If a deepfake tricked our employee into sending a wire, will the bank reimburse us?

Usually not. Commercial wire transfers are governed by UCC Article 4A rather than the consumer protections of Regulation E. Under Article 4A, if the bank and the customer agreed on a commercially reasonable security procedure and the bank accepted the payment order in good faith and in compliance with that procedure, the loss generally falls on the customer. Critically, a payment order authorized by a deceived but genuine employee using genuine credentials is not an unauthorized order in the way a stolen-credential attack is — the instruction really did come from inside the company.

Does our cyber insurance cover deepfake-induced transfers?

Often only partially, and the relevant coverage is usually the crime policy rather than the cyber policy. Losses where an employee voluntarily transferred funds after being deceived fall under social engineering or fraudulent instruction coverage, which is frequently a sublimit far below the main policy limit and typically conditioned on the insured having followed a specified verification procedure. Many claims are reduced or denied not because deepfakes are excluded but because the required callback or dual-authorization step was skipped.

Our policy is to call the executive back to verify. Is that still enough?

Voice callback alone no longer is. The control assumed an attacker could not reproduce a known voice, and consumer-grade cloning has removed that assumption — and video calls with synthesized participants have removed the visual version of it too. A callback still works if it goes to a number from an internal directory rather than one supplied in the request, but the verification value now comes from the out-of-band channel and the pre-known contact record, not from recognizing the voice.

What control actually stops this attack?

Controls that do not depend on recognizing a person: dual authorization by two named individuals in the payment system itself, a shared secret or challenge phrase never transmitted over the requesting channel, mandatory delay windows on new or changed payee accounts, and a standing rule that no payment instruction is executed through a channel where it originated. Urgency and secrecy are the constants in every version of this fraud, so a policy that treats an urgency-plus-secrecy request as an automatic escalation catches attacks the technology cannot detect.

Are there disclosure or notification obligations after a deepfake fraud loss?

Possibly several, and they run on short clocks. A material loss may trigger securities disclosure obligations for a public company. If the attacker also accessed systems or personal data, state breach-notification statutes apply on their own timelines. Vendor and customer contracts frequently carry incident-notice clauses. Building this into the incident response plan before an event matters, because the first forty-eight hours will be consumed by attempted funds recall.

Deepfake Voice Fraud 2026: Who Pays When AI Impersonates Your CFO

The Attack Did Not Get Smarter — The Verification Got Weaker

Business email compromise has followed the same script for a decade: an urgent request from a senior executive, an instruction to keep it confidential because of a pending deal, a new payee account, and a deadline that leaves no room to check. The control most companies adopted in response was equally simple — call the executive and confirm.

That control rested entirely on an assumption that has quietly expired: that an attacker could not produce the executive's voice. Voice cloning from a few seconds of public audio is now a consumer product, and real-time synthetic participants on video calls have followed. The consequence is not that a new category of fraud appeared. It is that the single verification step standing between a payment request and an executed wire stopped working, in most organizations, without anyone updating the policy that relies on it.

How the Loss Actually Gets Allocated

Risk: Assuming the bank absorbs the loss

RISK

Commercial payments run under UCC Article 4A, not the consumer protections of Regulation E. Where the bank and customer agreed on a commercially reasonable security procedure and the bank accepted the order in good faith following it, the loss generally stays with the customer — and a wire entered by a real, deceived employee using real credentials is a much weaker case for the customer than a credential-theft attack.

Risk: Treating the cyber policy as the answer

RISK

Fraudulent-instruction and social-engineering losses usually sit in the crime policy, often at a sublimit far below the headline limit, and typically conditioned on the insured having actually performed a specified verification step. The gap between the loss and the sublimit is the company's, and the condition is where claims get reduced.

Risk: A written policy no one can prove was followed

RISK

Both the Article 4A analysis and the insurance claim turn on documented practice. A verification policy that exists in a handbook but leaves no record of individual verifications performed provides nothing to point at when the insurer asks what happened on the specific transfer.

Mitigant: Dual authorization inside the payment system

MITIGATES

Requiring two named approvers to act in the banking platform itself — not to confirm to each other over a channel an attacker controls — is the control that survives synthetic media, because it never asks anyone to recognize a person.

Mitigant: Callbacks to directory numbers only

MITIGATES

A callback retains its value when it goes to a number retrieved from the internal directory rather than one supplied in the request. The verification comes from the pre-known contact record and the out-of-band channel, not from the voice that answers.

