In short (TL;DR)
Deepfake fraud has reached the treasury departments of mainstream companies. According to data released by the FBI in April 2026, AI-related fraud losses reached $893 million in 2025, and in one high-profile case, a victim lost over $25.6 million after a video call in which every participant was a deepfake. The bad news: On average, humans can recognize a high-quality fake video only 24.51% of the time. The good news: Protection doesn’t depend on a sharp eye, but on five simple process rules that you can implement today.
How does deepfake invoice fraud work?
The classic invoice fraud relied on a fake PDF and a hasty victim. AI added three new weapons:
- Voice cloning: three seconds of public audio (interview, webinar, social media video) is enough to call with the driver's voice and order a "quick payment".
- Deepfake video call: Scammers invite employee to a meeting where the „CFO” and „colleagues” are all AI-generated. This is how one company lost $25.6 million with 15 transfers.
- Perfect fake invoices: AI creates an invoice that mimics the supplier's design, numerical logic, and tone: the only change is the bank account.
How big is the risk really?
The numbers come primarily from the US, but the pattern is global. According to the FBI's Internet Crime Center In 2025, 22,364 complaints were received involving AI fraud, with a total loss of $893 million. Finance departments are the first target, because speed, amounts, and routine intersect there. And while separate deepfake statistics for Estonia have not yet been published, the fraudulent letters and „manager calls” here work according to exactly the same pattern.
Why doesn't "I recognize my driver's voice" work?
Because you don't. In controlled tests, people can identify a high-quality deepfake video with an average accuracy of 24.5%, or less than one in four. More than half of managers admit that their employees have not received any training on deepfakes. Vigilance is necessary, but insufficient as a defense strategy: a fraudster only needs one busy afternoon, you need to be infallible every day.
Five process rules that stop fraud
- Callback rule: Every payment or account change request that comes in via phone, video, or text is verified on another channel using a previously known number. Not the number the scammer wrote down.
- An IBAN change is always a red flag: a change in the supplier's bank account is never taken from the invoice or letter, but is confirmed by the supplier's official contact.
- Double confirmation: Payments above an agreed limit (for example, 1,000 euros) are confirmed by two people. Rushing only overrides the rule in favor of the fraudster.
- Code word with management: if the „manager” asks for an extraordinary payment, ask an agreed-upon question that cannot be answered on the public internet.
- Train your team with examples: one 30-minute session with real cases will raise awareness more than any memo.
How do e-invoices and AI contribute to protection?
The fake invoice lives in an email attachment. If your invoices are sent via an authenticated e-invoice channel, for example, in the Estonian operator network or In the Peppoli network, the sender's identity is verified by the network and the "same invoice, new IBAN" trick loses ground. It won't completely replace PDFs, but every supplier that switches to e-invoices is one less email fraud channel.
The second layer is a machine that doesn't know the boredom of routine: AI compares the details, amounts and rhythm of each incoming invoice with the previous one and highlights any deviations before payment day. We describe how it works in the article AI-based invoice fraud detection.
Summary
Deepfake fraud is a real business risk in 2026: losses are measured in hundreds of millions and the human eye alone cannot detect high-quality counterfeits. Protection is a process: callbacks on a known number, separate checks for IBAN changes, double confirmation, a codeword and training – and a technical layer in the form of an authenticated e-invoice channel and AI anomaly detection. None of these steps cost more than one successful fraud.
Try Bilnex's free invoice processing, where AI checks the details of each incoming invoice.
FAQ
How common will deepfake scams be in 2026?
The FBI's Cybercrime Center recorded 22,364 AI-related fraud complaints in 2025, with a total loss of $893 million; the data was published in April 2026. The most famous case is the video call deepfake, which resulted in $25.6 million being transferred.
Can a human recognize a deepfake?
Bad: In controlled tests, humans only detect high-quality deepfake video with an average accuracy of 24.5%. Therefore, protection must rely on a process, such as calling back a known number, rather than recognition.
What is the most effective single protective measure?
Callback rule: Every payment or bank account change request is confirmed by a previously known contact on the other channel. This stops both voice clones, deepfake videos, and fake emails because the scammer doesn't check your other channel.
How do e-invoices reduce the risk of fraud?
In an authenticated e-invoicing channel, such as the Peppoli network, the sender's identity is verified by the network, so the trick of a fake PDF in an email attachment and a changed IBAN does not work. Every supplier that switches to e-invoicing means one less email fraud channel.