Machine learning predicts from your numbers; generative AI produces from your prompt. A machine-learning tool studies thousands of past examples — closed sales, historical rents, lease terms — and learns to estimate the next one: what this building is worth, what this suite should rent for, which of your leads is most likely to transact. A generative-AI tool takes an instruction you type and creates new language from it: a letter of intent, a lease summary, a market paragraph. Both get called “AI” in vendor demos, which is exactly why the distinction is worth ten minutes of your time. The two categories are evaluated differently, trusted differently, and expose your confidential data differently — and knowing which one you are being sold tells you precisely what to test before you rely on it.
A word on why a principal should care before the definitions. You do not need to build either kind of system to benefit from telling them apart — you need enough of the vocabulary to judge a vendor claim and pick the right first move for a lean team. This piece is the plain-English version of that distinction; the broader glossary of terms you will meet in demos is laid out in our field guide to the AI jargon every CRE principal should know, and the case for why a small firm should invest in this literacy at all is the argument behind the small CRE firm manifesto.
The One-Sentence Difference
Machine learning answers “what is the likely number?” Generative AI answers “write me the thing.”
That is the whole distinction, and it holds up under pressure. When a tool looks at your rent comps and returns an estimated market rent, that is prediction — machine learning. When a tool takes your deal points and returns a drafted letter of intent, that is production — generative AI. One is fundamentally a very sophisticated calculator that has learned from history; the other is a very sophisticated writer that has learned from language.
They are not opposites, and they are not competitors. Generative AI is itself built on machine learning — it is a particular kind of model that happens to produce content instead of a forecast. So the honest framing is not “which should my firm use,” but “these are two different jobs, and I should recognize which one a given tool is doing.” A single proptech platform might use both: a predictive model to score deals and a generative model to write the summary of each one.
Where Each One Sits
The terms nest inside each other, which is the source of most confusion. Artificial intelligence is the broad umbrella — any software that does something we would call intelligent. Machine learning is the branch of AI that learns from data instead of following hand-written rules. Generative AI is a slice of machine learning, built on a technique called deep learning, whose specific talent is creating new content.
Picture four rings, one inside the next: AI contains machine learning, machine learning contains deep learning, and deep learning contains generative AI. Everything generative is machine learning; not everything machine learning is generative. The spam filter that quietly sorts your inbox is machine learning with no generative component. ChatGPT is generative AI sitting at the innermost ring. When a vendor says “our platform uses AI,” they have told you almost nothing until you learn which ring they mean, because the two ends of that diagram behave nothing alike in practice.
Machine Learning in a CRE Firm: Prediction From Your Numbers
Machine learning shows up in commercial real estate as prediction and classification — a model trained on historical data that outputs an estimate or a category. You have almost certainly used it without calling it that.
Automated valuation and rent estimates. When a platform gives you an estimated value or a suggested market rent, a machine-learning model is behind it — it has studied comparable sales and lease data and learned the relationship between a property’s attributes and its price. Tools like CoStar’s analytics and HelloData’s rent estimates are prediction engines. The output is a number with a confidence range, not a sentence.
Deal and lead scoring. A model that ranks your pipeline — flagging which prospects are most likely to transact, or which listings will move fastest — is machine learning classifying each record against patterns from deals that did and did not close. It sorts; it does not write.
Foot-traffic and demand forecasting. Location-analytics platforms such as Placer.ai use machine learning on mobility data to estimate visits to a retail center or trade area. For a broker underwriting a retail deal, that forecast is a predictive input, the same category as an AVM even though the subject is people rather than price.
The common thread: machine learning takes structured inputs — square footage, location, historical rents, comps — and returns a quantitative estimate learned from thousands of past examples. It is only as good as the data it trained on, and its answers are probabilities dressed as numbers.
Generative AI in a CRE Firm: Production From Your Prompt
Generative AI shows up as drafting — you describe what you want in plain language and it produces new text. This is the category a small firm meets through general-purpose tools like ChatGPT, Claude, Gemini, and Microsoft Copilot, and it maps directly onto the language-heavy work that fills a broker’s day.
Drafting deal documents. Give a generative tool your deal points and it returns a first-draft letter of intent in your structure; the highest-value early build for most brokerages is packaging exactly that repeated task, which our explainer on custom GPTs for CRE teams walks through end to end.
Summarizing long documents. Paste a ninety-page lease and a generative tool will pull the rent schedule, term, options, and key clauses into a readable abstract in about a minute. It is drafting a summary, not calculating one — which is why a person still has to verify every figure against the source.
Client and market language. From a market write-up grounded in figures you supply, to the meeting recaps produced by the AI note-takers now sitting in on client calls — covered in our piece on how AI note-takers work — generative AI handles the prose while you own the facts.
