Deepfake Voice Fraud: The Voice on the Phone Wasn’t Theirs

A familiar voice used to give us a reason to trust a call. Deepfake voice fraud turns that instinct into a way to steal money. An attacker can sound like a colleague or business partner and ask for a payment without ever gaining access to the company’s systems. If the person receiving the call treats recognition as verification, the transfer can be authorized through the normal process.

The same problem extends to video. Seeing the chief financial officer and other colleagues on a call feels more reassuring than reading an email request. Today, their apparent presence cannot prove they made the request.

That is an uncomfortable shift for any business that moves money. We have spent years teaching people to spot suspicious links, protect passwords and question unusual messages. Now the suspicious instruction may arrive in a voice they have heard hundreds of times.

At a Glance

• Deepfake voice and video fraud can persuade an employee to authorize a payment without an attacker breaking into company systems.
• In 2024, Arup lost $25 million after an employee acted on instructions during a video call with people who appeared to be senior colleagues.
• A Swiss business owner lost several million Swiss francs in 2026 after callers used AI to imitate a trusted business partner.
• A LastPass employee stopped an almost identical attempt in April 2024 by noticing the request arrived through an unfamiliar channel and refusing to act under manufactured urgency.
• A payment request must be verified through contact details and a channel established before the request arrived. A second approver must make an independent check.

How did a fake video call lead to a $25 million transfer?

In 2024, a finance employee at global engineering firm Arup joined a video call that appeared to include the company’s chief financial officer and other senior colleagues. The people on the call were deepfakes. The employee followed the payment instructions and transferred a total of $25 million to the criminals.

Arup’s Chief Information Officer, Rob Greig, told the World Economic Forum that none of the company’s systems had been compromised and no data had been affected in the incident. People had been deceived into carrying out transactions they believed were genuine.

That distinction matters. A firewall could be working. Multifactor authentication could be enabled. The employee could be signed into the correct banking platform with their own authorised account. The loss occurs when a fraudulent instruction is accepted as genuine and the legitimate user makes the payment.

After the incident, Greig tried making a real-time deepfake video of himself with open-source software. It took him about 45 minutes. He said the result was not particularly convincing, but the speed of the experiment surprised him. He also observed that incidents like Arup’s happen “more frequently than a lot of people realize.”

The lesson I would take into a finance or leadership meeting is straightforward: we cannot make payment approval depend on whether the person on a call looks convincing. The Arup employee believed they were hearing from several senior people at once. That apparent group agreement became part of the deception.

Not every attempt succeeds, and the difference is instructive. In April 2024, an employee at LastPass, the US password-management company, publicly confirmed they had received a series of calls, texts and a voicemail carrying an audio deepfake of the company’s own CEO, delivered over WhatsApp rather than any channel the company normally used for that kind of request. The employee recognized the hallmarks of the attempt, including the forced urgency, ignored the messages and reported the incident to the internal security team. LastPass said the attempt had no impact on the company. The instinct that stopped it is the same one this checklist is built on: an unfamiliar channel and manufactured urgency are exactly the signals a person can catch, if given a moment to pause.

Why should smaller businesses pay attention?

Arup’s loss is large enough to make headlines. The mechanism works at a smaller scale, wherever one person can persuade another to release funds.

In January 2026, a business owner in Switzerland received a series of calls over roughly two weeks from people presenting themselves as a known business partner. The callers used AI to manipulate their voice and made the proposed payments sound like part of a confidential international deal. The owner authorized several transfers to a bank account in Asia. By the time the fraud was discovered, the total loss was several million Swiss francs. The precise amount has not been made public. Biometric Update, which first reported the case, said it went undetected until after several transfers had already gone through.

Notice the pace of this case. It was not necessarily one alarming call followed by an immediate transfer. Repeated contact gave the fraudsters time to build a believable explanation and make the next request feel like a continuation of the last one.

A smaller company may have fewer people between a plausible request and a bank transfer. Staff often know the owner’s business partners personally. That familiarity can make an unexpected call feel less suspicious, particularly when the caller knows the names and circumstances surrounding a real transaction.

What is deepfake voice fraud actually exploiting?

These scams exploit the way people use context to decide whom to trust. The voice sounds right. The caller knows about a deal. The request appears to come from someone with authority. On video, other familiar faces may appear to agree. Each detail reduces the recipient’s willingness to interrupt and check.

