THE AI EDGE · WEEKLY EDITION
The AI Edge, the weekly edition from Hub365 AI. Friday, October 23, 2026 · 9 min read · by Todd & Naty Ross.
This week's read
Google opened the week. On October 13, three people from Google Cloud wrote up a night they left Gemini 4 Argon alone with one of their big open-source projects. Google's account: it started at 9:11 PM and ran eleven hours with nobody watching. By morning it had proposed six improvements. Two shipped. Four got cut, and the write-up names the reason for each: one risked serving stale data, one depended on another library's private internals, one made the code harder to read without a measurable gain, and one only helped in test runs. Their own summary: "the agent explores, the maintainer subtracts."
The part of that story we could use wasn't the hours. It was the four "no"s, each with its reason written down. All week, pieces about very different things ended in the same decision: three possible outcomes for each thing, and someone who picks one.
Tuesday, it was a task. Since October 13, Gemini can run jobs on a schedule if you pay for Google AI Pro or Ultra, and for now only on a personal Google account. The task we gave you to copy pulls vendor invoices out of your inbox into a sheet every Friday, and its first run puts an event on your calendar for the fourth Friday from now. That is the day you decide whether the task stays, changes or gets paused. You do not have to remember that check - the template schedules it for you.
Wednesday, it was what AI hands you. One line in a phone note for every draft, with a letter: U if you used it as is, R if you redid it, T if you threw it out. On Friday you group the lines by task type, and each type lands on one of three moves: stop asking for it, rewrite the instruction, or turn it into a template. We tallied twelve of our own pieces. Not one shipped as AI wrote it: three needed small fixes that left the core alone, and nine we rebuilt in a place that mattered. In five of the twelve, the English turned out to have been written with the Spanish open alongside.
And today, it was a line on your website. This morning's Business Fix is about the audit we ran on our own homepage on October 20 and 21. Lines written at different times had stopped agreeing with each other: 8 seconds in one place, under 10 in another, "agency" here and "not an agency" there. Every number and every promise had to answer one question - if a client asked you to prove it today, what would you show them? - and got one of three fates: it stays, it gets rewritten, or it goes.
What makes that decision actually happen is giving it a date: the fourth Friday for the task, this Friday for the tally. Cutting things without a day on the calendar stays a good intention. With a day, it becomes a habit. And keeping a little is not failing. Google did not write up its two of six as a miss; in its telling, that was the process working.
Thursday showed where you cannot cut. On October 21 we walked a luxury client's route: Google, an AI assistant, videos and forums, in English and in Spanish. At every stop someone different was doing the talking: you on your profile, a directory, another vendor who posted a video from one of your weddings, a couple in a review. Your profile and your directory listing are yours to write. What the other vendor and the couple say, you cannot delete or rewrite. You ask: the vendor with a one-line message, the couple with a direct link to leave a review. Thursday's 2pm prompts write the vendor message for you, credit line ready to paste.
And Monday made the case for why the call has to be yours. Microsoft CEO Satya Nadella wrote on October 10 that "we simply can't outsource responsibility for what intelligence does on our behalf." Whatever goes out under your business name, whoever drafted it, goes out under your name. This week's play is the first set day for that decision, and it is today.
This week's play (2 steps, 15 minutes)
Monday left you one job: log each thing AI handed you as used as is, redone, or thrown out. Today is when you judge, using Wednesday's method.
You walk away with a U, R and T count for each task type and language, plus one change on paper for Monday. Next week, the only number to measure yourself against is this one.
If you have not logged anything since Monday: start today, five seconds per line in the moment, not from memory - by Friday memory only keeps the wins. Your first reading is Friday the 30th.
The full walkthrough, with examples, is in Wednesday's post.
This week on the blog
Monday · When AI acts on its own, who has to answer for it? Google's model working eleven hours unattended, Gemini taking on tasks, a voice agent that issues credits, the false tip an Anthropic model sent to Philadelphia police, Nadella on who answers for AI, outside evaluators inside Anthropic, and Character.AI in Chapter 11. It ends on one job, and today's play grows out of it.
