A competitive market survey used to eat a day. You pulled the comparable properties, chased down asking rents and availabilities, dropped it all into a spreadsheet, then spent the afternoon writing it into something a client would read. The reason this now fits in an afternoon is narrower than the hype suggests, and worth being precise about: AI collapsed the part where you turn a pile of facts into a clean table and a finished narrative. It did not collapse the part where you get the facts, and it will happily invent them if you let it. Build a survey as a five-stage sequence — scope, source, structure, synthesize, verify — and you can hand three of those stages to an assistant and finish before the end of the day. Hand it the wrong stages and you will ship a confident, fabricated number under your own name. This is the workflow, stage by stage, for a lean firm with no analyst and no IT department.
What a market survey is, and why it took a day
A market survey is the structured look at everything competing with the asset in front of you: the comparable buildings in the submarket, their asking rents, available spaces, recent leasing or sale activity, concessions, and the trend those numbers describe. A listing broker builds one to price a space and win the pitch. A tenant-rep broker builds one to run a tour. An acquisitions principal builds one to screen a deal before spending real diligence hours on it. The document is the same shape; only the question it answers changes.
It took a day for a boring reason: two slow jobs stacked on top of each other. First you sourced the facts — pulling comparables and confirming rents and availabilities from your data platform, public listings, and calls. Then you synthesized them — sorting the comp set, formatting the table, and writing the paragraphs that tell the client what the numbers mean. The second job quietly ate more hours than anyone budgeted, and it is the job that just changed.
The five-stage assembly line
Break the build into five stages and it becomes obvious which ones AI touches:
| Stage | What it produces | Who does it |
|---|---|---|
| 1. Scope | The exact question the survey answers | You |
| 2. Source | Verified facts: comps, rents, availabilities | You + your data sources |
| 3. Structure | A clean, consistent comp table | AI, from your raw data |
| 4. Synthesize | The written narrative and takeaways | AI, from your table |
| 5. Verify | Every figure traced back to its source | You |
The middle two stages are where the day-to-afternoon compression happens. Stages 1, 2, and 5 stay human because they are judgment and accountability, not language work. This is the same division that runs through every disciplined use of AI in a lean firm’s deal work, and we lay out the full version of it in our playbook for screening and underwriting more deals with a small team. Get the boundary right and the tool is a force multiplier. Blur it — ask stage 3 or 4 to do stage 2’s job — and you have imported a liability.
Stage 1 — Scope the survey before you touch a tool
The fastest way to lose the afternoon is to start gathering data before you have decided what the survey is for. Scope is five decisions, and they take ten minutes:
- The asset and the question. Are you pricing a 12,000 SF office suite, positioning an industrial listing, or screening a retail acquisition? The question dictates everything downstream.
- The competitive set boundary. Which submarket, which property type, which size and class band. A survey that includes everything includes nothing useful.
- The metrics that matter. Asking rent and available SF are table stakes; add concessions, TI allowances, lease terms, or recent absorption only if they bear on the decision.
- The time window. Trailing twelve months of activity is a common default; a fast-moving submarket may want six.
- The audience. A survey for an internal screen can be terse. One going into a client pitch needs the narrative and the house format.
Write these five down in a sentence or two each. You will paste them into the assistant later as the brief that keeps stages 3 and 4 on target. Skipping this step is why so many AI-assisted documents come back generic — the tool was never told what decision it was serving.
Stage 2 — Source the facts (the part AI does not do)
This is the stage that did not get faster, and pretending otherwise is where firms get burned. The comparable rents, the current availabilities, the ownership, the recent trades — these are facts, and a general assistant does not have them. It has no live connection to your market and no proprietary database behind it. Ask ChatGPT, Claude, Gemini, or Microsoft Copilot for “average asking rent in my submarket last quarter” and it will produce a specific, plausible, wrong number in fluent prose, because guessing is what a language model does when it lacks a fact.
So you source the way you always have, from places that actually hold the data. Your paid platform is the backbone: CoStar, Crexi, or Buildout for availabilities and comps; Reonomy for ownership; Placer.ai for foot traffic on a retail survey; HelloData or Yardi Matrix for rent data, depending on your lane. Public listings, broker relationships, and a few confirming calls fill the gaps. Confirm each platform’s current capabilities against its own documentation before you lean on a feature, because proptech data coverage and AI-branded add-ons change from quarter to quarter and the label on a button is not a guarantee of what sits behind it. Knowing which sources an assistant can genuinely reason over versus which it will confidently guess at is its own skill, and we map it in our field guide to the CRE data sources AI can actually use.
