You do not need to become technical to buy AI well. You need to stop nodding along. A proptech rep says their platform uses “retrieval-augmented generation on your document corpus with human-in-the-loop review and 95% extraction accuracy,” and a principal who cannot decode that sentence is negotiating blind — unable to tell a real capability from a spec-sheet boast, or a confidentiality risk from a harmless feature. The vocabulary is not hard; it has just been explained badly, by people writing for engineers or writing to sell. This is the plain-English version: the thirty AI terms a CRE decision-maker actually meets — in demos, on pricing pages, in confidentiality conversations — each one anchored to work you already do, so you can sit in the next meeting and ask the right question.
This vocabulary is an edge in the one room where a small firm can lose real money: the vendor demo. The wider case for why a lean shop can out-operate a national brokerage is laid out in the small CRE firm manifesto. Terms below are grouped by the situation you meet them in, not alphabetically.
The Basics: The Words Behind the Words
These six sit under every other term. Get them straight and the rest falls into place.
Artificial intelligence (AI). The broad umbrella: software that performs tasks we associate with human judgment — reading, writing, summarizing, deciding. In CRE it is not one product; it is a capability now built into tools you already touch, from your email to your listing platform.
Machine learning (ML). A method of building AI where the system learns patterns from examples rather than following rules a programmer wrote by hand. A tool that predicts rent comps by studying thousands of past leases is doing machine learning. Most “AI” in proptech is machine learning underneath.
Generative AI. The kind of AI that produces new content — text, images, summaries — rather than just classifying or scoring existing data. When a tool drafts a property description or summarizes a lease, that is generative AI. It is the category behind almost every AI feature a broker uses day to day.
Large language model (LLM). The engine behind generative AI for text. An LLM is a system trained on an enormous amount of writing so it can predict and produce fluent language. ChatGPT, Claude, and Gemini are products built on LLMs. When a vendor says “powered by AI,” they almost always mean an LLM is doing the work.
Neural network. The underlying structure that makes modern AI work — loosely modeled on how neurons connect, it is the math that lets a model learn patterns from data. You will hear the word but never need to touch one. Treat it as the engine block: real, important, not something you service yourself.
Training data. The examples a model learned from. This term matters more than it sounds, because it raises the question every principal eventually asks: is my confidential deal data becoming someone’s training data? Hold that thought — it returns below.
The Model Vocabulary: What You Meet in a Demo
These are the words a salesperson uses when showing you how their AI works.
Prompt. The instruction you give an AI — the request that produces the output. “Summarize this lease’s rent schedule, escalations, and renewal options in a table” is a prompt. The quality of what you get back depends almost entirely on the quality of what you ask, which is why this one skill pays back faster than any other. It is worth its own treatment, covered in our piece on why the prompt is the one skill every CRE professional needs.
Prompt engineering. The practice of writing prompts well — being specific, giving context, showing an example of the format you want. It sounds technical; in practice it is closer to writing a clear brief for a sharp new analyst, and your team can learn the useful 90% in a workshop.
Token. The unit an LLM reads and bills in — roughly three-quarters of a word. It matters for two reasons: pricing (usage-based tools charge per token) and limits (a model can only hold so many tokens at once). When a vendor quotes per-token pricing, they are quoting per fraction-of-a-word.
Context window. How much text a model can consider at one time, measured in tokens. A small window means the model “forgets” the top of a long document by the time it reaches the bottom; a large one can hold an entire lease, or several, at once. When a tool struggles with a 90-page loan document, a small context window is often why.
Temperature. A setting that controls how predictable or creative the output is. Low temperature gives consistent, conservative answers — what you want for extracting lease terms. High temperature gives more varied, inventive output — useful for brainstorming marketing copy, risky for anything factual.
Multimodal. A model that handles more than text — images, PDFs, and scans, not just typed words. This matters in CRE because so much of your work arrives as a scanned PDF or a photograph of a document. A multimodal model can read the scan directly.
Chatbot / conversational AI. An interface where you type back and forth in plain language. ChatGPT is a chatbot; so is the assistant a proptech vendor bolts onto its dashboard. The interface is simple — the value depends on the model behind it and the data it can see.
How It Uses Your Data: The Confidentiality Terms
This is the group that keeps principals up at night, and the one where the vocabulary directly protects you. You handle NDA-bound financials; you need to know what a tool does with them.
