Fake Voices, Real Losses: How Deepfake Voices Are Changing Customer Support Fraud

What is deepfake customer-support fraud and why should merchants be concerned?

Deepfake customer-support fraud is a fast-evolving tactic where scammers take small audio clips or publicly available recordings of a real customer, use AI voice-cloning tools to recreate that person’s voice, and then call a merchant’s customer-support line pretending to be them.

The scammer might request a refund, change card details, move subscriptions, or ask for account access, all the time sounding eerily convincing because it mimics tone, cadence and emotional cues from the original speaker.

This isn’t science fiction anymore: the National Cyber Security Centre (NCSC) and industry bodies such as UK Finance note that AI voice cloning and live deepfakes are an emerging fraud risk, and although UK Finance says the issue is being monitored rather than identified as a major driver of losses, authorities are increasingly alert to the threat.

If merchants regard a familiar voice as verified identity, they risk paying out refunds, unlocking accounts or changing payment details to those of the attackers. As a result, they are not only faced with financial loss, but also a decrease in customer trust, repeated chargebacks, operational headaches handling disputes, and possibly compliance issues where incorrect changes cause data protection problems.

Because the attack includes social media, voice messages and other everyday content, victims are often unaware that their voice was even taken until it’s used against them.

How are scammers collecting the voice samples they need?

Many voice-cloners need only a few seconds of clear audio to produce realistic results. Scammers gather that audio from a surprising range of sources including public social video posts and livestreams, voicemail greetings, WhatsApp and other messaging app voice notes, YouTube clips, TV or podcast clips, or even short clips recorded during company sales or onboarding calls.

Sometimes the clip is taken automatically at scale using scraping tools but at other time it’s elicited – for example a scammer might ask someone on social media to “say hi” as part of a simple, friendly request. Media investigations have shown how quickly a voice can be cloned and weaponised.  For example, in one BBC investigation, a reporter used an AI-cloned version of their own voice to test several UK banks’ voice-ID systems and was able to get through some of the strict authentication checks which are purposefully designed to verify real customers.

Because many people use the same phrases for example, “Yes, that’s fine”, or “My account number’s…”, and because caller ID and incoming numbers can be faked, a cloned voice paired with a fake number can now be used to bypass frontline checks. That combination is what makes the tactic extremely attractive to criminals: it exploits both social familiarity and the trust that humans place in spoken identity.

How exactly are cloned voices used to commit merchant fraud?

Attack methods vary, but there are a few patterns that merchants are seeing more often. One common approach is a cloned voice calling customer support to report an “unauthorised” purchase and then push for an immediate refund. Another is a caller posing as the genuine account holder and requesting changes to stored card details, billing addresses, or delivery information. Another involves urgent-sounding requests, for example, replacing a payout bank account or withdrawing stored credits, during which, the convincingly familiar voice is used to discourage staff from performing deeper verification checks.

Deepfake audio can also be used to create emotional urgency. A caller who sounds panicked or upset can push for a quick refund or account update, relying on staff instinctively trying to help rather than stopping to run additional checks.

International cases show just how damaging these attacks can be when a cloned voice is used along with forged emails, compromised credentials and weak internal controls. In one widely reported incident, manipulated audio and video in a video-conference call led to a multi-million-dollar corporate money transfer. Scammers created deepfake versions of several senior executives and used them in what looked like a routine internal meeting. The employee on the call saw familiar faces, heard familiar voices, and had no reason to question the instructions being given. As a result, the pay out was approved before anyone realised that the entire video call (apart from that one employee) had been faked. The company ultimately lost more than $20 million, showing just how convincing deepfake impersonation can be when it mimics people we’re used to trusting.

The practical reality for merchants is that these scams can look completely legitimate at first glance. If your team relies heavily on voice recognition or keeps verification steps light, it becomes much easier for attackers to slip through. And because deepfake calls can be replayed or refined, with scammers repeatedly testing different scripts until something works, the threat isn’t just a single attempt, but a persistent, ongoing risk.