The common thread here is the mirror image of machine learning: generative AI takes an unstructured instruction and returns new language. It does not estimate a value; it composes a document. That difference in what each one produces is exactly what determines how you have to check it.
The Difference That Actually Matters: What to Distrust in Each
The reason a principal should care about this distinction is not taxonomy — it is that each category fails in its own way, so each demands a different kind of scrutiny before you rely on it.
Machine learning fails quietly and statistically. A predictive model gives you a number that looks authoritative but carries an error range the interface may not show. Its weak spots are stale or unrepresentative training data and thin comps — an AVM in a submarket with few recent sales is guessing more than estimating. The right scrutiny is the scrutiny you would apply to any analyst’s model: ask what data it learned from, how recent it is, and how it performed against known outcomes. Treat the estimate as an input to your judgment, not a replacement for it.
Generative AI fails loudly and specifically. A generative tool will state something false with complete confidence — an invented clause, a misremembered figure, a plausible citation to a document that does not exist. This behavior is called hallucination, and it is the single most important concept for anyone using these tools on real work. The right scrutiny is verification: every number, date, and legal term in a generated draft gets checked against the source before it leaves the office. The tool drafts; a person owns the facts.
Put simply: with machine learning you interrogate the data behind the number; with generative AI you verify the claims inside the text. Same word — “AI” — two entirely different review disciplines. A firm that applies the wrong one to the wrong tool either distrusts a useful forecast or ships a confident fabrication.
Which One Your Firm Meets First
For a 4–20 person firm, the practical order is nearly always the same: you meet generative AI first, deliberately, and machine learning second, usually without noticing.
Generative AI is the cheap, immediate entry point. A paid seat on ChatGPT, Claude, Gemini, or Microsoft Copilot costs about the price of a phone plan, works on day one, and applies to tasks every person on your team already does — drafting, summarizing, rewriting. It is where fluency starts and where the fastest return lives for a lean shop, because it attacks the language work that quietly eats hours.
Machine learning, by contrast, usually arrives bundled inside proptech you already pay for. You do not build an automated valuation model; CoStar, Buildout, or Placer.ai includes one. For a small firm, predictive machine learning is mostly something you buy inside a platform and evaluate, not something you stand up yourself. That is the reverse of how most technology press frames AI, and getting the order right matters: generative fluency compounds across your whole team, so a deliberate ninety-day effort pays back there first. The full sequence for building that fluency is laid out in the CRE AI training playbook.
The Confidentiality Line Runs Differently Through Each
The two categories expose your confidential data in different ways, and conflating them is how firms make avoidable mistakes.
With generative AI, the exposure is what you paste in. You hand the tool a document — a lease, an offering memorandum, a rent roll — and it processes that text. Two questions govern safety: whether the vendor trains its models on your input (business and enterprise tiers of the major tools state they do not, by default) and whether your NDA even permits handing the document to a third party at all. The rule that keeps a firm safe is to work on business-tier accounts and keep genuinely NDA-bound material out unless the agreement clearly allows it.
With predictive machine learning inside a proptech platform, the exposure is different: you are typically feeding structured inputs — an address, square footage, comps — into a vendor-hosted model, and the sensitivity depends on what those inputs reveal and what the vendor’s data-use terms say. The failure mode is less “I pasted a confidential document into a chat window” and more “I did not read how this platform uses the portfolio data I upload.” Both deserve care; they are simply not the same risk, and a single blanket policy that treats them identically will either over-restrict a safe tool or under-protect a sensitive one.
Decoding “AI-Powered” in a Vendor Demo
When a proptech vendor says “AI-powered,” treat it as an incomplete sentence and finish it with one question: which kind, and therefore what should I test?
If the answer is prediction — valuations, scoring, forecasting — your follow-ups are about the data: What did the model train on? How recent and how local is it? How does it perform where comps are thin? If the answer is generation — drafting, summarizing, extracting — your follow-ups are about verification: How do we catch hallucinations? What is the human-check step before this touches a client or a contract? A vendor who cannot say clearly which kind of AI their feature uses is either not being precise or does not know, and both are reasons to slow down.
This one habit — asking which kind of AI, then applying the matching scrutiny — is worth more than memorizing any definition. It turns a buzzword into a decision.
Where to Start
You do not need to pick a side between machine learning and generative AI, and you do not need to become technical to use the distinction well. The practical move for a lean firm is to get fluent with the generative tools your team touches daily, and to evaluate the predictive machine learning already bundled in your proptech with the right questions about its data.
A free AI-readiness assessment is a short working session that looks at your firm’s actual mix of brokerage, management, and acquisitions work, separates the language tasks where generative AI pays off first from the predictive tools you are already buying, and points you at the fastest, safest starting points for your team. Book a free AI-readiness assessment and you will leave knowing which kind of AI belongs on which task, what to test before you trust each one, and where the confidential-data line sits for your deals.
