← All Issues
My Techie Broker

Issue 021 · October 30, 2026

A couple of weeks ago I told you my own brokerage didn't show up in my own AI search result. Here's what I did about it. Over the last six days, 159 posts went out with my face and my voice on them. I recorded none of them.


There is a version of me that speaks Spanish on camera. I wasn't there when it was recorded.

I cloned my voice. Trained on my own audio, my own accent, the way I talk when I'm not being careful. Then I built an avatar that looks like me and moves like me.

Now I approve a script and a video exists. No camera, no lighting, no take four. A man on a screen with my face and my voice, explaining a down payment program in Spanish, who is me in every way that matters legally and no way that matters physically.

The first one I watched was unsettling. Not because it was bad — because it was good, and because he said it better than I would have on a Tuesday at the end of a long day.

I'm telling you it's AI, in print, while the videos are running. That's not a disclaimer. That's the point. I'd rather you hear it from me than notice it yourself and wonder what else I've been quiet about.


Three weeks ago, Ryan Serhant did the same thing.

On October 8 he announced a channel that posts AI avatar content every day for the next twelve months, built on his likeness and his voice. On the same software I used.

Let me be straight with you, because you'll go look. His is better than mine. Noticeably. He has a production team, a formal partnership, and people whose entire job is making it look right. Mine is one broker at a kitchen table, one week, and twenty-one dollars.

I'm not putting those side by side to flatter myself. The gap between his production and mine is real and it isn't closing.

The gap that closed is the other one. Six months ago the choice for an agent like me was a professional operation or nothing at all. Now the floor is something that actually runs, that you can stand behind, for the price of a dinner. That's the part worth knowing, and most of the people reading this don't know it yet.


One sitting. A hundred and fifty-nine posts.

A week ago Friday I sat down once and scheduled 159 posts. Twenty-one videos across six days, four slots a day — seven in the morning, nine, noon, and a quarter to six. Every one of them has now gone out, and I did not touch the thing once while it ran.

One script I approve becomes one video. That video becomes a post on every platform I'm on, at a time I picked, without me opening an app. Twenty-one approvals turned into a hundred and fifty-nine appearances.

Compare that to August: clipping a podcast by hand, posting when I remembered, going dark for a week every time a deal got loud. That's two years of my life in one sentence, and it's why my name wasn't in that search result.


That week costs me under thirty minutes.

Here is the entire loop, start to finish, now that it runs.

ChatGPT writes the scripts. I give it the topics and the voice I want and it comes back with a batch.

I paste them into Claude. From there Claude does the work: generates the audio in my cloned voice, sends it to be rendered as video, builds out the posts, and loads the schedule.

I click approval links. Two videos at a time. I watch, I approve or I kick it back.

That's the week. Most of those thirty minutes are me reading scripts to make sure nothing in them is wrong, which is the only part of this job I'd never hand to anything.

I want to be careful here, because this is the sentence that gets newsletters like mine in trouble. Thirty minutes a week is true. It is also the least important thing I'm going to tell you.


The thirty minutes is on the other side of a wall.

Every article you'll read about AI video starts at minute thirty-one. They show you the easy part because the easy part is what sells the subscription.

Building this took about a week, and it broke the entire time.

The best example is the one that cost me an evening. My cloned voice worked perfectly where it was made — I typed a sentence, it came out in my voice, flawless. I connected it to the video side. The import succeeded. The system could see the voice, list it by name, everything looked right. Then I asked it to actually speak and it failed, with an error message that said essentially nothing.

I checked permissions. Rebuilt the key. Deleted the voice and re-imported it. Checked which engine was selected. Read documentation at an hour when nobody should be reading documentation.

It was a spending cap I had set on that key myself, months earlier, so a runaway job couldn't drain the account. It did exactly what I told it to do. The failure wasn't a bug. It was a guardrail, working, that I'd forgotten I installed.

That's the shape of the whole build. Not one impossible problem — a dozen small ones, each of which looks like a dead end at ten at night. Most of what looks broken in these stacks is something you configured and didn't write down.

So when somebody tells you this takes thirty minutes, they're not lying. They're just starting the clock after the hard part. The hard part is real, it's front-loaded, and it's where almost everyone quits.


What it actually costs, both kinds.

Four pieces, none exotic. A voice clone from ElevenLabs. An avatar from HeyGen that lip-syncs to it. A script I approve. A scheduler, OpusClip, that posts on a calendar. Replies land in Follow Up Boss like anything else.