Attackers also benefit from ordinary working pressures. A finance employee may be managing a payment deadline. An owner may be travelling. A supplier may be waiting. A senior colleague may say a transaction must remain confidential. Under those conditions, asking for a separate confirmation can feel like delaying legitimate work.

I would want staff to understand that the interruption is part of legitimate work. Moving money to an account specified during a call is a consequential action. Taking a few minutes to verify the instruction protects the business and the employee who has been asked to act.

This is why the phrase “I recognized their voice” can no longer close an investigation into a payment request. A voice is useful context. It is not an independent identity check. The same applies to caller ID, a familiar video image and a reply within an existing message thread. An attacker may impersonate those signals or use a compromised channel to make the request appear routine.

The FBI’s guidance on AI-assisted financial fraud advises people to end a suspicious call and contact the purported caller through a number found independently. That advice is sound for businesses because it changes the source of trust: the recipient chooses a previously verified route back to the real person.

Financial regulators are treating this as a systemic risk, not an isolated case. In November 2024, the Financial Crimes Enforcement Network issued an alert naming AI-generated deepfake media as a specific method attackers use to circumvent identity verification, authentication and payment due-diligence controls at financial institutions.

Where do payment controls fail?

Many organizations already require two approvals for large transfers. That helps, but the second signature has value only if the second person checks the underlying instruction independently.

If the first approver says, “I was on a video call with the CFO,” and the second approver signs because that sounds convincing, the organization still has only one unverified account of the request. Both approvals rest on the same call.

A sound process separates three decisions:

1. Is the payment legitimate? Check the invoice, agreement or business reason against records the company already holds.
2. Is the destination correct? Compare the beneficiary and account details with previously verified records. Investigate every change, however plausible the explanation sounds.
3. Did the authorized person actually request it? Contact that person through an established number or separate company channel. For a significant transfer, the independent approver should perform or witness a separate verification.

The payment amount should influence the level of scrutiny, but a low amount does not make a changed bank account harmless. An attacker may test a process with a smaller request before attempting a larger one. A request split into several payments deserves attention for the same reason.

What should happen if a payment has already been made?

Time matters. The employee or owner who spots the problem should contact the bank immediately using a known number, explain that the transfer may be fraudulent and ask what can be done to stop or recall it. They should then alert the organization’s finance and security leads so other pending payments can be checked.

Keep the evidence: call details, meeting invitations, messages, invoices, account numbers and a record of who approved each step. The business may need these to investigate the attempt, report it to the appropriate authorities and determine whether any accounts or communications were also compromised.

Reporting must be encouraged even when someone has only a small doubt. Staff are less likely to raise a concern promptly if they expect blame for having taken the call or begun the approval process. A report made before the transfer is released may prevent the loss altogether.

FAQ

The payment-verification checklist

The aim is to make one behavior routine: verify the instruction separately before money moves. Finance staff, business owners and executives should all be able to use this checklist without needing to decide whether a voice sounds fake.

• End the call and call back. Use a number already held in your company records. Never use a number supplied during the call, in a text, or in the email making the request.
• Check through a second channel. Use an established company chat account or speak to the person in person. Replying to the same email thread or continuing the same call is not a separate check.
• Pause when someone creates urgency. “Do it now” is a reason to complete verification promptly, not a reason to skip it. The same applies to requests for secrecy.
• Investigate every change in payment details. Confirm a new bank account or beneficiary using contact details already on file, even when a familiar voice explains the change.
• Treat voice and video as context, not proof. A caller sounding or looking like a colleague does not establish that the colleague authorized the transaction.
• Require an independent second approver for wire transfers. That person must check the payment and its destination separately. Confirmation from the same apparent caller does not count.
• Report a concern immediately. Staff should alert finance and security if any part of a request feels wrong, even if they cannot explain why. If funds have moved, contact the bank without delay.

The hardest moment to apply this checklist is when the caller sounds exactly like someone you trust. That is the moment it was designed for.

Teams can use transaction monitoring and anomaly detection to help identify unusual amounts, destinations or payment patterns. Netsurit has discussed how AI can support fraud detection, while our approach to AI risk assessment considers how these tools fit into wider business decisions. Neither removes the need to authenticate a person’s instruction before releasing funds.

If your organization wants to test whether its payment and incident-response processes would stand up to an impersonation attempt, speak with Netsurit about managed cybersecurity.