Tuesday · Gemini now takes tasks. Give each one a review date The exact click from Chat to Task, a copy-ready task that schedules its own review, and four questions to bring to the quote when someone pitches you a voice agent. Heads up: for now it does not work with work accounts.
Wednesday · Your AI hands you a lot. Count how much of it you use Hours saved and pieces produced tell you what AI did. They do not tell you what you kept. A one-week tally with three letters, how to read it by task type and by language, and what turned up when we counted ourselves: the hidden work was in the English.
Thursday · Your luxury client checks four places before calling you. Each one shows a different list We walked the route of someone looking for a luxury wedding planner in Miami, in English and Spanish, and no business showed up in all four places. The "your four voices" framework, with a one-minute fix for each voice.
Thursday, 2pm · When another vendor posts your work: 2 prompts to ask for the credit, with the credit line ready to paste One for work that is already posted, one for a credits sheet before your next event. Tested in Claude on a made-up planner, with every failure spelled out.
Friday · Your website still promises things that stopped being true, and AI repeats them Our own homepage needed 20 fixes, among them two different response times for the same chat and a 24/7 promise that never said who was answering. Claude, asked to analyze the site, caught the same mismatches. The rule we used on every line: confirm it or cut it.
Tool of the week
Gemini tasks (Task), from Google. It was Tuesday's topic, and it makes this week's point better than anything else we covered: the useful part is less that it runs on its own than that its review can be on your calendar from day one. Everything below is from Google's announcement and its help pages; we have not set up a Gemini task of our own yet.
What it does. Since October 13, right in the chat window, you can give Gemini a job and tell it when to run: every Friday at a given hour, one time on a specific date, or when a particular email comes in. Google says a task can work with your email, calendar, Docs, Sheets and Drive, and can create files.
The detail that decides it. The account. It takes Google AI Pro or Ultra, and for now it works only on a personal Google account; work accounts are out. If your business email runs on Google Workspace, it is not for you yet. To set one up: at gemini.google.com, click Task under the text box; active tasks live under Scheduled, and from More you can Run now or Pause.
What it is for this week. Tuesday's invoice task. Every Friday it copies vendor invoices and quotes from your inbox into a sheet you own. It does not pay, reply to or delete anything, and its first run schedules its own review. Do one pass with Run now and watch it work, then match two or three rows to the emails they came from.
What it does not do. Google labels it experimental. Tasks eat into the same usage limit as everything else you do in Gemini. Inside Google Tasks it can add, edit or delete items in bulk without checking with you first. And an email can hide instructions that Gemini reads and you never see. The task we gave you writes to one sheet you own, answers nobody, and makes its first run while you are watching. We cover it in Tuesday's blog.
This week's worksheet
One word from Wednesday unlocks two things: the guide How to Subtract: What to Stop Asking AI For, which walks you through naming your task types, reading a tally that ends in a tie, and rewriting an instruction around whatever you keep fixing; and The Weekly AI Tally, a tool that does the adding for you and returns your numbers by task type and language, what each type is saying, and your first change.
Today's play gets you one decision for Monday. The guide and the tool help with next Friday's, and the one after that.
Send KEPT on WhatsApp and both land in your messages:
👉 https://wa.me/17866869497?text=KEPT
For you, who made it this far
You read all the way down on a Friday afternoon, and we noticed. What we are asking this week fits in a phone note: three letters, fifteen minutes of adding, one calendar event.
We have one rule at Hub365 AI: if it will not change anything in your business, we do not build it. This week a team that kept two of six and called it a good result reminded us why. Choosing what stays is work too, and it is the one job you cannot hand off.
Thank you for reading. The owners who go far are not the ones who know it all. They are the ones who keep learning when everyone else stops. That is why we do this, and what we do for our clients we treat as if it were our own business. We do not create just to create. We create to make an impact.
