Skip to content

Lesson 01 · 13 min read

What AI Can and Cannot Do in Commercial Real Estate

How AI chat assistants actually work, the commercial real estate tasks they handle well, the ways they fail, and the rule that keeps you safe: AI drafts, you verify and sign.

AI assistants can read a 60-page offering memorandum, draft a follow-up email, and write a spreadsheet formula in less time than it takes to find the file. They can also invent a tenant, misplace a decimal, and cite a page that does not exist, all in the same calm, confident tone. The skill is not getting the tool to talk. It is knowing which answers to trust, which to check, and which never to ask for.

This lesson explains in plain terms what these tools are, where they help in commercial work, how they fail, and the one rule that runs through the rest of the course. It ends with a worked example on a small Kissimmee retail strip and a preview of the confidentiality rules covered in Lesson 7.

What a large language model actually is

The general assistants you have heard of, such as ChatGPT, Claude, Gemini and Microsoft Copilot, are built on large language models. A large language model is a program trained on a very large amount of text. From that training it learns patterns: which words, numbers and ideas tend to follow which. When you type a question, it produces a response one piece at a time, each piece chosen because it is a likely continuation of everything before it.

That one idea explains most of the behavior you will see.

  • It is not a database. When you ask for the cap rate on a specific sale, the model does not look it up. It writes something that looks like a cap rate. If the real answer was never in its training text, or was there only faintly, it may still produce a number.
  • It writes fluent text by default. Fluency is the product. A wrong answer and a right answer come out in the same polished paragraphs.
  • It has a knowledge cutoff. Training stops on a date. Anything after that date, such as a new zoning change, a recent sale, or this quarter's lending conditions, is missing unless the tool can search the web or you give it the document.
  • It works best on what you hand it. When you paste a lease or upload a rent roll, the model works from that text in front of it. Results improve sharply, though it can still misread or skip parts.

Some tools now add web browsing, file uploads, or connections to your email, calendar or cloud drive. Those features change what the tool can see. They do not change what it is. A model that browses can still pick a stale page, misread a table, or blend two sources into one wrong fact.

What AI does well in commercial work

Language models are strongest when the job is to transform text you supply into a different shape. They are weakest when the job is to supply facts you do not have.

StrengthCommercial real estate exampleWhat you still check
SummarizeTurn a 60-page OM into a one-page deal sheetEvery number against the OM page it came from
DraftFirst version of an owner letter, tour recap, or LOI cover emailFacts, tone, licensed firm name, anything promised
Extract into tablesPull tenant names, suite sizes, and lease dates from a rent roll into columnsRow counts, totals, and any cell that looks off
ExplainWalk through what a co-tenancy clause does, in plain EnglishThe actual clause; explanations are general, leases are specific
Write formulasBuild an Excel or Google Sheets formula for DSCR or a rent scheduleTest the formula on a case you can do by hand
CompareList differences between two lease drafts or two LOIsRead both documents; a missed difference is silent
RewriteShorten a property description, change reading level, fix grammarThat meaning did not change in the rewrite
BrainstormQuestions to ask a seller at a first meetingRelevance to this property and this owner

Notice the pattern in the right column. The tool speeds up the first draft. It does not remove the review.

How AI fails

These failure modes show up across every general assistant. Newer versions reduce some of them. None has eliminated them.

Invented facts and citations

The polite name is hallucination. The model produces a statement that sounds right and is not supported by any source. In commercial work this looks like an invented tenant lease expiration, a made-up zoning designation, a cap rate "from a recent report" with no report behind it, or a statute section number that does not say what the model claims. Citations are especially risky. A model asked for sources may produce realistic-looking titles, page numbers or case names that do not exist.

Arithmetic slips

Models predict text, and digits are text. They often get simple math right and then drop a line item, transpose two digits, or round in the middle of a calculation. Some tools run calculations in a separate code tool, which helps, but you still need to see the inputs it used. Any number that feeds a price, an offer, a loan request or an ad gets recalculated by you or by a spreadsheet you control.

