ChatGPT for Commercial Real Estate: Source-Checked Reports and Comps

A commercial broker can use ChatGPT to draft market reports, summarise comps and write owner outreach, provided every figure comes from material you paste in and a person checks each number against its source before anything is sent. ChatGPT writes and reshapes; it does not know your submarket, and it will happily invent a plausible cap rate.

That split decides everything below. The model is good at turning a pile of your own notes into a readable page, and poor at being a data source. Most broker horror stories come from asking it for facts it was never given. The method here keeps facts on your side of the line and gives ChatGPT only the writing.

Follow me on Instagram@sagnikteaches

What ChatGPT does well and badly in a broker's week

Sort your tasks into two bins before you write a single prompt. In the first bin, the text already exists in your files and the job is to reshape it: a comps spreadsheet that needs a three-paragraph summary, a call-note dump that needs to become a follow-up email, a 40-page offering memorandum you need condensed to a one-page brief. ChatGPT is reliable here because every claim can be traced to the pasted input.

Connect on LinkedInSagnik Bhattacharya

In the second bin, the facts have to come from somewhere else: current asking rents, vacancy by submarket, a recent trade, a zoning change, an owner's mailing address. ChatGPT can search the web on every plan, Free included, but the pages it cites are often brokerage marketing pages, quarterly reports that are a year old, or a figure lifted from the wrong geography. Anything in this bin is unverified until you have matched it to a primary source such as your own transaction database, a public property record, a listing service you subscribe to, or a published report from a research team you trust.

Subscribe on YouTube@sagnikteaches

The rule that follows is simple and worth printing for the team: no number leaves the building unless it has a source you can name in four words. "Research-service export, 3 Nov" counts. "ChatGPT said so" does not.

Setting up so deal data stays protected

Pick the plan before you pick the prompt. OpenAI's business-data page says inputs and outputs from ChatGPT Business, Enterprise and Edu, and from its API, are not used to train its models by default. The individual plans work the other way round: the pricing page marks content on Free, Go, Plus and Pro as used for training unless you opt out, so a solo broker should open Settings, then Data controls, and turn off "Improve the model for everyone" before pasting anything about a client. Even then, treat the chat box as something a colleague, a future support request or a court order could reach, and paste accordingly.

Three habits cost nothing and prevent most leaks. First, create a Project (a ChatGPT folder that holds instructions, files and chats together) for each ongoing assignment, so the instructions and source documents sit in one place instead of being pasted again each time. Second, name every file and prompt with the deal code rather than the owner or tenant. Third, keep a one-page "never paste" list pinned in the team chat (see the last section).

If you have not yet confirmed that ChatGPT's terms allow professional work for clients, this check on commercial use takes ten minutes and is worth doing once per firm.

A source-checked market report, step by step

Here is a main worked example. It is illustrative: the figures are invented to show the shape of the method. Say a broker covers light-industrial space and wants a one-page quarterly note for owners in a submarket of about 90 buildings.

Step 1: assemble the inputs yourself

The broker exports four things: the quarter's lease transactions from the firm's own records (14 deals), the submarket vacancy figure from the research service the firm already subscribes to (say 5.8 per cent, down from 6.4), the active listings count (11 buildings) and two paragraphs of personal observation from tenant calls. Everything goes into one document called "Q3 inputs", with the source and date written beside each figure.

Step 2: give ChatGPT the writing job and the boundary

You are helping a commercial broker draft a one-page quarterly note for light-industrial owners.

Use ONLY the figures in the attached "Q3 inputs" document. Do not add any market statistic, rent, cap rate, news item or comparison that is not in that document. If the note needs a figure I did not supply, write [NEED: description] instead of estimating.

Structure: a 2-sentence headline; three short paragraphs (leasing activity, vacancy and supply, what owners should do this quarter); a closing line. Plain English, no hype words, under 300 words.

After the note, list every number you used with the line of the input document it came from.

