All insights

SECURITY & CONTROL

The Fideuram AI Voice Scam: How Enterprises Can Defend Against Deepfake Fraud

A CTO and CIO playbook for reducing executive impersonation risk with transaction controls, independent verification, and tested response plans.

AI deepfake fraud prevention and payment controlsExecutive security playbookFor CTOs, CIOs, treasury leaders, security teams, and finance executives

Finance leaders independently verify a high-value transfer through a dual-approval workflow to reduce AI voice-cloning fraud risk.

DIRECT ANSWER

The short answer

AI deepfake fraud prevention should not depend on employees recognizing synthetic media. Organizations should require independent identity verification, transaction-specific approvals, separation of duties, risk-based payment holds, and a tested response process so a convincing voice or video cannot authorize a high-risk transfer on its own.

Key takeaways

  • Treat a familiar voice, video, or message as a request signal, not proof of identity or payment authority.
  • Use independent verification, transaction-specific approvals, separation of duties, and risk-based holds.
  • Deepfake detection can support triage, but should never replace treasury controls or incident readiness.

When a familiar voice is not enough

Reuters reported on September 25, 2026 that an impersonation scam had targeted Fideuram’s then-chairman, Paolo Molesini, beginning in February. According to sources cited by Reuters, a WhatsApp message impersonated Intesa Sanpaolo CEO Carlo Messina and a subsequent call used an AI-replicated voice to impersonate a senior lawyer. Sources put the transfers at approximately EUR 95 million, with roughly EUR 53 million recovered and EUR 36 million still missing at the time of reporting. The banks declined to comment. These are attributed reporting details, not a bank-confirmed technical investigation or an updated account of recovery. Reuters report, September 25, 2026 (via MarketScreener).

The sequence matters: executive impersonation, urgency, a familiar intermediary, and a request routed through a legitimate business function. These are established social-engineering techniques. AI voice cloning can make the story more convincing, but the organization’s authorization design determines whether a persuasive impersonation can become a payment. For CTOs and CIOs, the key question is not whether every employee can identify a synthetic voice. It is whether an unverified communication can trigger an irreversible, high-impact action.

Design controls around the transaction

A call, video meeting, email, or chat may initiate a payment request; it should not approve the payment by itself. High-value transfers should be created and authorized in a governed treasury or ERP workflow, with an auditable record of the requester, reviewers, beneficiary, amount, and release decision. This separates the identity signal from the authority to move funds. A person who sounds like an executive may still be an impostor, while the payment system continues to enforce the same control requirements.

Six controls that make impersonation harder to monetize

1. Verify through a pre-established independent channel. For high-risk requests, use contact details maintained in a trusted corporate directory or approved procedure. Do not call a number supplied in the suspicious message, and do not continue verification in the same chat. A second channel is useful only when it is independently established.

2. Require separation of duties. Use maker-checker controls so the person preparing a transfer cannot release it alone. Set approval thresholds according to amount, beneficiary risk, jurisdiction, and deviations from normal behavior. Apply the policy to every role, including the CEO, chairperson, and CIO. Any emergency exception should require a separate approver, a documented reason, and a post-event review.

3. Bind approval to the payment details. Show the beneficiary, account, amount, currency, purpose, and recent changes in the approval interface. Avoid generic prompts such as “Approve transaction?” Where the platform supports it, use a cryptographically protected approval tied to the specific transaction. Account login or a one-time code may confirm access, but does not necessarily demonstrate that the approver reviewed the exact recipient and amount.

4. Put friction around beneficiary changes. A new payee, changed account, unusual destination, or altered payment procedure should trigger additional review and, where business needs allow, a cooling-off period. Maintain trusted beneficiary records and alert treasury when a change and a high-value transfer occur close together.

