AI Voice-Cloning Fraud: How to Protect Your Firm

How an AI voice-cloning scam works

The mechanism is simpler than it sounds. An attacker collects audio samples of the person they wish to impersonate — a few minutes of public recordings, interviews, YouTube videos or social-media clips are enough — and uses them to train a voice-synthesis model. The result is a synthetic voice capable of delivering any text with the intonation, rhythm and phonetic detail of the original. The quality achievable today makes recognition by the naked ear practically impossible for anyone.

In the Fideuram case documented by FederPrivacy, the attack combined two channels: a written WhatsApp message — easier to fake — to lower the target's guard, followed by a phone call featuring the synthetic voice of a trusted lawyer to deliver the final push towards authorising the wire transfer. This combination is now the hallmark of the most sophisticated voice fraud: a text channel prepares the ground, a voice channel closes the scam.

The technical term is *vishing* (voice phishing), but in its AI-enhanced form it represents a qualitative leap beyond the old phone calls with a foreign accent. This is no longer about exploiting inattention: it is about exploiting trust in a familiar voice.

Why professional firms and SMEs are ideal targets

Large banks and listed groups have spent years building multi-factor authentication, escalation protocols and automatic audit trails. A five-person accountancy practice, or a thirty-employee manufacturing SME, has not. And that asymmetry is precisely what fraudsters look for.

Three characteristics make an Italian professional firm or SME particularly exposed. First: management of significant financial flows through oral, informal decision-making. The word of the owner, partner or trusted accountant carries enormous weight — an operational advantage that becomes a critical vulnerability when faced with a cloned voice. Second: access to highly sensitive third-party data — balance sheets, tax returns, payslips, court documents — making a successful breach even more valuable to an attacker. Third: deep, long-standing mutual trust between people who have known each other for years, which lowers the verification threshold further still.

As the Cybersecurity360 analysis on contemporary cyber risk highlights, today's risk does not originate inside the organisation — it spreads through trust relationships between interconnected parties. A professional firm is by definition a trust hub: the point where relationships between clients, banks, suppliers and public bodies converge. Whoever compromises it can reach all of them.

How to spot a fake AI-generated call

The honest answer is: often you cannot — and you should not fool yourself into thinking your ears will save you. The latest generation of synthetic voices passes perceptual tests even with trained listeners. That said, there are contextual signals that should always trigger an alarm, regardless of how authentic the voice sounds.

The first signal is artificial urgency: the request must be acted on immediately, there is no time to verify, something terrible will happen if you do not act at once. This pattern is identical in every fraud, digital or otherwise, because time is the greatest ally of anyone who wants to check and the greatest enemy of anyone who wants to deceive. The second signal is an unusual combination of channels: a request that arrives first on WhatsApp and then by phone, or vice versa, from numbers that do not match exactly with those stored in your contacts, is already an anomaly that cannot be ignored. The third signal is a request that falls outside normal procedure: if your partner has never asked you to authorise a wire transfer by phone without written confirmation, and today they do, the voice — however convincing — is not enough.

On the technical side, synthetic-voice detection tools exist, but they are still rarely deployed in the context of Italian professional firms and require integration into communications systems. The most practical and immediate defence remains procedural: deciding in advance what can and cannot be authorised verbally, before the fake call arrives.

Internal procedures to prevent AI voice fraud

Effective prevention does not start with a technology tool: it starts with the decision about what can never be authorised on a verbal request alone. That decision must be made before the call arrives, written into an internal protocol, communicated to the entire team and reviewed at least once a year. This is not red tape: it is the difference between demonstrating you did everything possible and having to answer for a loss.

A minimum anti-fraud protocol for a professional firm or SME has at least three levels. The first is a list of actions that can never be authorised by voice alone: wire transfers above a defined threshold, changes to client bank details, access to management systems, sharing of credentials. The second is a dual-channel procedure: for any request on that list, confirmation must arrive via a second, independent channel — not WhatsApp on the same phone, but a corporate email or a call-back to a verified number. The third is a tracked log: every authorisation must be recorded in a system that shows who approved it, when and on what basis, so that due diligence can be demonstrated if the matter is ever challenged.

This third point also carries the most direct legal implications. If an employee executes a wire transfer on the instruction of a cloned voice, liability falls on the firm or company owner if there are no documented procedures that restricted that possibility. As Cybersecurity360's analysis on cybersecurity governance points out, the most significant regulatory violations do not stem from the absence of formal controls, but from the inability to demonstrate ongoing, informed risk management — an excellent summary of what any firm seeking to protect itself actually needs. For a deeper look at how AI affects professional liability, see our article on verifying AI output in professional firms.

Voice-fraud prevention protocol: minimum checklist for firms and SMEs

• ✓
  Blacklist of verbal authorisationsWritten and defined: which operations can never be carried out on a phone or chat request alone.
• ✓
  Wire transfer thresholdMaximum amount authorisable without written confirmation on an independent channel, signed by the owner.
• ✓
  Dual-channel procedureFor every critical request: confirmation by corporate email or call-back to a verified number from your contacts.
• ✓
  Authorisation logA tool or register that tracks who approved what, with date and time, kept securely and retrievable.
• ✕
  Annual team trainingAt least one session per year with up-to-date examples of voice fraud and practical simulations.
• ✕
  Protocol updateAnnual review of the protocol at a minimum, or following any incident or reported attempt.

GDPR, AI Act and liability: what the rules say about voice fraud

AI voice-cloning fraud touches the European regulatory framework on at least two fronts. The first concerns the GDPR: voice is biometric data. When an attacker clones the voice of a lawyer or an accountant to commit fraud, they are unlawfully processing a special-category personal data. But the professional's own liability does not end there: if their firm has not adopted adequate technical and organisational measures to protect its communication channels and client data, it may face challenges even as a victim of a breach, not merely as a controller.

The second front is the EU AI Act, which has been progressively entering into force across the European Union. Voice-synthesis and voice-cloning systems used to deceive natural persons fall within prohibited or high-risk AI practices, but this does not automatically protect those who fall victim to them: the professional must still have done their part in terms of preventive measures. Also relevant here is Italian Law 132/2025, which introduces specific obligations for Italian professionals who use AI tools — a regulatory context that Ovia has examined in detail in the dedicated article on Italian Law 132/2025 and client disclosure obligations.

In short: the law alone is not enough to protect you if you do not have documented procedures in place. And documented procedures are not enough if they are not backed by tools that leave a verifiable trail of every decision and authorisation. This is the point where strategy meets technology — and where a professional firm can move from being exposed to being protected.

Tools for tracking and verifying call-based authorisations

The market now offers tools capable of automatically transcribing, summarising and archiving phone calls and meetings, linking each conversation to the correct client or matter. This is not only useful for productivity: it is essential for building a verifiable audit trail of every interaction — one that, if a dispute arises, demonstrates what was said, by whom and when.

A call management and transcription system integrated into the firm's daily routine turns every phone call into a searchable record: who called, when, what was decided, what actions followed. If a request for a wire transfer comes in by phone, the log exists. If it does not exist in the log, that absence is itself a detectable anomaly.

In parallel, an operational CRM with structured authorisation workflows makes it possible to define which actions require formal approval, and to record every authorisation step with date, time and user. This is not digital bureaucracy: it is about making visible and provable the due diligence that the law requires and that, in the event of fraud, can make the difference between a covered and an uncovered loss. It is the principle that Cybersecurity360 summarises as: making the invisible visible, connecting technical risk to its real business impact.