The final instruction matters most. Asking for a source line per number turns the output into something you can audit in two minutes.

Step 3: audit the draft against the source list

A realistic first draft might say "vacancy fell 60 basis points to 5.8 per cent" and "average asking rents rose 4 per cent". The first is correct arithmetic (6.4 minus 5.8). The second has no source line: the inputs only held deal-level rents, not an average asking figure. The model smoothed two numbers into a statistic nobody supplied. The broker deletes it or computes it from the 14 deals themselves.

The checking pass takes about ten minutes and has three questions per sentence: is the number in my inputs, is the arithmetic right, and does the sentence claim more than the number supports ("surged", "tightest on record")? Words like "record", "highest since" and "strongest" need a source with history behind them, and a single quarter's input cannot supply one.

Step 4: sign it

A named broker reads the final version top to bottom and signs off by initialling the source list, which stays in the deal folder. If a client later queries a figure, the answer is on file.

Comps summaries that do not drift

A comps summary is where numbers silently change. The model rounds, reorders and occasionally merges two rows. The safest approach is to keep the arithmetic in a spreadsheet (see the ChatGPT and Excel guide if your firm lives in workbooks) and let ChatGPT describe results, never compute them.

Example: the sale comps table

Say you paste a table of six sales with columns for building size, price, price per square foot, cap rate where disclosed, and sale date. Ask for two things only: a neutral two-paragraph description of the range, and a flagged list of anything that looks inconsistent.

Below are six sale comps. Do not recalculate or change any figure. Write two paragraphs describing the range (price per sq ft low to high, cap rate range where given, date range). Then list any row where a figure looks inconsistent with another in the same row (for example price divided by size does not equal the stated price per sq ft), quoting the row.

This second request is the real value. Asked to cross-check, ChatGPT will catch that row four says 41,200 sq ft at $5.1 million but $142 a foot, which would be $5.85 million at that rate. Sometimes the table is wrong, sometimes the model's arithmetic is, so you open the source record and settle it. Either way a transposed digit surfaced before a client saw it.

Example: the lease comps with missing fields

Lease comps are messier because terms differ: free rent, tenant improvement allowances, who pays operating costs. Tell the model to label each rent as "gross", "net" or "unknown" exactly as you labelled it and to refuse to compare unlike bases. A summary that says "rents ranged from $9.50 to $14.20" across mixed bases is worse than no summary. A better output reads "Net deals: $9.50 to $11.80 (4 deals). Gross deals: $12.90 to $14.20 (3 deals). One deal basis unknown, excluded." If the model blends them anyway, add the sentence "Never average or compare rents with different bases" to the Project instructions.

Example: explaining a comp that does not fit

Owners push back when a comp looks low. Paste your own notes on why (a sale-leaseback, a building with a short remaining lease, a distressed seller) and ask for a neutral paragraph explaining the adjustment. Because the reasons come from you, the model only phrases them, and the paragraph can go straight into the broker opinion of value after your read-through.

Owner outreach that is accurate and not mass-mailed

Outreach to property owners fails two ways: it reads like a template, or it contains a wrong detail ("congratulations on your 60,000 sq ft building" when it is 36,000). ChatGPT helps with the first problem and can worsen the second, so the facts in each email must come from your own research notes.

Example: the first-touch email

Feed the model a short research note per owner: building type, approximate size from the public property record, the one reason you are writing (a nearby lease you handled, a tenant whose lease expires next year according to public listing data), and a single question. Ask for an email under 110 words with no claims beyond the note.

Draft a first email to an owner. Use only these facts: [paste research note]. Do not mention rent levels, valuation or anything about the owner's finances. One specific reason for writing, one question, no attachments, under 110 words. Sign off as [your name]. Offer to stop if not relevant.

A realistic output opens with "Hello [first name]" and the nearby-lease reason, and ends with the question. The edits you will usually make: remove an invented compliment about the property, delete any adjective about the market, and cut a sentence that promises a result.

