Back to Blog

When a Cloned Voice Asks Finance to Pay, What Should Your Rule Say?

A familiar voice is no longer proof of who is calling. A two-channel rule, a payment-change table and a hypothetical example for owners.

Two identical cream rotary telephones on a dark navy desk. The left one's handset is off the hook and its line cord lies loose on the desk, cut and connected to nothing. The right one rests in its cradle, its braided brass cord running out of frame. Headline: A familiar voice is not proof. Subline: Call back. Two channels. Credit: Dr. Jonah Tebaa.
Direct answer

When a Cloned Voice Asks Finance to Pay, What Should Your Rule Say?

When facing a cloned voice requesting payment, Dr. Jonah Tebaa advises establishing the Two-Channel Rule: any request that moves money, changes bank details, or releases data must be confirmed through a second channel initiated by staff using existing internal records. Under this three-tier framework, new bank details require a callback and a 24-hour payment delay. In his hypothetical example, 30 callbacks a year at 5 minutes each take 2.5 hours, set against one assumed loss of USD 18,000.

On a Friday evening, a finance manager's phone rang. It was the owner's voice, asking her to pay a supplier's new bank account before the weekend. The owner was on a flight and had called no one.

The call that sounded exactly right

I should be plain about that scene: it is a hypothetical I constructed for illustration. It names no real firm, person or bank, and it is not a client story. I use it because every detail in it is ordinary. The voice was familiar. The tone was a little hurried, as the owner's often is. The request was reasonable, because suppliers do change banks. And the deadline, before the weekend, left no room to think.

For most of the history of business, a recognised voice was a genuine control. If you knew the person, hearing them was strong evidence that they were the one speaking. That is no longer safe to assume. Voice and video imitation is now cheap and convincing enough that a short clip of a person can be enough to produce a passable copy. I will not quote figures or name products, because the specifics change month to month. The general point is stable: recognition alone cannot be trusted.

The firm's problem is that its habits were built when recognition worked. Staff obey a familiar voice because they were taught to, and they were right to be taught it.

What a recognised voice used to prove, and what it proves now

When I raise this with owners, the first answer is usually, "I would know my own boss's voice." Perhaps. But the question is not whether you would know it on a good day, with time to listen. It is whether a clerk finishing her week, hearing a hurried voice she has heard a hundred times, will stop and doubt it. A control that depends on someone's judgement in the worst moment is not a control.

So I stop asking whether the voice is real and look at the request instead. Whoever is asking, a few features should raise the bar:

  • Money or data moves. A transfer, a changed account, a password, a customer list.
  • Urgency. Before the weekend, before the bank closes, before the meeting.
  • Secrecy. "Do not mention this to anyone yet."
  • A new channel or new details. A different number, a fresh thread, a changed bank account.

None of these proves a fraud attempt. All of them together are the shape of one. And there is a human side that matters as much as the money. The clerk who delays a payment because the owner sounds irritated risks a scolding. The clerk who pays risks the cash. A good rule takes that choice away from her, so that she is following a procedure instead of betting her standing on a hunch.

The Two-Channel Rule

The rule is one sentence: any request that moves money, changes bank details or releases data is confirmed through a second channel that you start, using contact details already in your own records. Never confirm on the number or thread the request arrived on, since that is the channel the imitator controls. I scale the checking to the risk, in three tiers.

TierWhat it coversThe rule
Tier 1: money or credentialsAny new or changed bank account, any transfer above a set limit, any password or access requestCall back on a number from your own records, never the number or thread the request came on. New bank details wait 24 hours before the first payment.
Tier 2: data or documentsCustomer lists, contracts, staff recordsConfirm through a second channel or a second named person before sending.
Tier 3: information onlyQuestions, updates, general requestsNo check needed. The rule should not slow ordinary work.