How can customer support teams spot the red flags of a deepfake call?

Of course, human listeners are still useful, but increasingly unreliable against high-quality clones. Therefore, it’s increasingly important to train staff to recognise behavioural and procedural red flags rather than rely on the sound of a voice alone. Useful warning signs include:

• Inconsistent account details – mismatched addresses, incorrect answers to security questions, or information that does not align with the customer’s usual history.
• Unusual urgency or pressure – callers who push for immediate action, attempt to rush staff, or try to create a sense of urgency that feels out of character.
• Resistance to authentication – callers who argue with, delay, or try to bypass the usual verification steps.
• Scripted or unnatural vocal cues – emotion that sounds rehearsed, unnatural pauses before speaking, or phrases that resembles a public social-media clip.

Operationally, it helps to build in multi-factor verification for any sensitive account changes. This could include sending a one-time passcode to the customer’s registered mobile or email, asking knowledge-based questions that can’t be guessed from public information, or using a mandatory callback to the phone number already on file.

It’s also worth having a documented call-escalation process so that any request to update payment or payout details is automatically reviewed by another team member. For higher-risk actions, such as large refunds or account transfers, you might introduce a short delay and send a confirmation email to the registered address before proceeding. These behavioural and procedural checks offer far more protection than voice recognition alone.

What technical and process investments make the biggest difference?

There’s no single silver bullet, but an effective layered defence combines people, processes and technology:

• Strengthen identity checks – require at least two independent verification elements. For example: something the customer knows (eg. a password or pin) and something they have (eg. a mobile phone for a one-time passcode) before approving any sensitive account changes.
• Monitor for unusual patterns – flag accounts showing sudden request changes, high refund activity, or repeated calls from different numbers. Linking device data, IP information, timing, and payment behaviour can help surface synthetic-voice attacks that look normal at first glance.
• Use anti-manipulation tools where appropriate – voice-analysis tools and call-risk scoring services can detect signs of synthetic audio or a false caller ID.
• Tighten controls on outbound changes – for example, switching a payout bank account could require signed documentation or a micro-deposit that the customer must verify before the change is applied.
• Train and empower staff – provide agents with simple scripts, clear red flags, and the authority to pause or escalate a call without penalty, whenever something doesn’t feel right.

Alongside the above measures, merchants should also plan for what happens after an attack. Even with strong controls in place, some attempts may still slip through the net. That’s why it’s important to have solid processes for disputes, record-keeping and chargeback responses; these can significantly reduce the financial impact when fraud does occur.

What should merchants do today to reduce the risk, and what should they expect going forward?

Merchants can take a series of immediate and medium-term actions to defend against synthetic-voice fraud.

In the short term  (over days to weeks) – review and tighten verification processes for refunds and payment-detail changes, enforce mandatory second-factor authentication for high-value requests, implement a callback policy for suspicious calls, brief all agents with concrete examples and a clear escalation flow, and log call audio (with appropriate privacy safeguards) to so that further investigations can be made if necessary.

Over the medium term (weeks to months) – integrate call-risk scoring, improve analytics to detect unusual refund patterns, and review the new-customer sign-up process to prevent attackers from capturing fresh voice samples. Clear communication with customers is particularly crucial: advise them not to post short vocal clips containing personally identifying phrases and remind them of official channels for high-risk changes, such as email confirmations or account portals.

Looking ahead, AI voice cloning will continue to improve and become more accessible, making reliance on voice as a single authentication factor increasingly risky. Industry guidance already reflects this reality by urging organisations to treat synthetic-voice incidents as a mainstream fraud vector rather than simply a niche threat. Therefore, with fraud tactics evolving rapidly, merchants can’t afford to be reactive. They need to stay vigilant, invest in proactive safeguards, and adapt quickly to emerging threats in order to reduce losses and maintain the trust that underpins their brand.