If you only get one of those right, make it the voice. A face that moves a little strangely, people forgive — they've seen worse on a listing video. A voice that isn't yours, they don't.

Fixed cost: three monthly subscriptions, the voice, the avatar and the scheduler. Call it around a hundred dollars a month, and you can come in under that, because none of the three needs its top tier to do this. I pay it whether I publish one video or a hundred.

Variable cost: the rendering, billed by the second out of a prepaid balance. Every video, every test, every do-over since I started has come to $21.34.

That second number matters for what it didn't do. Going from a handful of videos to a hundred and fifty-nine posts barely moved it. The expensive part is standing it up. The volume, once it runs, is nearly free.

One warning, because it shows up on a statement: the rendering is billed separately from the monthly subscription. Your plan credits can't pay for it. And every cost estimate I was handed before I measured turned out wrong — the first figure I budgeted from was nearly double what I pay. Don't budget from an article, including this one. Load twenty dollars, render one video, read the balance before and after.


If you're going to build it, build it in this order.

Budget the build, not the subscription. Give it a week, and expect most of those nights to end in something broken. That's not failure, that's the job. Nobody gets to the thirty minutes without it.

Record clean audio first. Thirty quiet minutes of you talking normally. Everything downstream inherits the quality of that file, and no setting fixes a bad one.

Post five by hand before you automate one. If you won't record and post five yourself this week, you won't post thirty through a pipeline. The pipeline multiplies discipline, it doesn't create it.

Write down every limit you set. Spending caps, permissions, which key does what. Date it. I lost an evening to my own forgotten guardrail and it is the single most common way these builds stall.

One home per platform. Pick a single account per network. The same video on four accounts of one platform splits your own audience and can read as duplicate content.

Decide your disclosure line before video one. Write down what you'll say when a client asks whether that was really you, and make sure you'd be comfortable if they already knew.


Why Spanish, and what I won't let it say.

The videos are in Spanish because that's the language my family speaks and the language I already host a podcast in. It isn't a strategy I picked off a whiteboard. It's the voice I actually have.

Which makes the next part matter more, not less. The people who send me these questions are asking about credit, about down payments, about papers — the things you only ask someone you trust. Put a machine in that conversation and you had better have decided in advance what it is not allowed to say.

Mine doesn't promise outcomes. Not that you'll qualify, not that prices will do anything in particular. No legal or tax advice, and nowhere near immigration advice — that belongs with a licensed attorney and nowhere else. Every video is educational, carries my brokerage name, and ends by pointing at a real person.

That isn't a legal department talking. It's the floor. Nobody should get a confident answer about their money from something that can't be held responsible for it, and the responsible one is me.

The fair criticism came fast, and it came within days of Serhant's announcement. A trade publication called what's coming a flood of AI slop, and one line in it stuck with me: an avatar of a specific agent saying something generic is still generic.

They're right, and I'd rather say so than pretend I didn't read it. The machine solved my throughput problem. It did not give me one thing worth saying. That part is still mine, and it is still the hard part. Publishing daily is about to be something anybody can do. Being worth watching isn't.

And I've proven nothing yet. Thirty videos, thirty days, and I don't know if one of them lands. On day thirty I'll publish what happened in this newsletter — including if the answer is nothing. That's the version almost nobody writes, which is exactly why you should want to read it.


Before you go.

Would you put your own face on something you didn't record? Hit reply and tell me where your line is. I read every one, and I suspect this is the question our business argues about for the next two years.

And if you want to see what we've built, Danae runs twenty-minute Zoom demos of the tools our agents use daily. Watch one, answer six questions after, and six CE hours are yours at no cost — two full classes from the Arizona school where the courses are taught 100% by AI. ADRE-approved, three hours each, taught start to finish by Ada, who answers your questions mid-lesson straight from the Arizona Law Book. Pick a time →

— Josh Marquez
Designated Broker, HomePros
Director, Avance School of Real Estate (S22-0004)


Sources referenced in this issue
  • Everything here is my own build, October 2026 — the voice, the avatar, the schedule, the failures and the forgotten spending cap.
  • The 159 posts and 21 videos are read off my own publishing schedule. The $21.34 is read off my own prepaid rendering balance — balance before, balance after — as of October 21, 2026, and covers rendering only, not the monthly subscriptions.
  • No results are claimed. The thirty-day test had not finished when this was written.