The secret nobody selling you an AI writing tool mentions: when every firm prompts the same few models the same way, the output converges. Your listing descriptions, market write-ups, investor updates, and client emails stop sounding like your firm and start sounding like the industry average — competent, correct, and interchangeable. The copy did not get worse; it got generic, which in a relationship business is worse than worse. A tenant rep who has read forty AI-flattened notes this month reads yours as the forty-first and discounts it before finishing the first line. This is the sameness problem, and it is not a grammar issue you fix with a better sentence. It is a differentiation issue you fix by changing where AI sits in your workflow — and by protecting the thing that made counterparties trust you: your voice and your read of the market.
Why AI content converges on one voice
Sameness is not a defect in one tool. It is the predictable result of how these tools are built and how most people use them. Three forces push every firm’s output toward the same middle.
The first is model convergence. A small number of general models — ChatGPT, Claude, Gemini, Microsoft Copilot — write most of the AI-assisted content in commercial real estate today. Each one is tuned to be a broadly helpful, inoffensive assistant, which means each one defaults to a neutral, agreeable register that reads as competent to almost everyone and distinctive to no one. That default is a feature for a general audience and a liability for a firm whose edge is being not average.
The second is prompt sameness. When a broker in Denver and a broker in Atlanta both type “write a professional listing description for a Class B office,” they steer the same model with the same instruction toward the same answer. The words differ; the skeleton and the tone do not.
The third is the safe-median effect. A model trained to satisfy the widest range of readers gravitates to prose that offends no one — which is, by construction, the average of all the writing it learned from. Average prose has no edge, because edge is deviation and deviation is exactly what a “helpful, professional” default sands off. You did not ask for the industry mean; you asked for a professional description, and the industry mean is what “professional” resolves to.
Where sameness shows up in a brokerage
The flattening is easiest to ignore because it hits every surface a little, rather than any one surface a lot. Named plainly, here is where a 4-to-20-person firm loses its voice.
Listing marketing. CRE listing copy was cliché-ridden before AI existed — “prime location,” “well-appointed,” “won’t last.” AI does not invent new clichés; it amplifies the ones already in the training data, because those phrases are what “listing description” statistically looks like. Tools built into Buildout, CoStar, and LoopNet generate a clean description from your property data in seconds, and it reads like every other machine-assisted description built from similar data. The convenience is real. So is the cost: on a portal where a prospect scans thirty listings, the one that sounds like the other twenty-nine earns no second look.
Market write-ups and commentary. A quarterly submarket note is where a firm demonstrates judgment — the read on absorption, the risk nobody else is pricing, the contrarian call. Ask a model for “a commercial real estate market update” and you get sound, sourceable, forgettable macro framing any firm could have published. The facts may be right; the judgment — the reason a client reads your note instead of a data provider’s dashboard — is absent.
Client email and CRM-drafted communication. An AI CRM for real estate, or a copilot drafting from records already in HubSpot or Microsoft Copilot, produces house-less prose by default: correct, polite, and indistinguishable from the copilot output of the firm across the street. The reply-rate consequences on cold outreach are their own subject, covered in our analysis of why AI-written outreach decays and what restores reply rates; here the concern is broader — every touch a client receives sounds a little less like the person they hired.
Investor and LP updates. For a firm that raises capital, the update letter is a trust document; its voice signals that a specific, capable operator is watching the portfolio. Flatten it into generic reporting prose and you quietly tell your investors that no particular person is home.
Why sameness is a business problem, not a style preference
Treating sameness as an aesthetic quibble is the mistake that lets it compound. In a relationship business, voice is not decoration. It is a competence signal, and sophisticated counterparties read it as one.
Differentiation is the product. A brokerage does not win on data — the data is commoditized and largely public. It wins on judgment, relationships, and a distinctive read of a market. All three travel through the words you send. When those words converge on the industry average, you have handed away the only part of the offering a competitor cannot replicate. The listing that reads like the portal-generated one has already lost the argument that you add something the portal does not.