Example: the follow-up after silence

Three weeks on, a second message should add something, not nag. Give the model the first email and one new fact, such as a deal that closed in the submarket. Ask for a three-sentence follow-up. Check that the new fact is a public record and not something you heard in confidence from another client, which is the leak brokers most often make without noticing.

Example: a decline-friendly reply

When an owner says "not selling", paste the reply and ask for a courteous answer that keeps the door open and asks permission to send one quarterly note. This is pure tone work, which is where the model earns its keep. Add the owner to your prospecting routine so the quarterly note is actually sent.

Keep contact data out of the chat. Names and email addresses belong in your CRM; paste "[first name]" and the research note instead.

A verification table for the whole workflow

Print this and attach it to each project folder. It lists what ChatGPT produced, what can go wrong and how to check.

Condensing an offering memorandum without losing the catches

Long documents are where ChatGPT saves the most time and does the quietest damage. When you upload a 40-page offering memorandum and ask for a summary, the first thing to disappear is the awkward clause: a tenant termination option in year four, an environmental report flagged as pending, a ground lease that expires inside the hold period. A summary that reads smoothly can hide all three.

Ask for a structured extraction first and prose second. Use a prompt that names the traps: "List every lease expiry date, termination or early-exit option, rent escalation clause, ground lease term, pending litigation reference and any statement that a report is outstanding. Quote the page number for each. If a category is absent, write NONE FOUND." Then read the quoted pages. The page references let you verify in a minute, and "NONE FOUND" is a claim you should spot-check against the table of contents.

Example: a two-tenant building

Say a small office building has two tenants. The extraction returns: Tenant A, expiry March 2029, annual escalation of 3 per cent (page 14); Tenant B, expiry June 2027 with a one-time termination right in June 2026 (page 17). The termination date has already passed in this sketch, so the broker asks the seller whether it was exercised. A smooth prose summary would have said "two tenants with leases into 2027 and 2029" and the question would not have arisen.

A mistake that got through, and what stopped the next one

Here is an illustrative failure. A broker asks ChatGPT for a "short market update for an industrial owner" and attaches only last quarter's note as context. The draft says a nearby 52,000 sq ft building "recently traded at a strong price per foot". Nothing in the attachment mentions a trade. The model has pattern-matched what such notes usually say. Because the broker had not asked for a source line, the sentence read as plausible and went to the owner, who called a neighbour and discovered no such sale had closed.

The fix was procedural, not technical. Three rules went into the Project instructions: every sentence containing a figure ends with a bracketed source from the inputs; any sentence without a bracket is deleted in review; and the model must write [NEED: ...] rather than fill a gap. The second note the broker produced was shorter and slightly duller, which is the right trade.

A tenant-side example: summarising requirements for a search

Everything so far describes landlord and owner work, but tenant representation uses the same discipline. Say a tutoring centre wants 3,000 to 4,000 sq ft of retail space with parking and a lease of five years. You record a 20-minute call, transcribe it with a tool your firm has approved, and paste the transcript with names removed.

From this call transcript, produce a one-page requirements brief with these headings: space size range, preferred location types, parking and access, lease length and break options, budget as stated, deal-breakers, open questions. Use only what the client said. Where they did not say, write NOT STATED. Quote their exact words for budget and deal-breakers.

The brief is useful because the "NOT STATED" lines become your next call's agenda. The quotes protect you if the client later says they never mentioned a ceiling. Send the brief back to the client for confirmation before you start a search, since a correct one-page summary agreed in writing saves weeks of showing the wrong spaces.

Style rules that keep drafts credible

Add a short style block to your Project instructions so every draft starts closer to final. Ban the adjectives owners distrust: "prime", "unparalleled", "unique opportunity", "booming". Require numbers beside every comparison word, so "tighter" becomes "vacancy 5.8 per cent against 6.4 last quarter". Ask for one idea per paragraph and for the recommendation first, because owners read the top and skim the rest. Finally, tell the model to write in the first person singular if the note goes out under one broker's name, and in the plural for a team note, so the voice matches the signature.