Pull your comps into a rough export or a plain list. It does not need to be clean or formatted. Messy is fine — cleaning it up is exactly the job you are about to hand off.
Stage 3 — Structure the pile into a clean comp set
Here the afternoon starts paying off. Take your raw export — the ugly one, with inconsistent labels and mixed units — and hand it to an assistant with your stage 1 scope. Ask it to normalize the data into one consistent table: same rent basis across every row, square footage formatted the same way, property names de-duplicated, and anything outside your competitive-set boundary flagged for removal.
A capable assistant will turn a fifteen-row mess into a clean comp table in a minute or two. It will catch that one comp is quoted gross while the rest are triple-net and ask you which basis you want, standardize “12,000 sf” and “12k SF” and “12,000 square feet” into one format, and sort the set the way you asked — by rent, by size, or by distance from the subject. What it must not do is add a comp you did not give it or fill a blank cell with an estimate. Tell it explicitly: leave gaps as gaps, never infer a missing rent. A blank you can go confirm; a fabricated fill you might never catch.
The output of this stage is the spine of the survey — a table you trust because every number in it came from your stage 2 sourcing, not from the model.
Stage 4 — Synthesize the narrative in your firm’s voice
A table is not a survey. The client pays for the interpretation: where the subject sits in the set, what the trend is, what it means for the pricing or the tour or the offer. This is the language work that used to take the back half of the day, and it is what collapsed to minutes.
Feed the assistant your clean table and the stage 1 scope, and ask for the written survey to your house structure — a positioning summary, the comp discussion, and a clear recommendation or read. Give it a past survey as a format example and the voice comes with it. In a few minutes you get a first draft that would have taken an hour to write from scratch: the submarket story, the subject’s place in the rent band, the concession trend, and the one-line takeaway a principal can act on. The structured approach to turning market facts into a repeatable, publishable document is worth internalizing on its own, and we break it down in our framework for producing market reports on a small-firm budget.
The draft is a draft. It will be well-written and it will be confident about everything, including the parts it should not be. That confidence is precisely why the next stage exists.
Stage 5 — Verify before it leaves the building
Verification is not optional and it is not a formality. A market survey’s errors do not stay contained — a wrong asking rent or a mis-stated availability goes out under your firm’s name to a client who may price a decision on it. The failure has one shape: a number that reads exactly like it came from a database but was either fabricated by the model or transposed wrong in the handoff.
Two checks hold the line. First, every figure in the survey must trace to a source you control — your platform export, a confirmed listing, a call note. If you cannot point to where a number came from, it does not ship. Second, spot-check the figures the assistant carried from your table into the prose, because a transposition turns a real number into a wrong claim just as easily as a hallucination does. This takes fifteen minutes on a survey you sourced properly, because you are confirming facts you already gathered, not researching new ones. The accountability never moved: the broker or principal who signs the survey owns every number in it, whether a person or an assistant typed it. The same verification reflex governs everything downstream, from the market read to the deal model, a point we make in our account of how AI is changing CRE market research.
What an afternoon actually buys you
Be honest about what compressed and what did not. You did not skip data collection — stage 2 took as long as it always has. What you saved is the assembly and the writing, stages 3 and 4, which used to be the back half of the day and now take fifteen minutes each. The survey is not lower quality for being faster; if anything it is more consistent, because the table is normalized and the narrative follows a fixed structure every time.
The strategic point is bigger than one document. Synthesis was the layer where an institutional firm’s analyst bench gave it an edge — the big shop had people to write the survey and the boutique did not. When that layer collapses to minutes, the edge shrinks, and a 6-person firm that pays for one solid data source and prompts an assistant well can produce a survey that reads like a research department wrote it. The gap that remains is data and judgment, and a lean firm can compete on both. That is the whole thesis behind how small shops out-operate much larger competitors, which we lay out in the small-firm operating manifesto.
Getting there is mostly a fluency problem, not a software problem. Most firms already pay for a business-tier assistant, which runs roughly $20–60 per user per month, or can add one cheaply. The constraint is that the team has never been taught to prompt it well for the documents they actually produce. Focused training on exactly that — prompting for surveys, comp tables, and market write-ups — runs about $2K–15K in the current market and tends to pay back faster than a second data subscription. A custom automation that pulls from your specific sources and writes to your exact template is a real option later, at a market range of roughly $25K–150K to build, but it only earns its place once the same survey workflow repeats often enough to justify it. Learn the manual afternoon workflow first; automate it when volume demands.
FAQ
What is a CRE market survey?
