A client calls to move £180,000 from their ISA into cash. The voice sounds right. They know the account number. They pass the security questions. The instruction goes through. Three days later, the real client rings to ask why their portfolio has changed.
This is not a hypothetical. It is a pattern that is becoming structurally predictable. According to the Pindrop 2025 Voice Intelligence and Security Report, deepfake voice fraud attempts grew by 300% year on year and jumped from roughly one attempt per month to seven attempts per day across 2024. The technology to clone a voice from a few minutes of audio is now freely available. The barrier to attempting this kind of fraud is close to zero.
The FCA and Bank of England have jointly named frontier AI as material operational risk. If your firm has not yet reviewed how client instructions are authorised over the phone, that review is overdue.
The gap is in your workflow, not your voiceprint scanner
The instinct is to reach for a technology fix: a deepfake detection tool, a biometrics layer, a new authentication platform. Some of these have genuine value. But the deeper vulnerability is not a technology gap. It is a workflow gap.
Most firms rely on individual staff judgement under pressure. A client services associate takes the call, completes the standard security questions, notes that the voice sounds a little different, but the answers are right, and the client seems impatient. The system says proceed. So they proceed.
Fraudsters do not need to fool your entire compliance framework. They need to fool one person, once, in a moment of reasonable doubt. The Pindrop data underlines this: attackers are trained to exploit service-oriented culture, using partial correct information and social pressure to move quickly past hesitation. The vulnerability is the gap between “something feels slightly off” and “here is a structured reason to escalate.”
The question is not whether your staff can detect a deepfake. It is whether your workflow gives them anywhere to go when they are not sure.
What a redesigned verification workflow looks like
The framework here is sequential and does not require AI detection software to work.
Step one: map the high-risk actions. Not every client interaction carries the same risk. Changing a bank account, liquidating a position above a defined threshold, updating an address ahead of a withdrawal, adding a power of attorney — these are the actions that warrant a second gate. Identify them explicitly. Write the list down. Make it short.
Step two: require a second gate for high-risk actions. A second gate means a second person or a second channel, not a second question asked by the same person in the same call. That could mean: a callback to a pre-registered number, a confirmation sent to a verified email address and requiring a reply, or a brief hold while a colleague reviews the instruction. The point is that no single person, in a single real-time interaction, can authorise a high-risk action alone.
Step three: give staff a low-friction escalation path. “I’m going to need to verify this through our standard process for instructions of this kind” is a complete sentence. Staff should not need to accuse the caller of fraud to trigger additional checks. They need a script and a policy that makes escalation the default for certain actions, not an exceptional intervention that requires personal courage.
Step four: own the three-team problem. The Pindrop research frames this as a shared ownership question. Security defines the risk thresholds. Operations designs the workflow. Customer experience measures the friction. If those three conversations are not happening together, the workflow will either be too porous (operations optimising for speed) or too heavy (compliance optimising for coverage without operational buy-in). Someone in the firm needs to hold all three.
What Consumer Duty adds to the picture
Consumer Duty does not make this easier. Firms must act to deliver good outcomes, which includes protecting vulnerable clients from fraud. At the same time, adding friction to legitimate instructions creates its own harm, particularly for clients who find it difficult to navigate multiple verification steps.
The answer is proportionality. A low-value, routine instruction through an established channel does not need a second gate. A high-value, unusual instruction from an unfamiliar number does. The firm’s job is to map that distinction clearly in advance, not to adjudicate it case by case in real time.
The FCA’s current direction of travel, as signalled in recent consultations, is outcomes-focused rather than prescriptive. That means firms have more discretion, but also less cover if something goes wrong and the workflow design is inadequate. Document the decisions. Write down why the thresholds are set where they are. If there is a fraud attempt and the workflow held, you want a paper trail that shows the design was intentional.
A practical starting point
If you do not have a written list of high-risk instruction types and a defined second-gate process for each of them, that is the first thing to build. It does not require external software. It requires an afternoon with your head of operations and a compliance lead, a short policy document, a staff briefing, and a review date.
The firms that handle this well will not necessarily have the best voice authentication technology. They will have the clearest internal agreements about what happens when something feels off.
If your firm is working through what this looks like in practice, a discovery call is a sensible place to start.