Sophisticated recipients discount generic content fast. Landlords, acquisitions officers, tenant reps, and institutional investors receive AI-flattened material constantly and pattern-match it in seconds — the same reflex that files a mail-merge letter unread. A message that sounds like everyone else’s is a small withdrawal from the trust account, because it signals either that no senior person touched it or that the firm has nothing particular to say.
There is also a slower cost: commoditization of the firm itself. Voice is how a client distinguishes the broker who genuinely knows the submarket from the one who ran a query. Erase that distinction across enough touches and the client’s honest conclusion is that the choice of firm does not matter much. The ten rules of AI-assisted client communication exist because each message is a deposit or a withdrawal, and sameness makes every one a small withdrawal.
How brokers keep their voice with AI in the loop
The goal is not to abandon AI and go back to typing everything by hand. It is to keep the productivity while refusing the sameness. Four moves do most of the work.
Move AI upstream, from writing to preparing. Whole-message generation is what produces the flattened register. Point the model at research and structure instead — summarize the ownership and lease history, assemble the comps, outline the note — and let a person write or rewrite the sentences that carry the firm’s judgment. This is the same reframe that governs prospecting once AI drafts the first pass, which we cover in rethinking prospecting when AI writes the first draft: the tool makes specificity affordable; the human supplies the voice and owns what goes out.
Feed the model your voice, not “professional.” A model steered by a generic instruction returns generic prose. A model given three or four examples of your firm’s best past writing as context, and a prompt that names your register — plain, direct, opinionated, no clichés — returns something far closer to house style. Voice capture is the single most valuable habit here, and it costs nothing but assembling a folder of your own good writing.
Apply the specific-fact rule. Sameness lives in general statements; distinction lives in specifics no template can carry. A real comp, a named submarket shift, a lease expiry, a risk you are pricing that others are not — one verified fact per piece is worth more than a paragraph of polished generality, and it is the content the industry-average model structurally cannot produce. It requires a person who looked at this asset or market and found something true.
Run a read-aloud editing pass. Before anything goes out, read it as the recipient and ask one question: does this sound like us, or like everyone? The passages that sound like everyone are the AI default leaking through, and rewriting only those — usually a handful of sentences — is a five-minute pass that restores most of the lost voice. It is also where a confidently wrong claim about a property or a market gets caught before a sharp-eyed counterparty catches it for you.
A lightweight voice system a small team can run
None of this requires a marketing department, a custom-built assistant, or new software. A firm of four to twenty people can stand up a voice system in an afternoon, and the payoff is that AI speeds up the work without homogenizing it.
- A one-page voice reference. Half a page on how your firm writes — register, banned clichés, sentence length, the difference between how you write a listing and how you write an LP letter — pasted into the model as context on every substantial piece.
- A folder of exemplars. Five to ten pieces of your best past writing, by content type, used as few-shot examples. This teaches the model your voice more effectively than any description of it.
- A prompt template per content type. A saved prompt for listings, for market notes, for client email — each one specifying “prepare and structure, I will supply the judgment,” not “write the finished thing.”
- A human sign-off rule. Nothing client-facing leaves without a person running the read-aloud pass and inserting the specific fact. The rule is the control that keeps the default register from shipping.
This works because it depends on fluency, not tooling. A broker who understands why content converges and where voice comes from will get more from a plain chat window than a firm running an expensive platform that still generates whole listings at volume. That sequencing — capability before software — is the spine of the small-firm operating manifesto, and it is what short LLM-fluency training covers: prompting for research, market notes, and drafts a person then grounds in your voice. When the work connects across inbox, CRM, and listing marketing as one system, the communications playbook for AI across those surfaces frames how the pieces fit. Custom automation to wire a voice reference into every draft is a later, optional step — market rates for that kind of build run from roughly $25K to $150K — well after fluency has done the cheap work first.
Frequently asked questions
Why does AI-written content sound the same across different firms?
