Issue 014 · September 7, 2026
The number everyone skipped.
The stat you've seen a hundred times comes from the National Association of REALTORS 2025 Technology Survey: ChatGPT is used by 58% of agents, Gemini by 20%, Copilot by 15%. Every vendor deck in the country has that slide.
Here's what's sitting right underneath it. Asked what AI has actually done for their business, 17% said significantly positive. 33% said moderately positive. And 46% said no noticeable impact at all.
Read that again. Nearly half the agents using AI report that it changed nothing.
That is a strange result for a technology this good. And the explanation everybody reaches for — they picked the wrong chatbot — is wrong.
You can't get an edge from a task where everyone gets the same answer.
The same survey tells you what agents are actually doing with it. 46% report using AI-generated content, mostly listing descriptions.
So picture the math. Half the profession, same tool, same kind of prompt, same task. The outputs converge. They have to. That isn't a knock on any model — it's what happens when thousands of people ask the same question the same way.
A listing description you generated in nine seconds is a listing description your competitor generated in nine seconds. It saved you twenty minutes. It didn't move you an inch.
The 46% who felt nothing aren't using the wrong AI. They're pointing it at the wrong thing. They were told AI writes listing copy, so they used it to write listing copy, and it wrote listing copy. The tool worked exactly as advertised. The advertising was just aimed low.
There's a line, and almost nobody has crossed it.
On one side you're using AI: you open a chat window, ask for a thing, get the thing, close the window. Tomorrow you start over from zero.
On the other side you're building with it: you make something that runs the same way every time, that knows your market and your rules and your voice, that a person who isn't you could operate.
Since February I've put at least thirty hours a week into that second side. On top of running a brokerage. A school with an AI instructor that answers from the Arizona Law Book, a set of tools my agents use every day, course material, quiz banks — systems that run whether I'm awake or not.
I'm not a software engineer and I still can't write code from scratch. That stopped being the requirement. What I can do is describe a problem precisely and keep pushing until the thing works — and it turns out that's the actual skill now.
And that's where the survey stops being useful. It measures which chat window agents type into. There's no line on it for the agents who quietly built something.
Now the part where you should discount my opinion.
Straight out, before I say anything else: Ada runs on Anthropic's Claude. I built our school on it. So when I tell you what I prefer, weigh that accordingly — I'm not a neutral party and I'm not going to pretend to be.
With that on the table: for the building work, the tools are not interchangeable. Same request, and one comes back with something structured, designed, and close to finished, while another comes back with something that needs an hour of cleanup. At thirty hours a week, that difference stops being a preference and starts being months. For me Claude has been the one that reasons through a problem instead of just answering it, and the gap shows up most in the things people will actually look at.
But here's the part that isn't opinion, and it's the reason I'm not telling you to go switch: that gap barely exists on the task most agents use AI for. Ask any of them for a listing description and you'll get four good listing descriptions. If that's your whole use case, the 58% is fine and you should stop reading vendor comparisons.
Which tool you pick is the small decision anyway. Here's the big one.
It isn't a better prompt. It's rules that hold.
I've built five of these for my own agents, inside a platform we call TrueTopProducer. A comps tool. An inspection desk that turns findings into a repair request. A deal desk that answers contract questions against the actual form. Something that works your database instead of storing it. And Ada, for Arizona license law.
Every one of them is a chat window with a fence around it.
The comps tool isn't allowed to hand over a number without telling you which comps it threw out and why. The deal desk answers from the form in front of it, not from memory. Ada has to trace every duty to an Arizona statute, an ADRE rule, or federal law — never to an association's material — and when the answer isn't in the Law Book she says so instead of guessing.
Those rules aren't in my head. They're written down in a file, and the tool reads them before it reads your question. Which is why my newest agent gets the same answer my best one does, at two in the morning, on the four-hundredth run.
That's the whole difference. Not a cleverer question — a set of constraints. What it may cite. What it must never say. What it returns. What it does when it doesn't know.
Build that once and two things happen a chat window can't do. It stops drifting, and you can hand it to somebody else. A prompt lives in your head and dies with your attention. A system with rules runs while you're at a listing appointment.
Adoption stopped being the question.
For three years the industry asked whether agents were using AI. That question is answered and it turned out not to matter much, because the answer was yes and 46% of them felt nothing.
The gap that's opening now isn't between agents who use AI and agents who don't. It's between the ones typing questions into it and the ones building things with it. And that second group is small enough right now that being in it is still worth something.
That won't hold. In a few years the survey question won't be which chatbot you use, it'll be what you've built — and by then everyone will have built something. The advantage is in the gap, and the gap is open right now.
So the question isn't which AI is best. It's what you'd make if you understood that you could.
Before you go.
What's the thing you do every week that's never the same quality twice? Not the task you hate — the one that comes out great when you're rested and rough when you're not. That's the one worth building.
And if you want to get good at this, our CE classes are the cheapest place to start — three hours each, taught by an AI instructor we built ourselves. You'll be learning Arizona law and watching the thing work at the same time.
— Josh Marquez
Designated Broker, HomePros
Director, Avance School of Real Estate (S22-0004)
- 2025 Technology Survey — National Association of REALTORS, published September 18, 2025. Survey conducted July 2025; 49,233 active REALTORS sampled, 1,241 usable responses.