Picture a finance controller on a video call with the CFO. The face is right, the voice is right, they reference a live deal — so an urgent payment gets released. Except the CFO was never on the call. That's not a thought experiment any more: 62% of organisations have now faced at least one deepfake attack, and about 55% lost money to deepfake or AI-voice fraud in the past twelve months — with a mean loss above $280,000 per incident, and nearly one in five of those victims losing more than $500,000. AI-cloned voice and real-time video are now routinely used to authorise fraudulent wire transfers and bypass biometric authentication; deepfakes turn up in one in five biometric fraud attempts.
The curve is steep. Reported deepfake-fraud losses have passed $3.7 billion, and roughly 89% of that was booked in 2025 and the first half of 2026 alone. The FBI's IC3 broke out AI-enabled fraud as its own category for the first time in the 2025 report — 22,364 complaints and $893.3M in adjusted losses — and in H1 2026 a single threat dataset verified 821 attacks against 15,736 victims. Every published figure is an undercount, because most victims never report. The defining security shift of 2026 isn't a new malware family — it's that your people can no longer trust their own eyes and ears as proof of identity.
Why "train people to spot it" won't save you
The instinct is to teach staff to detect fakes — the tell-tale glitch, the unnatural blink. That advice is already expiring. Real-time generation is getting good enough that spotting fakes is a losing game, and you don't want your controls to depend on a stressed employee out-analysing a nation-state's toolkit at 4:55pm on a Friday. Seeing is no longer believing, so the defence has to assume the media is convincing and stop the fraud anyway. That means moving the control from human perception to process.
The defence is mostly process
The good news: the fixes are things a mid-market firm can implement without exotic technology. The core move is to break the request and the verification onto different channels:
- Out-of-band verification for money and identity changes. Any request to move funds or change bank details gets verified on a known, trusted channel — a callback to a number you already hold, never the number or link in the message. This single control defeats most deepfake-enabled fraud.
- Dual authorisation above a threshold. One person can be fooled; two people, verifying independently, is far harder to defeat. Make large or unusual payments require two sets of eyes.
- Remove "urgency + secrecy" as an accepted workflow. That combination is the attack pattern — "I need this done now, don't tell anyone until it's closed." Treat it as a red flag that triggers controls, not a reason to bypass them.
- Phishing-resistant MFA everywhere. So a cloned voice can't talk its way past a login or a helpdesk reset — the same identity-first foundation that stops credential abuse.
- A pre-agreed verification step for executive requests. A code word or a mandatory second channel for anything coming "from the CEO." Cheap, and remarkably effective.
Technology has a supporting role — deepfake-detection tooling, and content provenance/watermarking to verify genuine media — but detection is the backstop, not the plan. The plan is designing your money-movement and identity processes so that no single voice or face is ever sufficient on its own.
This compounds with your other AI-era risks
Deepfakes don't arrive in isolation. Synthetic identities are also used to infiltrate organisations — fake personas passing interviews, or feeding the pile of non-human identities already hard to police. The through-line is the same: verify explicitly, trust context over appearance, and never let a convincing signal replace a real control.
The bottom line
The camera and the phone stopped being proof of identity, and pretending otherwise is how firms lose seven figures to a single convincing call. The defence isn't smarter humans — it's process designed for a world where the media is fake: out-of-band verification, dual authorisation, no urgency-and-secrecy exceptions, and phishing-resistant MFA underneath. Building that resilience — and pressure-testing it before an attacker does — is central to our cybersecurity and resilience work.
Want to know whether a convincing fake could get a payment released today? Let's stress-test it.