Mitigant: Cooling-off windows on new payee accounts

MITIGATES

A mandatory delay before the first payment to a newly added or changed bank account defeats the urgency the entire fraud depends on, and it costs a legitimate vendor a day rather than costing the company the transfer.

Detection Is the Wrong Place to Spend the Budget

The instinctive response to synthetic media fraud is to buy synthetic media detection. It is a reasonable layer and a poor primary control. Detection accuracy on real-world calls degrades with compression, packet loss, and unfamiliar languages, generation quality improves faster than detection does, and any detector deployed as the gate creates a false confidence that a passing call is genuine.

Process controls do not have that problem, because they never try to answer the question the attacker is manipulating. Dual authorization does not care whether the voice was real. A payee cooling-off window does not care how convincing the video call was. A rule that payment instructions are never executed through the channel they arrived on removes the attacker's channel entirely. These controls are unglamorous, cheap, and the ones an insurer and a court will actually ask about afterward.

A Treasury Control Review for the Deepfake Era

Re-read the security procedure in your bank agreement

Find what procedure you agreed was commercially reasonable, whether you are actually following it, and whether the bank offers a stronger option you declined. Declining an offered procedure materially weakens your position under Article 4A.

Check the social engineering sublimit, not the policy limit

Locate the fraudulent-instruction sublimit in the crime policy and read the verification conditions attached to it. That number and those conditions, not the headline limit, are what a deepfake wire loss recovers against.

Remove voice and video recognition from every control

Audit each verification step for a hidden assumption that someone will recognize a person. Replace with directory-sourced callbacks, in-system approvals, and pre-shared challenge phrases.

Log every verification, not just the policy

Record who verified what, through which channel, and against which directory entry. This record is the evidence for both the insurance claim and any argument with the bank.

Set a delay window on new and changed payees

Apply a mandatory hold before the first payment to any new or modified bank account, with a separate approval path. This one rule neutralizes the urgency the fraud requires.

Rehearse the first forty-eight hours

Funds recall, bank notification, law enforcement reporting, insurer notice, and any securities or breach disclosure all compete for the same first two days. Decide the sequence and the owners before an incident, not during one.

Frequently Asked Questions

Will our bank reimburse a deepfake-induced wire?

Usually not. Commercial transfers fall under UCC Article 4A rather than Regulation E's consumer protections. If the bank and customer agreed on a commercially reasonable security procedure and the bank followed it in good faith, the loss generally rests with the customer — and a wire entered by a genuine employee using genuine credentials is not unauthorized in the way a credential-theft transfer is.

Does cyber insurance cover it?

The applicable coverage is usually the crime policy's social engineering or fraudulent instruction section, not the cyber policy. It commonly carries a sublimit well below the main limit and conditions payment on having performed a specified verification step. Most reductions come from that condition, not from a deepfake-specific exclusion.

Is callback verification obsolete?

Voice recognition during a callback is obsolete. The callback itself still works when the number comes from your internal directory rather than the request, because the security then comes from the out-of-band channel and the pre-known contact record instead of from recognizing whoever answers.

Should we buy deepfake detection software?

As a supplementary layer, possibly. As the primary control, no. Detection accuracy degrades on compressed real-world calls, generation quality is improving faster than detection, and a detector at the gate manufactures false confidence in calls that pass. Process controls that never depend on identifying a person are more reliable and far cheaper.

What if the impersonated executive is our own CEO instructing an override?

That scenario is the fraud pattern, which is why the control has to be structural rather than discretionary. A policy stating that no individual — including the CEO — can bypass dual authorization, adopted by the board and communicated to executives in advance, removes the pressure the attack manufactures and gives the employee an unambiguous answer.

Do we have to disclose the loss?

Potentially under several regimes at once. A material loss may trigger securities disclosure for a public company, system or data access triggers state breach-notification clocks, and vendor or customer contracts often include incident-notice provisions. Map these in advance, because the first two days after an incident will be spent on funds recall.

Find AI Security and Governance Tools on RatedWithAI

RatedWithAI reviews AI security and governance platforms — including synthetic media detection, identity verification, and approval-workflow tooling that keeps payment authorization out of channels an attacker can impersonate.

Explore AI Legal & Compliance Guides