What not to paste

Write this list into your Project instructions and the team handbook, and read it aloud once at the next meeting.

• Tenant names and suite numbers from rent rolls and lease abstracts. Replace with "Tenant A", "Tenant B".
• Owner financials: loan balances, debt service, tax returns, bank details.
• Anything under a confidentiality agreement unless the agreement and your firm's rules allow it. Off-market deals are the classic case.
• Another client's strategy, such as a client's maximum price. A prompt for one client should never contain information from another.
• Personal contact data: mobile numbers, home addresses, private email addresses of owners.
• Credentials and access links to data rooms.

If you are unsure whether a document is allowed, ask your compliance lead or a qualified adviser before using it. The cost of asking is a short delay; the cost of a leak is a client.

Measuring time, corrections and replies after a month

Pick three measurable signs. Time: log how long the quarterly note takes before and after, including the verification pass; as an illustration, a note that took two to three hours might come down to under an hour including the checking pass, but only your own log tells you. Errors: count the number of corrections the checking pass makes per document, and treat a rising count as a sign the prompt needs tighter boundaries. Replies: compare the response rate to outreach emails against your old templates, with the caveat that a few dozen sends is too small a sample to read much into.

Also run a monthly audit. Pull one report at random, trace each figure to its source, and record the result. When two figures cannot be traced, the process has a leak, and the fix is usually a missing line in the prompt rather than a different tool.

For the broader picture of which tools suit which broker task, this comparison by job shows where ChatGPT is the right choice and where a specialist tool does better, and a first-pass underwriting model is the natural next step once your comps are clean.

Questions brokers ask before pasting deal data

Can ChatGPT pull current comps or market rents by itself?

Not reliably. It can only work from what you paste or upload, plus whatever web search it runs, and web results are often marketing pages or stale. Treat any rent, cap rate or sale price it produces unaided as unverified until you match it to your own comps database or a listing record.

Is it safe to upload an owner's rent roll?

Only if your firm has approved the plan you are on and the owner's confidentiality terms allow it. A rent roll names tenants and lease terms. Strip tenant names and suite numbers first, or summarise the figures yourself and paste only the totals the task needs.

Should I tell clients a report was drafted with AI?

Check your brokerage's policy and any client agreement. A sensible default is to be honest if asked and to make sure a named person has checked and stands behind every figure, whatever tool drafted the prose.

Which ChatGPT plan should a small brokerage use?

If several brokers share prompts and you want business data protections, a business plan fits best, since OpenAI says it does not use ChatGPT Business data for training by default. A solo broker on an individual plan should switch off model improvement in data controls and keep deal data out.

Further reads

• How to Use AI as a Commercial Real Estate Broker: A Deal-Week Routine — Where these drafts sit in a working deal week.
• ChatGPT for Insurance Agents: Ten Safe Uses and Three to Avoid — Another client-data-heavy profession and its paste rules.
• AI Tools for Estate Agents: A Stack by Job, Valuation to Completion — A stack by job for property professionals.
• Real Estate AI Chatbot Use Cases: Which Three to Start With — Where a chatbot fits alongside your own drafting.
• Best AI for Real Estate Asset Management: Reports and Lease Abstracts — Compare ChatGPT, Claude and Copilot for lease abstracts, variance commentary and investor report drafts, with the checks that keep numbers honest.
• Best AI for Real Estate Marketing: Ads, Email and Listing Video — Pick an AI tool by job for real estate marketing: ad copy, email nurture sequences and listing video, with fair housing and disclosure checks.
• Can AI Abstract a Commercial Lease With Page Citations? — A lease abstract template, a prompt that forces page-cited quotes, a spot-check method and an exceptions list for the clauses AI most often misreads.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI business data privacy page, ChatGPT pricing page, OpenAI help-centre articles on web search, data controls and projects (checked 4 November 2026).