The phone call you cannot trust anymore sounds exactly like your business partner. Modern voice technology can learn a person's voice from a very short sample of public audio — a voicemail greeting, a video on your company page, a podcast interview — and then hold a live, two-way conversation in that voice. This week's Scale with AI episode explains why a familiar voice is no longer proof, the November 25 opt-out deadline sixteen days out, the alleged $129 million deepfake-CEO fraud prosecutors charged on November 5, and the three-rule system protected businesses actually run.
The Shift
For all of business history, a familiar voice was proof. That rule is now finished. A voice is no longer evidence; it is a claim — and the same is true of a face on a video call and a perfectly written email with the right invoice attached. This matters because most business training teaches teams to spot the scam by looking for the flaw: the misspelled word, the strange address, the robotic voice. This scam has no flaw to spot. The email reads perfectly, because a machine that writes perfectly wrote it. The voice sounds perfect, because it was built from the real person's own recordings. The invoice looks perfect, because it is a copy of a real one, with one quiet change — the bank details. If your defense depends on your bookkeeper noticing something feels off, you no longer have a defense. You have to check the request, not the voice.
The Law
Sixteen days. On November 25, revised federal rules on opting out take effect. If a customer opts out of one kind of informational message, like appointment reminders, you may limit that opt-out to that kind of message. But if a customer opts out of marketing, that ends all of your marketing calls and texts, full stop. You may name specific ways for customers to opt out, but only if you state them clearly everywhere and they actually work when a customer tries them. Every opt-out must be honored within ten business days at the very latest, though the smart standard is the same day. The exposure is set in the law itself: five hundred to fifteen hundred dollars per violation, and those numbers add up across a call list very quickly. And if a customer tells a live member of your team, in plain words, to stop calling, that counts on its own — no form required. Your system has to hear it, record it, and honor it.
Who's Doing It
On November 5, federal prosecutors announced charges in a fraud case built entirely on this technology. Prosecutors say a small group targeted thousands of companies worldwide over about a year and a half, using email plus AI copies of a chief executive's voice — and in some cases the executive's face on video, plus a supposed lawyer — to persuade finance staff to wire money for a confidential deal. The reported total is about one hundred twenty-nine million dollars. The reported victims include hospital systems, a bank, and an investment company; they are not careless people. These are allegations and the defendants have not been convicted, but the method is the lesson, and it scales down perfectly: the small-business version is one urgent call, in the owner's voice, asking the bookkeeper to pay something quietly and quickly. The FBI's internet crime bureau received more than twenty-two thousand AI-related complaints in 2025, with reported losses above eight hundred ninety-three million dollars. The businesses getting it right have stopped trying to spot fakes. They run a system instead: the call-back rule, two-person approval, and a code word.
Your Move
Three moves, all free, all doable before Friday. First, write your call-back rule in one sentence and tell your team: any request to move money, change bank details, or pay a new supplier gets confirmed by calling the person back on a number the business already had — never the number in the message. A fake voice can call you; it cannot answer your real supplier's known phone. Second, pick your two-person limit: no single person approves a new payee or a payment above the amount you set, and urgency never skips the second person, because urgency is the weapon. Third, choose a code word for genuine urgent requests, share it in person, and never write it in an email. No code word, no action, no argument. Then tell your team, out loud, that following these rules will never get them in trouble — even when the person asking sounds like you, and even when the request turns out to be genuine and the delay costs an hour. The scam works because a good employee wants to be helpful. Your rules only protect you if your people believe you will thank them for checking. This is exactly the kind of protection we build into the systems we set up at The Agency — answering, follow-up, and reminders with the consent rules, opt-out handling, and verification habits built in from the start. Start with a free AI Audit and we will look at how your phones, messages, and payment requests actually flow, and show you where the gaps are. Check the request, not the voice.