5. Detect anomalies and pause before release. Treasury systems can flag unusual amounts, geographies, timing, channels, or approval patterns. Risk signals should route a payment to a human review or hold state, not make an unsupported accusation. Deepfake classifiers can contribute evidence, but should not be the sole control: they can miss convincing fakes, flag legitimate media, and become less reliable as methods change.

6. Harden identity and communications. Protect executive email, collaboration platforms, and payment administration with phishing-resistant authentication where supported, least privilege, monitored account changes, and strong recovery procedures. Limit unnecessary access to payment workflows and monitor suspicious sign-ins, forwarding rules, and newly registered devices.

A practical control matrix for technology leaders

A control is useful only when it has an owner, an observable outcome, and a response path. CTOs should work with treasury, finance, security, and business leadership to map each risk to a system control and evidence that it is operating. The matrix below is a starting point; thresholds and approval requirements should be set through the organization’s risk and regulatory process.

Implement in stages, then exercise the response

Begin with a short discovery across treasury, finance, security, identity, and executive support. Map how payment instructions arrive, who can create and approve them, which systems can release funds, and where beneficiary details are maintained. Identify high-value and hard-to-reverse paths first. Then close obvious gaps: remove single-person release paths, establish trusted callback details, add review for new beneficiaries, and make approval records searchable.

Next, connect risk signals to operational actions. Define which events require an alert, a hold, a second approval, or a bank callback. Ensure that logging captures enough to investigate without retaining unnecessary sensitive content. Set owners and service expectations for alerts so a high-risk payment does not wait in an unattended queue. Test the controls in a safe environment and confirm that legitimate urgent payments still have a governed route.

Finally, rehearse the first hour of a suspected incident with treasury, security, IT, legal, communications, and executive leadership. Staff should know who can pause pending payments, contact the sending and receiving banks, preserve messages and call records, and activate incident response. Tabletop exercises should combine channels: a message followed by a call, a familiar voice, a new beneficiary, and pressure to keep the request confidential.

Make trust verifiable

The Fideuram case is a reminder that AI can make impersonation more persuasive, but the deeper issue is relying on identity signals that were never designed to authorize high-value transactions. A resilient organization verifies identity independently, separates request from approval, binds authorization to payment details, monitors for anomalies, and rehearses recovery.

FIX Intelligence, part of FIX Solutions JSC, helps organizations assess AI-related risks and design secure workflows, governance, and operational controls around production systems. A focused review can map where AI-enabled impersonation intersects with payment, identity, and approval processes, then prioritize controls according to business impact. The goal is not to promise that every fake can be detected; it is to ensure that a convincing voice is still not enough to move money.

A practical control matrix

Risk signalControl responseEvidence to review
Impersonated executive requestIndependent verification using a pre-registered channelVerification record and channel used
Single-person payment releaseSeparation of duties and risk-based approval thresholdsApprover logs and exception reports
New or changed beneficiaryEnhanced review and a cooling-off period where appropriateBeneficiary-change audit trail
Unusual transfer patternRisk-based hold and human escalationAlert response time and disposition
Suspected fraud in progressBank recall procedure and incident responseTime to contact banks and preserve evidence

Sources and further reading

Frequently asked questions

Can deepfake detection tools prevent executive payment fraud?

They can provide a useful risk signal, but should not be the sole authorization control. Detection can miss convincing media or flag legitimate communications. High-value payments still need independent verification, transaction-specific approval, separation of duties, and a clear escalation path.

How should employees verify an urgent payment request from an executive?

Follow the organization’s established payment procedure and verify through a pre-registered, independent channel, such as contact details in a trusted corporate directory. Do not rely on a number or link supplied in the request, and do not approve a transfer solely because the voice or video seems familiar.

What should a company do if it suspects an AI-enabled payment scam?

Pause pending transfers where possible, contact the sending and receiving financial institutions immediately, and activate the incident response process. Preserve messages, call details, payment instructions, approval records, and relevant account logs. Notify treasury, security, legal, and leadership according to the incident plan.