A recent report cited by Agenorix says that 25% of Americans have received an AI-generated voice call. That is not evidence of a confirmed fraud in every case, and the source should be treated with appropriate caution. It is, however, a useful signal of exposure.
The practical question for a security leader is not whether an employee can hear that a voice is artificial in a calm demonstration. It is whether they can reliably verify a trusted voice while handling urgency, authority, a familiar caller ID and a live business process. The second question is whether the organisation can prove what happened afterwards.
For a regulated firm, a policy that says “verify the caller” is not the same as a tested control. It does not show whether a relationship manager challenged a supposed client, whether an operations colleague stopped a beneficiary change, or whether a service desk agent followed the approved route during a password reset.
The attack is a workflow problem
Voice fraud is often discussed as a question of detection. In practice, the failure usually occurs at the point where a plausible conversation meets an authority-sensitive workflow.
Consider a call from someone who sounds like a senior executive and asks for an urgent payment approval. Or a known client requesting a change to settlement instructions. Or a private client asking for an exception to a normal authentication step. The employee may know the policy. The difficulty is applying it when the caller sounds familiar, supplies credible context and frames delay as a business risk.
The Arup incident remains a useful reference point. In early 2024, a finance employee wired about 25 million US dollars after a video call involving AI-deepfaked recreations of the company CFO and colleagues. Arup’s CIO, Rob Greig, described the incident as technology-enhanced social engineering. No systems were compromised and no data was affected. The initial contact was a phishing email before the deepfake call, a reminder that voice fraud can be one stage in a broader sequence.
Deloitte’s 2024 poll of more than 1,100 C-suite executives found that 25.9% said their organisation had experienced one or more deepfake incidents targeting financial or accounting data in the prior 12 months. That is a specific survey finding, not a measure of all fraud, but it puts the control question firmly inside financial risk management.
Pindrop’s 2025 Voice Intelligence & Security Report records a 1,300% rise in deepfake voice-fraud attempts in 2024. This is single-vendor telemetry, so it should not be treated as a universal market estimate. It does, nevertheless, provide category validation, particularly for firms whose contact-centre and client-servicing processes depend on voice.
Verification alone is not enough
FinCEN Alert FIN-2024-Alert004, issued on 13 November 2024, warns financial institutions about generative-AI deepfake fraud and executive-impersonation vishing. The alert endorses live audio and video verification, while also warning that attackers can generate synthetic audio and video responses to live prompts. It is advisory guidance, not a binding rule.
That distinction matters. A challenge question can be useful, but it is not a complete control if the attacker can respond in real time. The control has to include the route used to verify the request, the separation of duties, the permitted approval channel and the point at which the employee must stop the interaction.
The BankSafeOnline guide is a further reminder that voice cloning is now part of the public fraud conversation. Awareness material has value, but awareness is not assurance. A board, insurer or regulator needs evidence that the organisation’s own people and workflows were tested against an authorised scenario, with the result recorded and the remediation owned.
What a useful test should examine
A meaningful exercise should not ask employees to identify a synthetic voice in isolation. It should test the decisions around the voice.
A controlled assessment might examine:
- payment approval and urgent transfer requests;
- beneficiary or settlement-instruction changes;
- password resets and account-recovery requests;
- VIP servicing and exceptional client instructions;
- escalation to a second approver or an independent callback;
- recording, case management and evidence retention.
The question is not simply whether a person was fooled. It is where the workflow permitted the request to progress, which control was missed, whether the escalation route was usable and whether the event would be reconstructable six months later.
This is also relevant to regulatory planning. DORA has applied since 17 January 2025 and Article 26 mandates Threat-Led Penetration Testing under TIBER-EU, explicitly targeting people, processes and technologies, with ethical boundaries for social engineering. For UK firms, the Payment Systems Regulator’s reimbursement regime has applied since 7 October 2024 and requires banks to reimburse victims up to £85,000 per Faster Payments scam claim. The PSR’s framework makes prevention a balance-sheet issue as well as a conduct and control issue.
The Vishing Readiness Assessment
DeepBlocker’s Vishing Readiness Assessment is a fixed-scope, authorised engagement lasting two to three weeks. An AI red team places controlled vishing calls against agreed members of staff and tests the workflows that matter to the firm.
The deliverable is a board-ready evidence pack. It records which scenarios were attempted, who was fooled or resisted, where the workflow broke, which controls operated as designed and what remediation should happen next. It is designed for discussion with the board, the insurer, internal audit and the relevant regulator.
The entry price is around £9,500. The aim is not to promise a detection percentage or to create a dramatic demonstration. It is to establish, with bounded testing and sealed evidence, whether your people and phone workflows would survive an AI voice-fraud attempt.
Book an assessment if you need to replace assumptions about voice verification with evidence from your own operating environment.