Deepfake Voice and Video Scams Against Companies: How to Defend

Deepfake Voice and Video Scams Against Companies: How to Defend

A company defends itself against deepfake scams by no longer treating a voice or a face as proof of identity. The control that works is a process control: confirm every request for money or sensitive data through a second channel, use a safe word agreed in advance, and cap how much any one person can pay alone. These three controls are cheap, do not depend on spotting the fake, and still work when the audio or video looks perfect.

Key takeaways

• Voice and video are no longer proof of identity. Anyone who approves payments needs a process that holds up even against a convincing imitation.
• Dual confirmation through an independent channel (calling back a number already on file) stops most fake-executive fraud.
• A safe word agreed outside digital channels gives a fast check for urgent, out-of-pattern requests.
• Payment approval limits define who can approve what, and turn a single order into something that needs two people.
• The weakest point is usually pressure and urgency. Training the team to slow down is worth more than trying to recognize the scam by ear.

How does a deepfake scam against a company work?

A deepfake is synthetic content that imitates the voice or appearance of a real person. For the fraudster, the technique itself is rarely the goal. The goal is to get someone in finance, administration or customer service to carry out an action that looks legitimate: make a payment, change a Pix key (Brazil's instant payment system), send a document, grant access.

The pattern repeats. The fraudster picks someone with the power to execute, gathers public information (the names of the director, the accountant and suppliers, clips from videos and interviews), creates a sense of urgency and asks for secrecy. The voice or video imitation is there to remove the doubt the person would have about a text message from an unknown number.

It is important to separate what is documented from what is hypothesis. It is documented that voice cloning and face-swap tools exist and are accessible. How much audio is enough for a convincing imitation varies by tool, and your company's defense should not bet on that. The safer premise is this: assume a good imitation is possible and design the process so that it is not enough.

In Brazil, the channel is often WhatsApp, where a short audio from a supposed director arrives from a new number with a copied profile photo. Quick video calls also appear, with deliberately poor quality that explains away glitches in the image. The low quality is part of the disguise.

Which fraud scenarios should a company rehearse?

It is worth training the team with concrete situations. Three scenarios cover much of the risk.

Illustrative scenario 1, the fake director asking for a Pix transfer: late in the day, a finance analyst receives a message from a new number saying it is the director, who has changed phones. Then an audio arrives in his voice, asking for an urgent payment to a supplier "to close a contract today" and asking for discretion. She has access to internet banking, and the pressure comes from someone she sees as an authority. Without a clear rule, the tendency is to comply.

Illustrative scenario 2, the video call with a swapped face: someone in administration is invited to a short video meeting with a "partner" and a "lawyer". The image freezes, the audio cuts out, and the partner asks that a service provider's bank details be updated. No one on that call needs to be real for the scene to work. It only has to look plausible for a few minutes.

Illustrative scenario 3, the cloned voice calling support or reception: someone calls claiming to be a manager and asks for a password reset, a customer spreadsheet, or the verification code that just arrived by SMS. Here the damage can be a leak of personal data, which brings in LGPD (Brazil's data protection law) as well.

All three have something in common: the person acting cannot prove, in the moment, who is on the other end, and the urgency keeps them from leaving the channel the fraudster controls.

Dual confirmation: the simplest and most effective control

Dual confirmation means validating any sensitive request through a different channel from the one it arrived on, using a contact the company had before the request. If the request came as a WhatsApp audio, the confirmation is a phone call to the person's registered number, or an in-person conversation. Replying in the same conversation does not count, and neither does calling a number given in the message itself.

To work, the rule has to be specific. A suggested wording:

• Any off-routine payment request, change of bank details, customer list transfer or access reset requires confirmation through an independent channel.
• The contact used for confirmation comes from the internal records, never from the message received.
• The confirmation is logged: who asked, who confirmed, through which channel and when.
• Requests marked as "confidential" or "too urgent to confirm" are treated as a warning sign, not as an exception.

The last line deserves attention. Whoever runs the scam knows the confirmation step exists and tries to dismantle it with arguments about secrecy and haste. That is why the rule needs explicit backing from leadership: nobody will be reprimanded for confirming, even if the request turns out to be genuine. If confirming costs embarrassment, the team will skip the step.

Safe word: how to create and use one without complicating things

A safe word is a spoken password, agreed in advance among the people who can request or authorize sensitive operations. It is useful when confirmation through another channel is hard in the moment, for example an unexpected phone call or a video call.

Good practices so that it actually protects:

• Agree on it in person or in an already verified conversation, never by email or open messaging.
• Use a phrase that does not appear on social media, in videos, in interviews or in company documents.
• Change it periodically, and whenever someone leaves the company or the circle of trust.
• Do not use it in WhatsApp groups and do not write it in systems that many people can access.
• Define what to do if the person does not know the word: end the conversation and trigger confirmation through another channel.

