THE AI EDGE · WEEKLY EDITION
The AI Edge, the weekly edition from Hub365 AI. Friday, October 2, 2026 · 8 min read · by Naty Ross and Todd Ross.
The read of the week
Monday's roundup opened on an uncomfortable idea: an email from the usual address no longer proves its owner wrote it, and a voice on the phone no longer proves there's a person talking.
Microsoft supplied the first half. On September 22 it took down EvilTokens, a subscription scam service that, according to the company, was tied to more than 12,000 compromised mailboxes across more than 10,000 organizations. It skipped the fake page entirely. It asked you to type a sign-in code on the real one. And once it was in, it wrote from the real account.
Meta supplied the second half. Its assistant, Muse, is already testing calls to US businesses for the people who use it, and an employee who called his insurer that way said they hung up as soon as they heard it was AI.
On Wednesday we read that line from the insurer's side. It didn't hang up on a machine. It hung up on a customer who had sent one. He knew, and he put it in writing. The insurer has no way to count those calls: a call you cut off doesn't get logged anywhere.
The rest of the week was the same crack in other doors.
On Tuesday we covered the friendly side of the same thing. On the 23rd, Google opened a feature that builds a synthetic copy of your voice from up to 30 seconds of audio; it requires the same person to record a consent line and marks everything it produces as AI-made. We covered it as a way to narrate what you've already written, and it still makes one thing clear about the phone: a voice that sounds like a person no longer proves that person is talking.
Thursday, at a real estate closing: according to Microsoft, as cited by Axios, EvilTokens used an AI assistant to search the inbox for payment conversations and work out who trusts whom. The email asking to change the account can land in the same thread, with the same signature. Reading it closely won't help: it was written to have nothing odd in it.
And today, the quietest message of all: the report that says "nothing new." A "nothing" on the screen looks exactly like data, and we caught it twice in September in our own operation.
Five blogs, one idea: the message no longer proves who sent it. Not the email, not the voice, not even the "nothing."
What does prove something is the path. An appointment an assistant asks for stays pending until the customer confirms from their own phone. An account change gets confirmed by calling a number you had before that email showed up. A "nothing" becomes checkable when it says where it looked, when, and what the newest thing there was: that doesn't make it right, but it lets you open that screen and see for yourself. And in all three, you go get the proof, down a path the message didn't bring.
For a fifteen-person business, this gets fixed with a conversation at the office, and the first agreement runs through the oldest door in the business: the front desk phone.
The question fits on one line and gets asked today: if a customer's AI assistant calls tomorrow to book an appointment, what does the person who answers your phone do?
This week's play (2 steps, 5 minutes)
It's the same play Monday's roundup closed on, and Wednesday added what to do with the answer.
What you leave with: where your front desk starts, before any script. "I'd hang up" says Wednesday's first decision is missing. "I'd help them" sends you to the other two: what a machine gets told, and when the appointment is real. And "I don't know" says none of them is made yet, which is an honest place to start.
And the warning, which we mean: don't turn it into a test. Correct them on the spot and next time you'll most likely get the answer you want to hear, and the one thing this exercise gives you is gone. Save the correcting for the script conversation.
The reasoning is in Monday's roundup, and what to do with that answer is in Wednesday's piece.
The one thing this week
Next to that answer, write one line that separates two cases that looked a lot alike this week: an appointment and a payment.
An appointment an assistant asks for gets confirmed by texting the customer's phone: the number in your records if they're already a customer, the one you're given if they're new. And that second number only proves someone at that phone wants the appointment. Nothing more.
An account change doesn't get a new number. It gets confirmed only at a number you had before the email arrived, and you place the call. An incoming call saying "yes, that's right" confirms nothing.
The difference is what's at stake. A wrongly booked appointment gets fixed the next day. A wire that's already gone often doesn't come back.
This week on the blog
Monday - The email and the phone call stopped proving who's there EvilTokens and the code typed on Microsoft's real sign-in page, Muse calling businesses, 200,000 candidates narrowed to 20 with a person reviewing at the end, Google's spam update, and an agent that treated "no" as one more obstacle. September 21 to 25, and the closing move is the play above.
Tuesday - Give ChatGPT the job out loud, and get it back in writing ChatGPT's voice now works inside Work, and if you hang up before it's done, it carries on in writing in the same conversation. Google's voice replication, with your recorded consent. And why the same PDF doesn't reach Claude and ChatGPT the same way. With the line to dictate and the script to record your voice in one take.
Wednesday - Your next customer won't call. They'll send their assistant Hanging up on the assistant is hanging up on the customer. Your front desk's three decisions: take the call on one condition, say only what's already public, and no appointment exists until the customer confirms from their own phone.
Thursday - The fake email doesn't have typos anymore. It has your closing instructions Wednesday's rule, now with money and stricter. Why checking the sender stops protecting you when the sender is real, and four questions before a single dollar moves on a real estate closing.
Thursday - Rehearse the call before it rings: 8 prompts to practice with your front desk Your own AI tool plays the assistant on the line, with made-up customers, and whoever answers your phone practices the three hard cases before they're needed. The hardest one is the polite caller who keeps pushing.
Friday - The report that says "nothing new" is the first one to check A "nothing" and a data point look identical on screen. We caught it twice in September, in our own operation, and the fix is one sentence with three blanks: where it checked, at what time, and what the newest thing there was.
Tool of the week
ChatGPT voice inside Work, from OpenAI. It came up Tuesday, and it's here because it's built for the gap between two appointments, without sitting down at a screen. It's paid: Work only comes with the plans that include it. And what follows comes from OpenAI's help pages; nobody here has dictated a job to it yet.
What it does. You hand ChatGPT a multi-step job out loud inside Work, its space for longer tasks: an email ready to send, a summary, a one-page proposal. If you hang up before it finishes, it carries on in writing in the same conversation.
The detail that decides it. Tell it up front not to answer until you say "done." Thinking out loud means pauses, and the voice can take any pause as its turn and start on half the request. OpenAI suggests exactly this.
What it's for this week. Getting ready for the script conversation, if your plan includes Work: between two appointments, dictate what has to come out of that conversation, and it leaves it in writing for you to read before you sit down. Not the line from step 2: that one you write by hand, exactly as you heard it.
What it doesn't do: it draws from two places at once: your voice limit, which on ChatGPT's Go and Plus plans is three hours in any rolling 24, and your Work usage. It only sees what's connected. And the part that ties it to the rest of this edition: voice doesn't approve. Nothing that needs your approval goes out on a spoken yes; you approve it on the screen.
This week's worksheet
Wednesday's is two pieces delivered together: the guide The Front Desk Script, with the full script, the table of what a machine gets told and what it doesn't, the confirmation text to copy, what to do when the customer doesn't reply, and the log sheet; and Is your front desk ready for an AI call?, eight questions in five minutes telling you which of the three decisions are already made and what your first fix is.
The guide gives you the lines. The eight questions tell you which of the three decisions your front desk is missing today, and they don't include the lines, which is why the two travel together.
Ask for them by sending the word FRONTDESK on WhatsApp, and they arrive immediately:
For you, who made it this far
Thanks for reading on Fridays. If you take one thing from this edition, make it the question for your front desk: this week the email and the voice stopped proving who's on the other end, and whatever the person answering your phone says is the first thing that gets tested.
At Hub365 AI we don't create for the sake of creating. We create to make an impact, and this week that meant not putting a number on something nobody has counted: we don't know how many of these calls a business gets today, and we won't make it up. The first real figure will be the one on your own log sheet.