Frequently Asked Questions
What is the difference between machine learning and generative AI in simple terms?
Machine learning predicts from your numbers; generative AI produces from your prompt. A machine-learning tool studies historical data — past sales, rents, lease terms — and estimates the next value, like an automated valuation or a rent forecast. A generative-AI tool takes a plain-language instruction and creates new content, like a drafted letter of intent or a lease summary. They are related, not opposites: generative AI is a type of machine learning that happens to write instead of forecast. The useful test is what the tool returns — a number is prediction, a document is generation.
Is generative AI a type of machine learning?
Yes. The categories nest inside each other: artificial intelligence contains machine learning, machine learning contains deep learning, and generative AI is a slice of deep learning. So everything generative is machine learning, but not everything machine learning is generative. A spam filter or a deal-scoring model is machine learning with no generative component, while ChatGPT sits at the innermost ring. That is why “our platform uses AI” tells you very little until you learn which part of that diagram the vendor means.
What are examples of machine learning in commercial real estate?
The clearest examples are predictive: automated valuation models and market-rent estimates (the numbers behind CoStar analytics or HelloData rent estimates), deal and lead scoring that ranks your pipeline by likelihood to transact, and foot-traffic forecasting from mobility-analytics platforms like Placer.ai. In each case a model trained on historical data takes structured inputs — square footage, location, comps — and returns a quantitative estimate. The output is a number with an error range, not a written document, and its quality depends entirely on the data it learned from.
What are examples of generative AI in commercial real estate?
Generative AI shows up as drafting: producing a first-draft letter of intent from your deal points, summarizing a ninety-page lease into a structured abstract, writing a market paragraph from figures you supply, and generating meeting recaps through AI note-takers on client calls. A small firm meets it through general-purpose tools like ChatGPT, Claude, Gemini, and Microsoft Copilot. In every case the tool takes a plain-language instruction and returns new text, which a person then verifies — the tool drafts, but a human owns the facts and figures.
Which is better for a small CRE firm, machine learning or generative AI?
Neither is “better” — they do different jobs, and a lean firm uses both without building either. You start with generative AI because it is the cheap, immediate entry point: a paid seat costs about the price of a phone plan and applies to the drafting and summarizing every person already does. Predictive machine learning usually arrives bundled inside proptech you already pay for, like a valuation model in CoStar or a forecast in Placer.ai, so your job there is to evaluate it, not build it. The fastest return for a small team comes from generative fluency first.
How do machine learning and generative AI fail differently?
Machine learning fails quietly and statistically: it returns an authoritative-looking number that carries a hidden error range, and it weakens when its training data is stale or the comps are thin. Generative AI fails loudly and specifically: it states something false with total confidence — an invented clause or a misremembered figure — a behavior called hallucination. The scrutiny differs accordingly. For machine learning you interrogate the data behind the number; for generative AI you verify every claim inside the text before it leaves the office. Applying the wrong review discipline is how firms get burned.
What does a vendor mean when they say their tool is “AI-powered”?
It is an incomplete claim until you ask which kind. If the feature predicts — valuations, scoring, forecasting — the right follow-ups are about the data: what it trained on, how recent and local it is, and how it performs where comps are thin. If it generates — drafting, summarizing, extracting — the follow-ups are about verification: how you catch hallucinations and what the human-check step is before output touches a client or a contract. A vendor who cannot say clearly which kind their feature uses is a reason to slow down, not speed up.
Is it safe to put confidential deal data into these tools?
It depends on the category. With generative AI the exposure is what you paste in, governed by two questions: whether the vendor trains on your input (business tiers of the major tools state they do not, by default) and whether your NDA permits handing the document to a third party at all. With predictive machine learning inside a proptech platform, you are usually feeding structured inputs to a vendor-hosted model, so the concern shifts to that platform’s data-use terms. A blanket policy that treats both identically will either over-restrict a safe tool or under-protect a sensitive one.
Do I need to understand this difference to buy AI tools for my firm?
You do not need to become technical, but you need enough vocabulary to finish the sentence “AI-powered” with “which kind, and what should I test?” That one habit turns a buzzword into a decision: prediction tools get data scrutiny, generation tools get verification scrutiny. Without it, you risk paying for a standard capability dressed up as proprietary, or trusting a confident fabrication because it looked like a calculation. The distinction is the difference between negotiating a proptech purchase and nodding along to one.
Where should a small CRE firm start with AI?
Start with generative AI on one real, repeated task your team already does — summarizing a lease, drafting a market write-up — using a business-tier account so your data is handled correctly. Get a few people genuinely fluent before packaging anything, and evaluate the predictive machine learning already bundled in your proptech with the right questions about its training data. A free AI-readiness assessment can map your specific workflows to the fastest, safest starting points and tell you which kind of AI belongs on which task.
Dirk Jan van Veen, PhD
