Last month I told you to get out of the passenger seat and start driving AI instead of just riding along. If you missed it, start with NAR’s 2026 Technology Report Says Agents Use AI. Most Are Barely Using It. A lot of you did. Good.
But a new survey says the next problem is already here. Plenty of agents are now driving fast with no brakes, and nobody is checking the mirrors.
The risk with AI for real estate agents is no longer whether you use it. It is whether you read what it wrote before your client does.
Speed without brakes
In a September 2026 survey of 400 US independent agents across real estate, finance, and insurance, 150 of them in real estate, 88% said they have sent AI generated content to a client without editing it. 57% said they usually or always do.
Read that again. These are not beginners. The survey reports that 96% use AI tools a few times a week or more. So this is not a story about people who are afraid of AI. It is a story about people who got comfortable before they got careful.
And the confidence is not there either. Only 31% said they are very confident their AI use meets legal and compliance requirements. Think about that. Most agents know something might be off, and they hit send anyway.
A fair warning on the source. The survey comes from CYPHER Learning, a company that sells training, and the release does not make clear which group each percentage is based on. So I would not quote these as hard facts about every agent in America. I would treat them as a very loud smoke alarm. The behavior is real. You have probably seen it in your own inbox.
Why training is not fixing it
You would think this is a training problem, and it is. But not the way most people think.
97% of the agents surveyed said their organization provided AI training. Sounds great. Then 55% called that training broad or general, not practical. And 53% said AI skills and workflow training is the single most valuable training they could get.
So they are being trained. It just does not stick, because it never shows them what to do on a Tuesday afternoon with a listing going live at five.
No surprise, then, where agents go with AI questions. Asked to pick all that apply, 79% said ChatGPT, 59% online searches and forums, 55% official training, and 45% peers. The tool is teaching the tool.
This matters because the training world is about to get louder. Bright MLS, with more than 100,000 subscribers, announced an AI Academy in September built with CYPHER Learning. It is due by the end of 2026, with role-based courses, assessments, certificates, and a conversational AI search over their knowledge base. Costs were not stated. That is a real step and I am glad MLSs are investing in it.
But notice the pattern. Product training teaches you where the buttons are. Nobody has to teach you where the buttons are. What agents are asking for is judgment: when to trust the draft, when to rewrite it, and when to throw it out.
That is the gap. And it is exactly where the 88% number comes from.
The five landmines in AI real estate writing
After you read enough AI drafts, you start to see the same five problems show up again and again. Learn these and you will catch most of the damage before it leaves your outbox.
- Made up facts. The draft says the home has a renovated kitchen, a new roof, or a fenced yard because those things are common, not because they are true. AI fills gaps with what sounds right.
- Fair housing language. Phrases that hint at who would fit in a neighborhood, or who would not, can slip into a friendly description without anyone noticing. The rules apply to AI generated content exactly like anything you write yourself. Check with your brokerage and your attorney on what that means in your market.
- Wrong tone. The email is polite, warm, and sounds like nobody you have ever met. Your past clients will notice before you do.
- Stale market claims. The draft says prices are rising, or rates are falling, with total confidence and no date. It may be describing a market from a year ago.
- Invented neighborhood details. The coffee shop that closed, the school that changed boundaries, the park that is really a parking lot. These are the ones that embarrass you in front of a buyer who lives there.
Picture this. You send a past client a market update. It says inventory in her neighborhood is tight, and she knows three houses on her street have sat for a month. You did not lie on purpose. But now she wonders what else in your emails is autopilot.
That is the real cost. Not a complaint. A little less trust, quietly, one email at a time.
The five minute brake: read it like a client
You do not need a policy binder. You need a habit that takes five minutes. Here is the one I would teach in a room.
- Read it out loud. If you would not say it to their face, rewrite it. Your ear catches robot language faster than your eyes.
- Highlight every fact. Prices, features, schools, dates, distances. If you cannot verify it in thirty seconds, it comes out.
- Hunt for people words. Anything that describes who lives in a place, or who would like it, gets a second look.
- Date the market claim. Where did that number come from, and when? No source and no date means no claim.
- Ask the client question. If this person called me tomorrow and asked where I got that, could I answer?
That is it. Five minutes. Compare that to the cost of one angry buyer.
Then add one more thing that almost nobody has: a one page personal AI rulebook. Three parts. What I will use AI for. What I will never paste into a free tool, like client financials, Social Security numbers, or MLS restricted data. And what I always read before it goes out. Confirm anything about client data with your brokerage before you put it in any tool.
AI writes the first draft. You own the last mile.
Why this gets more important next year, not less
Here is the part that should wake you up. The unedited email problem is a human sending a draft. The next wave is AI agents that do the sending.
RealAnalytica launched Atlas Agents in August: AI agents for lead follow-up, client engagement, listing and market analysis, and transaction monitoring, with recurring tasks that run on their own once approved, plus more than 30 integrations with tools like CRMs, MLS, email, and e-signature. In July, FirstTeam Real Estate in Newport Beach announced it is rolling out Purlin company wide, where agents set up their own AI agents by voice, text, email, or chat. Those are company claims from press coverage, and I have not tested either one. Pricing and rollout details were not given.
The direction is clear though. The more AI does on its own, the more that one word, approved, matters. If a human forgot to read an email, that is one email. If an agent is approved to send follow-ups every week, that is a pattern.
So before you hire your first AI employee, make a Go, Ask, Never list. Go: what it can do alone. Ask: what it can do only after I say yes. Never: what it never touches. Write that down. Most people skip this step and learn it the hard way.
Your move this week
Pick one. Do it before Friday.
Take the last three AI drafts you sent and run the five minute read on them. Not to feel bad, just to see what you find. If you catch even one thing, you just paid for this article.
Then write your Go, Ask, Never list: what AI does alone, what needs your yes, and what it never touches. One page. Put it where you will see it.
Now I want to hear from you. No one is grading. Have you ever sent an AI draft to a client without reading it all the way through? Tell me what happened. I will bet the room is bigger than you think.
If you want help building this into your daily workflow with real examples instead of generic training, let’s talk. Book a free 30 minute discovery call with me and we will build your Go, Ask, Never list together.
Sources
All figures come from press releases and press coverage. I did not test any of the products named, and the survey was run by a company that sells training.