Because a small number of general models write most of it, and each is tuned toward a neutral, agreeable register that reads as competent to everyone and distinctive to no one. When two firms give the same model the same generic instruction — “write a professional listing description” — they steer it toward the same answer; the words vary but the skeleton and tone converge. It is the default behavior of instruction-tuned models responding to default prompts, which is why the fix is the prompt and the workflow, not the tool.
Is generic AI content actually a problem, or just an aesthetic complaint?
It is a business problem. In a relationship business, voice is a competence signal counterparties read — “sounds like a specific capable operator” is taken as evidence one is involved. Content that reads as the industry average signals the opposite: that no senior person touched it. It also commoditizes the firm, because voice is how a client tells the broker who knows the submarket from the one who ran a query.
How is this different from the AI outreach reply-rate problem?
Related but distinct. Outreach decay is about reply rates on cold prospecting — templated shapes, deliverability, list quality — and is covered separately. The sameness problem is broader: voice and differentiation across every content type a firm produces, including listings, market write-ups, and investor letters, most of which are not measured by reply rate at all. Outreach decay is one symptom; losing your distinctive voice across all client-facing writing is the underlying condition.
How do I keep my firm’s voice when using AI to write?
Change where AI sits in the work. Use it upstream to research and structure, and have a person write the sentences that carry judgment. Feed the model three or four examples of your best past writing as context instead of asking for “professional” prose. Insert one verified fact per piece that no template could carry. Then run a read-aloud pass and rewrite only the sentences that sound like everyone rather than like you.
What is the fastest way to make AI content sound more like us?
Give the model examples of your own writing. A prompt that includes three to five pieces of your firm’s best past listings, notes, or letters as few-shot context steers output toward your voice far more effectively than any adjective like “professional” or “engaging.” Pair that with a one-page voice reference — register, banned clichés, sentence length — pasted in as context, and most of the generic default disappears before you edit a word. It needs no new software.
Do AI writing tools built into CoStar, LoopNet, or Buildout make the sameness problem worse?
They make it more likely, not inevitable. Those built-in generators produce a clean description from your property data quickly, and because many firms feed similar data through similar generators, the outputs converge. The risk is publishing the result unedited so your listing reads like the other machine-assisted listings on the same portal. Used as a first draft you then make specific and rewrite in your voice, they save time; used as the finished product, they hand your differentiation to the tool. Verify any tool’s current features against vendor documentation, as proptech capabilities change quarterly.
Does an AI CRM for real estate help or hurt a firm’s voice?
Both, depending on use. A CRM copilot that drafts from records in HubSpot or Microsoft Copilot saves time and keeps client history in one place. Left on its default, it produces house-less prose that sounds like every other copilot’s output, so every client touch drifts further from the person the client hired. The remedy is the same voice system: supply exemplars and a voice reference as context, keep a human sign-off on client-facing messages, and treat the draft as a starting point.
Do we need training or new tools to fix this?
Training first, tools later. The sameness problem is caused by how people prompt and where they place AI in the workflow — both fluency issues, not tooling gaps. A team that understands why content converges and how to capture voice gets strong results from a plain chat model, while one that bought a platform and still generates whole pieces at volume keeps producing average prose on expensive software. Short, task-focused LLM-fluency training pays back faster than any purchase because it fixes the behavior that caused the problem.
Where to start
AI did not make your writing worse. It made it average, which in a business built on judgment and relationships is the more dangerous failure — average is invisible, and invisible content earns no trust. The recovery is not cleverer prose. It is moving AI upstream to research and structure, feeding it your own voice as context, insisting on one verified fact per piece, and keeping a human sign-off on anything a client reads.
A free AI-readiness assessment is where that starts. A short working session reviews how your team currently uses AI across listings, market notes, CRM, and client email, shows where the default register is quietly costing you differentiation, and returns a plain plan for capturing your firm’s voice before you buy anything. Book a free AI-readiness assessment before your firm’s writing sounds like everyone else’s.
Dirk Jan van Veen, PhD
