The chair of Fideuram, the private banking arm of Intesa Sanpaolo, got a WhatsApp message that appeared to come from the group CEO, Carlo Messina. It asked for urgent help with an overseas transaction. A follow-up call used an AI-cloned voice imitating a senior lawyer. Believing the request was real, the chair told the finance department to make the transfers. €95 million went out, mainly to accounts in China and Hong Kong. Reuters reported it on September 25, 2026. About €53 million has been recovered, and the rest was converted to cryptocurrency.
The interesting part of this story isn't the AI. It's the process. A cloned voice convinced one person, and that person had the authority to move €95 million on the strength of a phone call.
Why would I fix the process before buying a detector?
I've been a CIO, so I know the instinct after an incident like this: ask the technology team which tool would have caught it. The better question is why one person could approve it alone. This was a finance-controls failure that happened to use AI.
Deepfake detection is a real and improving market. Modulate raised $25 million on September 28 and reports a 1.1% equal error rate on a public speech deepfake benchmark. Pindrop has screened contact-center calls for years. The roundup of deepfake detection tools on wetheflywheel.com covers who does what. But a detector is a probability, and attackers can tune against it. A rule that says "no transfer without a callback and a second approver" doesn't depend on catching the fake at all.
What are the four rules I'd put in place first?
If I were back in a CIO seat tomorrow, these would go in before any tool. They cost almost nothing, and none of them requires anyone to recognize a fake.
- Call back on a known number. Any urgent, confidential or unusual payment request gets verified on a number from the company directory, never one in the message. If the "CEO" can't take a callback, the payment waits.
- Two approvers above a threshold, no exemption for rank. In the Fideuram case, the instruction came from the chair. A rule that the most senior person can override protects you against everyone except the person attackers will impersonate.
- A challenge phrase for executives. Agree a phrase in person that never appears in email, chat or a meeting recording. A cloned voice can copy how someone sounds, but it can't know something that was never written down.
- Train for the pattern, not the technology. Urgency, secrecy and a request to bypass the normal process are the signature of this scam whether it arrives by email, WhatsApp, a call or a video meeting. Finance and executive assistants are the people to train first.
Why does process beat perception?
Endurance athletes learn this about fuelling. You don't eat when you feel hungry in a long race, because by the time you feel it, it's too late and your judgment is already off. You eat on a schedule you set before the start. Payment verification is the same. You don't verify when a request feels suspicious, because the best fakes won't feel suspicious. You verify on a rule you set before anyone calls.
When do detection tools earn their place?
When the volume is too high for people to check. A contact center taking thousands of calls a day needs automated voice screening. An onboarding funnel needs liveness and injection detection. A company whose executives are constantly impersonated online needs a takedown service such as Doppel. In each case the tool handles volume that no process could, and it flags what then needs a human decision.
The numbers say this is worth taking seriously. The FBI's 2025 Internet Crime Report tracked AI for the first time and counted 22,364 complaints and $893 million in losses, and it expects that to be an undercount. It's the same reasoning I apply to my own AI agents: decide in advance which actions need a second check, and don't rely on spotting the problem in the moment.
Sources
- Biometric Update, Italian bank loses $108M to AI messaging fraud (Oct 15, 2026), citing Reuters (Sep 25, 2026).
- TechCrunch, Modulate raises $25M (Sep 28, 2026).
- PYMNTS, FBI flags $893 million in AI-driven scams, on the FBI IC3 2025 Internet Crime Report.
Frequently Asked Questions
What is deepfake CEO fraud?
A scam in which criminals impersonate a senior executive, using an AI-cloned voice, a deepfake video or a convincing message, to get an employee to make an urgent payment or share credentials. It is the AI version of business email compromise. In the 2026 Fideuram case, a WhatsApp message posing as the group CEO and a call with a cloned voice led to €95 million in transfers.
How do you stop deepfake CEO fraud?
With a payment process that does not rely on recognizing a voice or face. Verify any urgent or unusual request by calling back on a number from your own records, require a second approver above a threshold with no exemption for seniority, agree a challenge phrase with executives, and train finance teams that urgency plus secrecy is the warning sign. Add detection tools after that, where call or meeting volume makes manual checks impossible.
Can you detect a cloned voice on a phone call?
Sometimes, with software. Detection vendors such as Pindrop and Modulate score calls for synthetic speech, and benchmark accuracy is high. A person listening cannot do it reliably, and McAfee found that three seconds of public audio is enough for an 85% voice match. Treat any detector as a prompt to verify, not as proof either way.
How common are AI scams against companies?
More common than reported. The FBI's 2025 Internet Crime Report was the first to track AI separately and counted 22,364 complaints and $893 million in losses, most of it investment fraud. The FBI expects that figure is an undercount, because many victims never report or do not realize AI was involved.
Need Expert Technology Guidance?
20+ years leading technology transformations. Get a technology executive's perspective on your biggest challenges.
