verification routine your team can adopt this week
The routine below does not need new software. It needs agreement from leadership that everyone follows it, including the executives being impersonated.
- Call back on a number you already have. Use the number in your own records, never one given in the request, and hang up first if the call came to you.
- Agree a code phrase for payment requests. The FBI suggests a secret word or phrase to confirm someone's identity. The same idea works inside a finance team, as long as the phrase is never shared by email or chat.
- Never change payment details on the strength of a call, message, or meeting. Bank detail changes should go through a separate, documented process.
- Treat urgency as a reason to slow down. A real deadline survives a five-minute callback. A fake one often does not.
- Make it safe to say no. Leaders should say openly that they expect to be verified, so nobody feels they are insulting the CFO by checking.
Controls that hold even when someone is fooled
The routine above depends on people. The stronger defense is a system where one convinced person cannot complete the payment alone.
That is the job of the maker-checker process: one person initiates a payment and a different person approves it. On Xtrm, approval policies let you designate approvers for transfers to a linked bank, wallet-to-wallet moves, mass payment files, and linking new banks, with an optional second approver for files and amount-based approval levels for transfers. Approvers are notified when a payment file is submitted, and it is not processed until they approve. A deepfake might fool the person who receives the call. It now has to fool a second person too, through a separate step. Our help center guide to approval policies walks through the full setup.
The second control takes away a frequent goal of these schemes. When payees maintain their own payee record behind their own login, a caller asking finance to "just update the bank details" is asking for something finance does not hold. Our guide to payout controls covers how these checks fit together.
If you think you have been targeted
- Stop. Do not send anything further, and do not reply to the original message.
- Call your bank right away if money has moved. The sooner they know, the better the chance of stopping or recalling the payment.
- Keep the evidence: emails, phone numbers, call times, meeting links, and any recordings.
- Report it to the FBI's Internet Crime Complaint Center at ic3.gov, even if the attempt failed.
- Tell your team. The same people may be targeted again, sometimes with a different voice.
FAQs about deepfake payment fraud
What is deepfake payment fraud?
It is a scheme that uses AI-generated audio, video, images, or text to impersonate someone the victim trusts, such as an executive, colleague, or vendor, in order to get a payment sent or payment details changed. It is an evolution of business email compromise that reaches beyond email to phone and video calls.
Can a deepfake voice fool someone who knows the speaker?
It can be convincing, especially on a short call under time pressure. That is why the defense should not depend on recognizing a fake. A callback to a known number, a code phrase, and a second approver all work regardless of how good the imitation is.
How should a finance team verify a payment request?
Confirm it through a channel you already trust, such as a callback to a number in your own records, and never one supplied in the request. Require a separate, documented process for any change to payment details, and use dual approval so no single person can release a payment alone.
Should we report an attempt if no money was lost?
Yes. Reporting attempts to the FBI's Internet Crime Complaint Center at ic3.gov helps investigators connect related cases. If any payment or account details were shared, tell your bank as well.
Conclusion: Trust the process, not the voice
Deepfakes change what counts as evidence. A familiar voice, a recognizable face, and a well-written email used to be good signs. Now they are just inputs, and any of them can be manufactured. The teams that hold up are not the ones with the sharpest ears. They are the ones where verification is routine, payment details cannot be changed by a phone call, and no single person can move money alone.
If you want to see how approval policies and payee-managed details would fit your payout process, we are glad to walk through it with you.
Last updated: October 19, 2026
Post by The Xtrm Team
Oct 19, 2026, 9:06:19 AM