The tiers, the 24-hour wait and the quarterly drill below are my working judgement from advising firms, not research findings. I would also ask your own bank what confirmation procedures it offers, and add them to yours. The rule does not make fraud impossible. It makes a perfect imitation insufficient, which is the most a procedure can honestly claim.

Build it in one afternoon

Nothing here needs new software. It needs a document and some agreement. Work through this list:

  • List every request type that can move money or release data: bank-detail changes, urgent transfers, payroll changes, customer lists, password resets.
  • Build a callback directory. For your top ten suppliers, your approvers and your bankers, record one verified phone number each, in a document that does not live in e-mail or WhatsApp threads.
  • Write one sentence in the owner's name: "I will never ask you to skip the callback." Staff then have cover.
  • Set the delay. New or changed bank details wait 24 hours and one callback before the first payment.
  • Agree a challenge question or private phrase between the owner and the finance lead, known to two people only and rotated each quarter.
  • Run one drill a quarter. Send a harmless fake request from a number nobody knows. Reward the person who calls back, and report the result to the team without blame.

Now the cost. These are hypothetical numbers for illustration, for a 60-person Beirut import distributor, and both inputs are assumptions I chose to be round. I make no claim about how often such attempts occur.

  • Assume 30 bank-detail change requests a year, and a callback takes 5 minutes.
  • Time spent: 30 x 5 = 150 minutes. 150 / 60 = 2.5 hours a year.
  • Assume one miss costs the next payment, USD 18,000.
  • Comparison: 18,000 / 2.5 = USD 7,200 protected for each hour of checking.

Change the inputs to your own and the conclusion rarely moves: checking costs hours, a miss costs a payment.

Your own AI must never become the cloned voice

There is a responsible-deployment turn here. If your firm runs a voice or chat assistant that contacts customers or suppliers, it must say it is an AI (I made that case in an earlier post on disclosure). It must never ask for or accept changed bank details or credentials. And it must hand any such request to a named person who uses the callback directory.

Then teach your customers and suppliers the same rule. When they know what a real call from you will never ask, a real call becomes recognisable by what it leaves out.

Monday: spend 30 minutes writing the callback directory for your ten largest suppliers, and tell finance in writing that no instruction, however familiar, replaces the callback.

Related evidence: Audio deepfake technology, also called voice cloning, uses artificial intelligence to generate speech that convincingly mimics a specific person, and the article records cases in which scammers used an impersonated executive's voice to direct a money transfer. (the reference overview of audio deepfakes and voice cloning)

The FBI warns that criminals use AI-generated audio, which it calls vocal cloning, to impersonate public figures or people close to the victim in order to obtain payments, and it advises agreeing a secret word or phrase to verify identity. (the FBI's public warning on AI vocal cloning used to request payments)

Frequently asked questions

Can someone really copy my voice well enough to fool my own staff?

Assume so. Voice and video imitation is now cheap and convincing enough that a short clip of a person can be enough for a passable copy. I will not quote figures, since they change quickly. The safe position is that recognition alone is no longer proof, so the request needs checking, not the voice.

Does a callback rule slow my team down on ordinary payments?

Only on the risky ones. Tier 3, information only, needs no check at all. Tier 1 applies to new or changed bank details, transfers above your set limit and credential requests. In my hypothetical example, 30 callbacks a year take 2.5 hours in total. Routine payments to known accounts carry on as before.

What if the owner is travelling and genuinely needs a payment made fast?

Then the owner is reachable through a number in your records, and the callback takes five minutes. If the owner truly cannot be reached, the payment waits. That is why I ask the owner to put in writing, ahead of time, that nobody will ever be asked to skip the callback. Staff then have cover.

Do I need special software to detect fake voices or video?

Not to start. The rule works because it does not depend on spotting a fake. A written procedure, a callback directory and a 24-hour wait make a perfect imitation insufficient. Detection software may have a place later, but I would not rely on it in place of the procedure, and I would ask your bank what it offers.

Written by Brian, Dr. Jonah Tebaa's AI partner, on his behalf.