A market survey is a structured comparison of the properties competing with the asset you are working on — the comparable buildings in a submarket, their asking rents, available spaces, lease or sale terms, and the trend those numbers show. A listing broker uses one to price a space, a tenant-rep broker to run a tour, and an acquisitions team to screen a deal. It answers one question: where does this asset sit in its market, and what does that mean for the decision in front of you?
Can AI really build a full market survey in an afternoon?
The assembly and the writing, yes; the data collection, no. AI compresses the two stages that used to eat the back half of the day — turning your raw comps into a clean, consistent table and writing the narrative around it — from hours to minutes. It does not collect the facts for you. If you already have your comps and availabilities sourced, a survey that took a day now takes an afternoon. If you count on the tool to supply the market data, you will get fabricated numbers.
Which AI tool should I use to build a market survey?
Any capable general assistant — ChatGPT, Claude, Gemini, or Microsoft Copilot on a business tier — handles the structuring and writing well, so use whichever your firm already pays for. The tool choice matters far less than the workflow around it: feeding it real data you sourced, giving it a clear scope, and verifying every figure before the survey ships. A better assistant does not fix a survey built on numbers the model invented.
How do I stop the AI from making up rents or vacancy figures?
Feed it data instead of asking it for data. When you supply the comp table and ask the assistant to structure and write from it, it works from real numbers. When you ask it for a rent or vacancy figure you did not provide, it guesses in fluent prose. Enforce two rules: every figure must trace to a source you control before it enters the survey, and you spot-check every number the assistant carried from your table into the narrative.
Do I still need CoStar, Crexi, or a data subscription if I use AI?
Yes, if you need the facts those platforms supply. An assistant writes about market data; it does not replace it. Platforms like CoStar, Crexi, Buildout, Reonomy, and Placer.ai exist because collecting and verifying availabilities, comps, ownership, and traffic is expensive, and no language model recreates that database by reasoning. The economical setup for most small firms is one data source matched to your dominant work plus a business-tier assistant to assemble and write everything up.
What should I check before I send an AI-built survey to a client?
That every figure traces to a source. Confirm each asking rent, availability, and comp against your platform export, the live listing, or your call notes, and spot-check the numbers the assistant carried from your table into the prose for transposition errors. On a survey you sourced properly this takes about fifteen minutes, because you are confirming facts you already gathered. The person who signs the survey owns every number in it, so this stage is never optional.
How is building a survey with AI different from a proptech tool’s built-in report generator?
A platform’s report generator pulls from that platform’s own data and formats it into that platform’s template. The AI workflow is source-agnostic: it structures and writes from whatever verified data you assemble, across every source you use, into your firm’s own format and voice. The two are not exclusive — you often source facts from a platform, then use an assistant to combine them with other sources and write the interpretation the generator cannot.
How much does it cost a small firm to start building surveys this way?
Less than most owners expect. A business-tier assistant runs about $20–60 per user per month, which most firms already pay. Getting the team fluent enough to prompt it well for surveys and comp work runs roughly $2K–15K in the current market. Data-source subscriptions vary by provider and market. A custom automation that pulls from your sources and writes to your template ranges roughly $25K–150K and only pays back at high, repeatable volume — most firms do not need it early.
Is AI going to replace the analyst who builds our surveys?
No — it changes what the role does. AI absorbs the assembly and drafting, so whoever builds your surveys spends less time formatting tables and writing paragraphs and more time sourcing, verifying, and applying market judgment. At a firm with no dedicated analyst, that means a broker or principal can now produce a research-grade survey themselves in an afternoon. What survives is the work the tool cannot do: knowing the submarket, choosing the right comps, and standing behind the numbers.
Key takeaways
- A market survey is a five-stage build — scope, source, structure, synthesize, verify. AI collapses the two middle stages from hours to minutes and leaves the other three human.
- The afternoon is real because the assembly and writing got fast, not because data collection disappeared. You still source the facts the way you always have.
- A general assistant has no market database. Feed it data you sourced; never ask it to supply rents, availabilities, or comps, or it will invent them in fluent prose.
- Verification is non-optional. Every figure traces to a source you control, every carried-over number gets spot-checked, and the person who signs the survey owns every number in it.
- The change favors small firms. When assembly and writing collapse, a lean shop with one data source and a well-prompted assistant produces a survey that reads like a research department built it — get the team fluent first, automate later.
Not sure whether your firm should start with fluency training, a new data source, or a custom build? A short assessment answers that faster than any tool comparison, because your deal mix, markets, and current workflow drive the choice. Book your free AI-readiness assessment →
Dirk Jan van Veen, PhD