Stale knowledge

The model's general knowledge has a cutoff date, and it may not tell you when that was. Market levels, interest rates, tax rules and local ordinances change. Never take a current cap rate, rate, rent or vacancy figure from a chat answer. Get market levels from current sources: lender quotes, appraisers, your brokerage's comps, paid data platforms, and published reports. The site's own cap rate pages are one starting point for context, not a substitute for deal-specific comps.

No access to your real data sources

A general assistant cannot see the MLS, CoStar, Crexi, LoopNet, the county property appraiser, the clerk's official records, or Sunbiz unless one of two things is true: you give it the document, or the tool has a connected feature that actually retrieves it. Even then, check that what it read is the record you meant. If you ask "who owns 1200 Example Parkway" without supplying a record, any name you get back is a guess.

Overconfidence and agreement

Models rarely volunteer "I don't know." They also tend to agree with the framing of your question. Ask "Why is this a great deal?" and you will get reasons. Ask "What could go wrong with this deal?" and you will get risks. Neither answer is analysis. You will learn to ask for both sides and to require the model to say when the source does not answer the question.

Missed content in long documents

Long files can exceed what the tool processes well. Pages may be skipped, tables flattened, scanned images ignored, and footnotes lost. The output will not always say so. Spot-check sections near the middle and end of a long document, not just the first pages.

The rule for this course: AI drafts, you verify and sign

Every workflow in the next six lessons follows the same order:

  1. You supply the source. The OM, lease, rent roll, T-12, public record or your own notes.
  2. AI drafts. A summary, table, letter, formula or checklist.
  3. You verify. Each fact and number goes back to the source. Anything unsupported is cut or labeled.
  4. You decide and sign. What goes to a client, a lender, or the public is your work product.

This is not just good practice. Your license makes it a requirement. A Florida licensee is responsible for the advertising and communications they put out, and false or misleading advertising is a ground for discipline under F.S. 475.25(1)(c). Your broker of record supervises your marketing and files. "The AI wrote it" is not a defense to a misleading cap rate in an email blast or a wrong lease date in a BOV. Course 27, Lesson 1 covers who is responsible for what in the brokerage.

Investors who are not licensed carry their own version of the same risk. If you sign a purchase contract based on an AI summary that missed a tenant's termination right, the tool will not be in the room when the tenant leaves.

Three kinds of AI tools you will meet

The tool landscape changes month to month, so this course stays tool-agnostic. Products, features, plan names and prices change often. What matters is the category and the questions you ask about it.

CategoryExamples of the typeGood forAsk before using
General chat assistantsChatGPT, Claude, Gemini, Microsoft CopilotDrafting, summarizing documents you upload, formula help, explanationsWhere does uploaded data go? Is it used for training? Can you turn that off?
AI features inside CRE data platformsSearch, summary and valuation features built into listing marketplaces, comp databases, and CRMsQuestions about the data that platform already licensesWhat sources does it draw from? What does your license let you republish?
Spreadsheet and office copilotsAI built into Excel, Google Sheets, Word, Docs, Outlook, GmailFormulas, cleanup of lists, email drafts inside your existing filesWhich files can it read? Who in your organization can see the output?

You will also see specialty tools: lease abstraction services, AI note takers for calls and meetings, and image tools for virtual staging and renderings. Each brings its own issues. Note takers raise recording-consent questions under Florida's all-party consent rule (see Course 27, Lesson 5). Image tools raise advertising and labeling questions, covered in Lesson 7.

For any tool, check the tool's current data and privacy settings before you upload anything that is not already public. Terms change. What was true when a colleague told you about a tool may not be true now.

Match the task to the risk

Not every task needs the same level of checking. Sort your uses by what happens if the output is wrong.

