Copy is the most obvious thing to hand to AI and the easiest to hand over badly. Ask a general model for a letter to motivated sellers and you get something fluent, enthusiastic, generic, and almost exactly what every other investor asking the same question received. It reads like marketing, which is the one thing seller copy must not do.
Used properly it is genuinely valuable, and the difference is entirely in what you supply and what you forbid.
Why the Default Output Fails
Three failure modes, all predictable, all fixable with constraints.
It writes for the category, not the situation. Asked for seller copy, a model produces the average of all seller copy it has seen: cash offers, any condition, fast closings, no fees. That is precisely the interchangeable language that gets sorted into recycling, and the whole argument in the offer is that a category is not a proposition.
It is relentlessly upbeat. Models default to enthusiasm, and enthusiasm is wrong for this audience. Someone dealing with a parent's estate or a foreclosure notice is not looking for excitement. Exclamation marks, "great news" and "amazing opportunity" all signal a sales letter to a reader whose first instinct is suspicion.
It overwrites. Asked to improve a paragraph, the output is nearly always longer. Good seller copy usually gets shorter as it improves, because the improvement is deletion.
What You Have to Supply
The model cannot know these and the output is worthless without them.
The specific situation. Not "motivated sellers" but "owners of properties that have been vacant more than a year, in a working-class suburb, where the owner usually lives out of state." The narrower the input, the less generic the output.
Your actual offer. What you genuinely provide, in concrete terms. As-is with contents left behind, their choice of closing date, no showings. If you supply "we buy houses for cash" you will get category copy back, correctly.
The register. Plain, unhurried, no urgency, no superlatives, the way you would speak standing in a driveway. Say this explicitly every time, because the default is the opposite.
A length cap. Under 150 words for a letter body, under 40 for a text. Without a cap you get expansion.
What to forbid. No exclamation marks. No manufactured deadlines. No claims about results. No mention of how you found them.
That last one matters more than it looks. Given a prompt saying these owners are in probate, a model will helpfully write "I understand you recently lost a loved one," which is the exact sentence that makes a reader feel surveilled, per the lead.
Where It Genuinely Earns Its Place
Volume of variations. Twenty headlines instead of five, because the useful ones arrive after the obvious ones are exhausted. This is the single best use, and it maps directly to the advice in headlines for seller marketing.
Niche adaptation. You have a letter that works for absentee owners. Adapting it for tired landlords, heirs and expired listings is mechanical work that used to eat an afternoon per version, and it is the thing that makes per-niche copy economically viable at all.
The adversarial read. Asking a model to read your letter as a suspicious recipient and say where it stops trusting you is more useful than asking it to write one. Models are agreeable by default, so you have to request the criticism explicitly.
Reformatting across channels. Turning a letter into a text, a voicemail script and a landing page, each respecting its own constraints. Tedious, mechanical, and the emphasis genuinely differs per channel.
Second-language and readability passes. Simplifying dense paragraphs without losing meaning.
Where It Should Not Touch
Anything asserting a fact you have not verified. A model will happily write that you have closed over two hundred properties, and it does not know whether you have. Every factual claim in generated copy needs checking, which is the standard in proof and credibility.
Anything making a commitment. Guarantees about price holding or closing timelines are promises you have to keep, and generated copy makes them casually.
Compliance-sensitive language, particularly in mail or texts to distressed homeowners, where several states impose specific disclosure requirements.
And the message to a seller you are already in conversation with. Once a relationship exists, generated follow-up is detectable and it costs more than the time it saves.
The Sameness Problem
A commercial risk specific to this moment, and worth thinking about before it bites.
Direct response works on differentiation. If every investor in your market generates copy the same way from the same models with similar prompts, the output converges, and the advantage that careful copy provides erodes for everyone.
Two defenses. Supply specifics no model has: your actual local knowledge, the street names, the real situations you have handled, your own numbers. And keep the voice yours, using generated text as a draft you rewrite rather than a finished piece you send.
The investors who will still have an edge in two years are the ones whose copy contains things that could only have come from them, which is also the argument for keeping a record of what actually worked in your market, per building a swipe file.
Three Prompts That Do Most of the Work
"Here is my offer in one sentence and the situation of the person receiving it. Give me eight headline options that name the situation rather than describing my service. No superlatives, no urgency, no company name."
"Rewrite this letter for a tired landlord instead of an out-of-state heir. Keep my offer identical. Change only what the situation demands, and keep it the same length or shorter."
"Read this as a suspicious 68-year-old who has had four similar letters this month. Tell me the exact sentence where you stop believing it."
The third is the most valuable and the least used. Models default to approval, so criticism has to be requested explicitly and framed as a role rather than a question, or you get a polite list of minor suggestions.
Editing Generated Copy: What to Look For
Specific tells, all quick to check.
Adjective density. Generated copy leans on words like seamless, hassle-free and stress-free, which are claims without content and which readers discount automatically.
Symmetrical sentences. Models produce balanced constructions at a rate humans do not, and several in a row read as machine-written even when each is fine.
The false summary. Generated pieces frequently end by restating what was just said. Delete it, and the piece almost always improves.
Hedged claims. Phrases like may be able to help and could potentially are inserted as caution and read as weakness. Either you do the thing or you do not.
Invented specifics. Any number, timeline or claim about your record that you did not supply is fabricated, and it will read as the most credible sentence in the letter.
The Workflow That Works
Write your own offer first, by hand, in one sentence. This is the input everything else depends on and it is the part you should not delegate.
Generate options rather than a finished piece. Headlines, openings, several angles.
Pick and rewrite. The output is raw material and the last pass should be yours, which is also what keeps the voice consistent.
Check every claim. Then read the draft out loud: if a line would embarrass you spoken face to face on the seller's porch, it has no business in a letter signed with your name.
Then test it against what you already run, because a version that reads better to you is a hypothesis until response says otherwise, per beating the control and split testing with low traffic.
The wider read on where this technology fits is in the guide to AI for real estate investors, and the prompts to start with are in prompts for real estate investors.