There is an honest limit here: a safe word protects against imitation, but not against coercion or against a leak. That is why it complements dual confirmation and approval limits instead of replacing them. In families and small businesses, a useful variation is a second, distinct word that signals the person is under pressure, but this only works if the team actually trains on what to do when they hear it.

Payment approval limits: who can approve what

Approval limits set boundaries by amount and type of operation for each person or role, and define when a second approval is required. They are the control that limits the damage when the others fail. If a single person can move any amount alone, a well-executed scam only has to convince one person.

A simple structure that fits small and mid-sized companies:

• Define tiers, without needing to publicize them: small, recurring operations follow the normal flow, larger operations require a second approval, and out-of-pattern operations require approval from a partner or director regardless of amount.
• Separate who registers from who approves. Whoever adds or changes a Pix key or a supplier account should not be the one who releases the payment. This separation of duties is a classic access control principle, described for example in OWASP's authorization guidance.
• Configure the limits in the bank itself. Banks and payment platforms generally let you set limits per operation, per period and per user, and require dual approval in business internet banking. Check with your account manager what is available in your case.
• Treat a change to a supplier's bank details as a risk event: independent confirmation, a waiting period before the first payment, and a notice to whoever owns the supplier relationship.

The ideal design is one where no voice, however familiar, can move money alone. This also protects the people on your team: when the process requires two people, nobody carries the blame for a mistake alone.

Step-by-step: putting the controls in place

Rolling this out does not require new technology. It requires a decision, training and review. A practical sequence, following the map, limit, pilot, transfer method:

• Map: list who in the company can make payments, change bank details, reset access or send customer data. Note which channels those people receive requests through (WhatsApp, email, phone, meetings).
• Limit: write the dual confirmation rule, define the approval limits and adjust the limits in the banks. Reduce what each person can do alone, including access held by people who rarely use it.
• Agree on the safe word among the few people who authorize sensitive operations, in a verified meeting.
• Pilot: choose one real flow, for example out-of-pattern payments, and apply the rules for a short period. Record how many confirmations were made, where there was friction and what the team found confusing. An internal simulation, announced to leadership and coordinated with HR, helps test whether the rule holds up under pressure.
• Transfer: document the process, name an owner for the review and decide, based on the pilot, whether to expand to other flows, adjust or stop. To train the team, also read about what to teach first in AI training for teams, since recognizing synthetic content is part of basic literacy today.

One note on AI agents: if the company uses an agent on WhatsApp or in customer service, it should not have permission to release a payment, change a financial record or reveal data based on who claims to be the manager. This is solved with action limits and handoff to a human, a topic covered in guardrails for AI agents and in what not to delegate to AI in business. When the company transcribes WhatsApp voice messages with AI, remember that a transcript only converts voice into text. It does not verify who is speaking.

To structure a broader risk assessment, the NIST AI Risk Management Framework offers a reference for identifying, measuring and treating AI-related risks, including malicious use.

What should you do if the scam has already happened?

The first hour counts. A lean checklist:

• Contact the bank immediately through its fraud channels and ask for the transfer to be blocked or disputed. Ask about the recovery mechanisms available for Pix.
• Preserve the evidence: audios, messages, phone numbers, call screenshots, receipts. Do not delete conversations.
• File a police report and inform leadership and legal counsel.
• If personal data was exposed, assess your LGPD obligations with the data protection officer or legal counsel. The website of ANPD, Brazil's data protection authority, gathers official guidance on security incidents.
• Change passwords and access that may have been compromised, and review what the fraudster knew about the company to understand where the information came from.
• Run a review without hunting for someone to blame: where the process failed and which rule would have prevented the problem.

The right question after a scam is not who fell for it, but which rule allowed one person to decide alone.

Frequently asked questions

Can you spot a voice deepfake by ear alone? Not always, and that should not be your main defense. Some imitations have noticeable flaws, such as odd rhythm or a lack of natural background noise, but others convince even people who know the speaker well. That is why the focus should be on verifying identity through process, not impression.

Are small companies also targets? Yes, and often with fewer barriers, because the owner or a single employee concentrates the payment decisions. The measures in this article are designed to fit that reality: one confirmation rule, one safe word and well-configured bank limits already change the picture a great deal.

Do I need a deepfake detection tool? Detectors can help in some contexts, but they are imperfect and do not replace process. Before investing in technology, put dual confirmation, separation of duties and approval limits in place. If the company later evaluates a tool, treat it as an additional layer and test it on a real flow before relying on it.

Is the voice of employees and directors personal data under LGPD? A voice identifies a person and can therefore be treated as personal data. When used for biometric identification, it may call for even more careful handling. If the company records meetings or calls, define purpose, access and retention period with legal support, not least because those recordings can become raw material for imitations.

Sources

• NIST: AI Risk Management Framework
• OWASP: Authorization Cheat Sheet
• ANPD: Autoridade Nacional de Proteção de Dados