Risk levelExamplesVerification
LowRewriting your own email for tone, outlining a presentation, brainstorming questionsRead it before sending
MediumSummarizing a public market report, cleaning a public-records list, explaining a lease conceptCheck key facts and a sample of rows against the source
HighNumbers in an offer, BOV, loan request, OM, listing ad, or client emailEvery number and fact traced to a source and recalculated
Do not delegateLegal conclusions, tax positions, choosing market rent or cap rate, deciding who to market to based on personal characteristicsHuman professional judgment, with counsel or a CPA where needed

A worked example: the Kissimmee strip summary

Illustrative numbers. Not market data.

An agent is asked by an investor client for a quick read on a three-tenant retail strip in Kissimmee listed at $2,700,000. The agent pastes the seller's T-12 operating statement into a general chat assistant and asks for a one-paragraph summary with NOI and cap rate.

The T-12 shows:

LineAnnual amount
Base rent$216,000
CAM and tax recoveries$24,000
Total revenue$240,000
Real estate taxes$28,400
Insurance$17,900
Common area maintenance$9,100
Management$3,200
Total expenses$58,600

The AI summary reads, in part: "The property produces NOI of $186,400, a 6.90% cap rate at the asking price. The center sits outside the flood zone and all three tenants are on long-term leases."

The agent checks it line by line.

  • NOI. $240,000 minus $58,600 is $181,400, not $186,400. The model's figure is $5,000 high, a dropped or misread expense.
  • Cap rate. $181,400 divided by $2,700,000 is 6.72%. The AI's 6.90% is just its wrong NOI divided by the price ($186,400 ÷ $2,700,000).
  • Flood zone. Nothing in the T-12 mentions flood zones. The model filled a gap with a plausible fact. That claim needs the FEMA flood map and the survey, or it gets cut.
  • "Long-term leases." A T-12 does not show lease terms. That needs the rent roll and the leases.

The agent rewrites the summary with the correct numbers, removes the flood and lease claims, and asks the listing broker for the rent roll and leases. The AI saved drafting time. The review caught three problems that would have gone to a client under the agent's name. For how to normalize a T-12 before you rely on it, see The T-12 and normalized NOI.

A preview of confidentiality

Before you paste anything into an AI tool, ask three questions. Lesson 7 covers them in full.

  • Did you sign something? A confidentiality agreement or NDA for an OM often limits who may see the materials. An AI service may count as a third party under its terms. Read the agreement.
  • Is it a client's private information? Seller financials, a buyer's maximum price, a tenant's sales reports, and guarantor statements are confidential to your client.
  • Does it contain personal information? Social Security numbers, bank details, ID copies and personal phone lists do not belong in a general chat tool.

When in doubt, remove names and identifying details, use a business-tier tool your brokerage has approved, or do the task by hand.

Common mistakes

  • Asking the model for facts you should source. "What is the cap rate in Lake Mary?" gets you a number with no basis.
  • Trusting the citation. A page number or report title in an answer is not proof the page or report says that.
  • Skipping the math. Recalculate any number that will be repeated to someone else.
  • Uploading under NDA without reading the NDA. Confidentiality obligations do not pause for convenience.
  • Letting fluent writing pass as review. Polished output still needs a line-by-line check.

What to take away

  • A large language model predicts likely text. It is not a database and does not look things up unless a connected feature does
  • It is strongest at reshaping text you supply: summarizing, drafting, extracting, explaining and formula writing
  • Its main failures are invented facts and citations, arithmetic slips, stale knowledge, missing access to your data sources, and overconfidence
  • Without the document or a connected tool, it cannot see the MLS, CoStar, county records or Sunbiz
  • The course rule: you supply the source, AI drafts, you verify, you decide and sign
  • Under your license you remain responsible for what you send or advertise, whoever drafted it
  • Know the three tool categories and check each tool's current data and privacy settings before uploading
  • Match your verification effort to the risk of the task
  • Never take market levels such as cap rates or interest rates from a chat answer

Next lesson: Prompting for Commercial Real Estate Work.

Get Market Insights Delivered

Weekly Central Florida CRE updates — cap rates, new listings, market trends, and investment opportunities. No spam, unsubscribe anytime.