Type “write a follow-up to a buyer who toured a house yesterday” into any AI assistant and you will get something polite, warm and completely interchangeable. Every agent who asks that way gets a version of the same thing.
That is the visible problem, and it is the one people try to solve. It is not the expensive one.
The expensive problem is that the same request has gaps in it, and a language model will often try to fill gaps. Its default response to a missing fact is to produce the most plausible-sounding one available, in the same confident register as everything else it writes. There is no reliable signal in ordinary output separating what you supplied from what the model added.
In real estate, five particular blanks come up constantly. It is worth knowing what they are before a draft goes out with your name on it.
The reason a fresh chat writes toward the middle of everything is that it knows one thing about you: that you asked for a real estate email. Not your market, not your niche, not whether you work with first-time buyers or move-up sellers, not how you sign off.
Give it that once, at the start of a working session, rather than encoding it into every prompt:
You are helping me, [name], a real estate agent. Remember all of this for everything that follows in this chat.
Brokerage and market area: [yours]. Types of transactions I handle: [listings, first-time purchases, relocations, investment properties]. My tone in three words: [yours]. My standard sign-off: [yours]. Phrases I never use: [yours].
Two minutes, filled in once, saved somewhere you can paste from. Instructions can still be missed or outweighed later in a long conversation, so restate the important ones when the task changes.
That handles the voice. Now the part that matters more.
Ask what similar homes in a neighborhood sold for and a general model may produce a specific answer even when it is not connected to a current, approved market-data source.
Every number comes from you or another approved, verified source. There is no exception worth carving out here, because the numbers are exactly the part a client will act on.
This is the one agents do not see coming, because it does not feel like a fact at all.
Ask for warm copy about a neighborhood and a model will produce warm copy about the people in it. Family-friendly. Safe. Great for young professionals. Walkable to churches. A quiet, established community. These are phrases a model reaches for naturally, and they sit in exactly the territory that fair-housing training teaches agents to think hard about.
Whatever general safety training a model has had, it does not know your brokerage’s standards or your local ones, and nothing in the output flags that a line has been approached. Your brokerage has rules about listing and marketing language, and those are the rules that apply here rather than anything a blog can tell you. Find out what they are and read AI-drafted marketing copy specifically against them.
This is the category where the draft looks most finished and is most likely to need changing.
What the other side said. What your seller will actually take. Why they are moving. What the inspection turned up.
Two separate problems live here. The first is that information entered into an unapproved tool may be processed, retained or reused under vendor terms your brokerage has never reviewed. The second is subtler: a model that has been told your seller’s real floor will use it. It will write a genuinely helpful, well-phrased sentence built on a number the other agent should never see.
The roof is newer. The basement has never flooded. Zoned for [school]. The square footage. What the permits cover.
Hand a model a hint and it will state any of these plainly and confidently. Each one is a representation about a property, and the person whose name is on the message owns it, not the tool that drafted it.
Whether they should put it in a trust. What the capital gains will be. Whether they will qualify. What the contingency actually obligates them to do.
This is a boundary AI is particularly cheerful about crossing, because the questions sound like general knowledge and the answers are readily available in a plausible form.
Those five are why the setup block has a second half, and the second half matters more than the voice half.
My guardrails: never invent a price, a days-on-market figure, a comparable sale, a commission, or a timeline. Never state that a property will sell or predict an outcome. Never describe who a home or a neighborhood is suited to. Never give legal, tax or lending advice. Use only the facts I provide, and when something is missing, ask me one specific question instead of assuming. Mark anything I need to verify before sending.
It makes a question more likely than an invention. Often you get “what did the comparable sell for?” instead of a number. Not always, which is why it reduces how much you have to catch rather than removing the need to look.
A buyer follow-up, asked for cold:
Hi Sarah! It was so great showing you 14 Elm yesterday. I could tell you loved the kitchen! This home is a fantastic value in a wonderful family-friendly neighborhood and I don’t think it will last long on the market. Let me know if you’d like to see it again or write an offer. Homes like this are moving fast in this price range!
Four sentences containing a claim about value, a characterization of a neighborhood by who lives in it, a prediction that the property will sell quickly, and a claim about the market. None of it came from the agent. All of it is now in a message with the agent’s name at the bottom.
The same request, with the setup block in place and two facts supplied:
Hi Sarah, good to see you at 14 Elm yesterday.
Two things you said afterward stuck with me: the kitchen worked better than it looked in the photos, and you were not sure about the stairs.
On the stairs, I can ask the listing agent whether a rail was ever looked at, and I can get you into two others on your list with a similar layout this week so you have something to compare against.
No rush from me. Do you want to see it again, or see those two first?
[Sign-off]
The agent supplied exactly two things the model could not know: what Sarah actually said, and what the agent is offering to do next. Everything else was assembly.
That is the trade the whole approach rests on. You are not asking the tool to know your business. You are asking it to write down what you already knew, faster, and the blanks are where that arrangement breaks.
Before this becomes part of how you work rather than something you tried:
Which tools, if any, are approved for client communication. What client and transaction information is permitted to go into them. Whether AI-assisted marketing copy goes through a different review than anything else. And whether anything needs to be disclosed to a client about how their communication is produced.
Ask before rather than after. The answers are far easier to obtain before a problem occurs than to reconstruct afterward.
The Claude Workflow System for Real Estate Agents opens with this setup block, filled in and explained, then applies it across the whole client-communication cycle, with the five blanks marked in every workflow that touches one.
Get the Real Estate Guide →Free tools, no strings: the Starter Toolkit is 5 copy-paste AI prompts for validating, listing, and pricing digital products. Free on Gumroad, and it puts you on the list that hears about new systems first.
Get the Free Starter Toolkit →