The Boss's Voice on the Phone Asks for a Transfer, and It's a Deepfake

Late afternoon, a finance staff member gets a WhatsApp call from a new number. The voice sounds exactly like the company owner: "I'm in a meeting, can't talk long. Please transfer this to the vendor now, it's urgent, I'll explain later." The tone is rushed, a little irritated, and very convincing.

Scams like this used to be easy to spot because the voice sounded off. Now AI can copy a voice, and even a face on a video call. What's left as a safeguard is procedure.

This isn't a movie plot

Arup, 2024. In an interview with the World Economic Forum (February 2025), the Chief Information Officer of Arup, a UK engineering firm, described what happened in early 2024: an employee transferred company money after a video call with "senior management" who all turned out to be deepfakes. The loss was US$25 million. He called it technology-enhanced social engineering: no systems were breached and no data was affected, the person was deceived. Afterward, he tried making a real-time deepfake of himself with open-source software, and it took about 45 minutes. The result wasn't very convincing, but he was surprised how much could be done in so little time.

Indonesia, 2025. Indonesian police arrested a scammer who used deepfake videos of President Prabowo, Vice President Gibran, and then Finance Minister Sri Mulyani to offer fake government aid. Victims were directed to a WhatsApp number, then asked to pay an "administrative fee" of Rp250,000 to Rp1 million (about US$15 to US$60). Police recorded 11 victims, and the suspect admitted to running it since 2020. The targets were the general public, not companies, but it shows the same technique is already used by local scammers.

Why this old trick works better than ever

The pattern is old: authority (the request comes from a boss), urgency (it has to be now), and secrecy (don't tell anyone yet). What used to foil it was a voice or face that felt off. AI removes that clue. So what needs to change isn't employees' ability to guess whether a voice is real, but rules that don't depend on guessing.

The FBI, in its public warning about fraud using generative AI (December 2024), suggests a few relevant things: create a secret word or phrase to verify identity, hang up and call back on a number you look up yourself, listen closely to tone and word choice, and limit recordings of your voice and face online.

A simple procedure for the finance team

These are my suggestions, and all of them can be put in place this week:

1. An out-of-the-ordinary transfer is never processed on a phone call, voice note, or video call alone. It needs confirmation through a second channel.
2. Call back on the number already saved, not the number that called or a number sent in a chat.
3. An internal spoken passphrase for urgent requests. Agreed in person, never written in a chat.
4. Two people approve transfers above a set amount, through a channel different from the original request.
5. Changes to a vendor's bank account must be re-verified through the official contact that already existed before.
6. Delaying is allowed. Agree on it from the top: the real boss won't be angry that staff checked first.
7. A short drill. Tell the finance and admin team about cases like these, then run a simulation once.

The principle is the same as the split in At What Point Should Humans Still Check AI and Automation Results: transferring money is an action a person decides on, with checks. Even OpenAI's Dots agent asks its user to take over for money transfers, as I cover in ChatGPT Now Has Dots, and What's Interesting Isn't How Smart It Is but Who Holds the Access.

If the transfer has already gone out

Move fast. Contact the bank right away to ask for the funds to be blocked or held, gather evidence (numbers, recordings, chats, transfer receipts), report it to the police, and also report it to OJK's Indonesia Anti-Scam Centre (IASC), the financial regulator's scam reporting service. After that, check whether any company account was also compromised, because fake requests sometimes come from real accounts that were taken over. How to secure them is in If Your Company's Google Account Got Hacked Tomorrow Morning, What Would You Lose?.

Another scam that exploits trust, this time through AI answers on Google, is covered in Asked AI Overview for a Call Center Number, Got Connected to a Scammer.

Wrapping up

We can no longer rely on our ears and eyes to be sure who's speaking. What we can rely on is a procedure that doesn't care how convincing the voice is: call back, two approvals, and a culture that lets staff check first.

If you'd like to tidy up the payment approval flow in your company, tell me about the current flow first.