A vishing call used to be easier to classify. It might arrive as a generic robocall, a poorly disguised request for credentials, or an obvious attempt to pressure a member of staff.
That is no longer a safe operating assumption. A caller can now present as an executive, a supplier, a relationship manager or a member of the IT team. The request may be familiar. The language may be plausible. The voice may belong to someone the recipient knows.
Recent explainers on AI voice cloning describe the same practical problem: the technology is not the fraud by itself. It is an accelerant for existing social-engineering workflows, including payment requests, data access and account recovery. One overview describes how cloned voices are used to make an urgent request feel genuine, while another sets out how voice cloning can be combined with familiar deepfake warning signs and verification steps.
For a regulated firm, the question is therefore not whether an employee can spot a strange recording. The question is whether the organisation can complete a sensitive workflow safely when the person on the phone sounds right.
The control failure is usually behavioural
Consider a beneficiary change, a payment approval, a password reset or a VIP client request. Each workflow may have documented controls. There may be a callback requirement, dual approval, a secure portal or a recorded confirmation.
Under pressure, however, controls can become informal. A senior executive asks for an exception. A supplier claims that a payment is holding up a transaction. An IT caller says an account must be reset before close of business. A familiar voice supplies just enough context to make the request feel internally consistent.
This is where authority, urgency and familiarity interact. Authority reduces the perceived need to challenge. Urgency compresses the time available to verify. Familiarity makes the interaction feel lower risk than it is.
A policy can say, “verify independently”. It does not tell you whether a particular employee will do so when the caller sounds like the chief executive and knows the name of the current transaction. That is an evidence gap, not necessarily a training gap.
The evidence is moving from theory to risk management
The issue has been recognised by financial-crime authorities. FinCEN Alert FIN-2024-Alert004, issued on 13 November 2024, warns financial institutions about generative-AI deepfake fraud and executive-impersonation vishing. It endorses live audio and video verification, but also warns that attackers can generate synthetic responses to live prompts. Verification remains necessary, but it cannot be treated as a complete defence when the verification channel itself can be manipulated.
The same alert is important for another reason. It frames the problem as a typology involving people, processes and technology. That means the relevant control test is not simply a model benchmark. It is a controlled examination of what happens when a real person receives a credible request inside a real workflow.
Deloitte’s 2024 poll of more than 1,100 C-suite executives found that 25.9 percent said their organisation had experienced one or more deepfake incidents targeting financial or accounting data in the prior 12 months. That scope matters. The finding is not a claim that every employee is being targeted every day. It is evidence that the risk has entered executive-level reporting.
Pindrop’s 2025 Voice Intelligence & Security Report recorded a 1,300 percent rise in deepfake voice-fraud attempts during 2024. This is single-vendor telemetry, so it should not be treated as a universal market measurement. It is nevertheless a useful signal that synthetic voice is being observed in operational fraud environments, not just discussed as a future possibility.
Deloitte’s May 2024 projection puts the potential scale in financial terms: US generative-AI-enabled fraud losses could reach 40 billion US dollars by 2027 in its aggressive scenario, up from 12.3 billion in 2023. Deloitte also presents a more conservative scenario of around 22 billion dollars. These are projections, not reported losses, but they explain why boards and insurers are asking for evidence of preparedness.
What a proper test should establish
A useful assessment should not ask staff to identify an artificial voice in isolation. It should test whether the workflow holds when the voice is embedded in a plausible scenario.
That means defining authorised scenarios such as:
- a payment approval requested by a senior executive;
- a beneficiary or bank-detail change presented by a supplier;
- a password reset requested by an apparent IT colleague;
- a sensitive disclosure requested during VIP servicing.
The assessment should record what happened, which control was invoked, whether an independent channel was used, who approved the action and where the process allowed an exception. It should distinguish a person who challenged the request from a person who noticed something unusual but proceeded anyway.
The output should be sealed evidence: the scope, authorisation, call records, decision points, affected workflows, control failures and remediation priorities. That is more useful to a board, insurer or regulator than a general statement that staff have completed awareness training.
A practical route to readiness
The DeepBlocker Vishing Readiness Assessment is a fixed-scope, authorised engagement delivered over two to three weeks. Our AI red team places controlled vishing calls against selected staff and workflows, then produces a board-ready evidence pack showing who was fooled, which safeguards worked and where the process broke.
The entry price is around £9,500. The purpose is not to embarrass individual employees or claim perfect detection. It is to answer a narrower and more consequential question: would your payment, servicing, access or approval workflow survive a convincing impersonation attempt?
For firms operating under DORA, this kind of testing also fits the regulation’s emphasis on threat-led penetration testing against people, processes and technologies, with ethical boundaries defined for social engineering. In the UK, the Payment Systems Regulator’s reimbursement regime, live since 7 October 2024, requires banks to reimburse victims up to £85,000 per Faster Payments scam claim. Prevention therefore has a measurable balance-sheet dimension as well as a conduct and resilience dimension.
The voice may be synthetic. The decision is real. Subscribe to the DeepBlocker intelligence for evidence-led guidance on vishing readiness.