Training on your data. Whether a vendor uses what you type to improve their model. This is the central confidentiality question. The business and enterprise tiers of the major AI products state they do not train on your content by default — but “by default” and “which tier” matter, so verify it in writing for the specific plan you are on. The full decision framework sits in our confidentiality guide for client deal data.
Retrieval-augmented generation (RAG). How an AI answers a question about your documents by looking them up first, then writing an answer grounded in what it found — instead of guessing from general knowledge. This is the technique that lets a firm sitting on twenty years of lease files ask “which of our tenants have cotenancy clauses?” and get a real answer with sources. When a vendor says “trained on your data,” they often actually mean RAG, which is safer — the documents are retrieved, not absorbed into the model.
Fine-tuning. Genuinely adjusting a model’s behavior by training it further on a specific set of examples. Unlike RAG, this changes the model itself. It is expensive and rarely necessary for a small firm; if a vendor pitches fine-tuning, ask why RAG would not solve the same problem for less.
Embedding / vector. The behind-the-scenes way RAG finds the right document — text is converted into a numerical fingerprint so the system can match your question to the most relevant passage. You do not need to manage this, but you will see “vector database” on technical proposals. It is plumbing, not a feature to pay a premium for.
Data residency. Where your data physically lives and which rules govern it. For a firm handling confidential US deal data, it is fair to ask a vendor where their servers are and who can access what. A vague answer is itself an answer.
Judging a Claim: The Buyer-Defense Terms
These six are the vocabulary of skepticism. They are how you separate a real capability from a marketing number.
Hallucination. When an AI states something false with complete confidence — an invented clause, a wrong square footage, a citation that does not exist. This is the single most important term in the list, because it is the reason no AI output touching money or legal terms should ship unchecked. A tool that cannot show you its source is a tool that can hallucinate one.
Accuracy / precision. How often the AI gets it right — but the number is meaningless without the test behind it. “95% accurate” on what documents, measuring what, against whose answer key? A precise-sounding figure with no defined test is a marketing claim, not a performance guarantee. Always ask what was measured and on what sample.
Human-in-the-loop. A workflow where a person reviews or approves the AI’s output before it counts. This is not a weakness; for CRE work it is the correct design. Any vendor selling “fully automated” lease review or underwriting for high-stakes decisions is selling you the hallucination risk without the safety net.
Ground truth / benchmark. The verified correct answers a tool is measured against. When someone claims accuracy, ground truth is the answer key they claim to have beaten. No ground truth, no credible accuracy claim.
Confidence score. A number the AI attaches to its own output estimating how sure it is. Useful when it is calibrated and honest, misleading when it is not — a model can be confidently wrong. Treat a confidence score as a hint about where to look harder, not as permission to skip review.
Guardrails. The limits built in to stop an AI from doing or saying certain things — refusing out-of-scope requests, flagging low-confidence answers, staying inside a defined task. Good guardrails are a sign of a serious product; their absence is a sign to slow down.
Getting Work Done: Agents and Automation
The last group is the vocabulary of AI that does not just answer, but acts — the terms behind custom projects and the more ambitious vendor pitches.
AI agent. A system that can take a goal, break it into steps, and carry them out — using tools, reading documents, and chaining actions rather than answering a single question. An agent might monitor a broker-blast inbox, screen each deal against your criteria, and file the survivors into your pipeline. The concept, and where it genuinely helps a small firm, is unpacked in our plain-English guide to AI agents for property professionals.
Workflow automation. Connecting steps so a process runs without someone shepherding it — an intake form that fans a new listing out to nine channels, or an approved invoice that posts itself. AI makes automation smarter by handling the judgment steps (reading, classifying, drafting) that rule-based automation never could.
Integration / API. The connection that lets two tools talk — the reason your AI can pull from your CRM or push into your accounting system. An “API” is just the doorway software uses to exchange data. When a vendor says “we integrate with Yardi,” they mean an API connects the two. When they cannot, you become the integration, copying data by hand.
OCR (optical character recognition). The technology that turns a scanned image of a document into machine-readable text. It is the unglamorous first step behind reading any old lease PDF, and it is where naive document tools quietly break — a smudged 1990s scan defeats weak OCR before the AI ever sees the words.
Copilot / assistant. An AI that works alongside a person inside an existing tool, suggesting and drafting rather than acting alone. Microsoft Copilot inside Outlook and Excel is the common example. The framing is honest — it assists, you decide.
Custom GPT. A version of ChatGPT you configure with your own instructions and reference documents — a lightweight way to build a repeatable assistant without writing code. A firm might build one that drafts LOIs in the house style. It is a useful on-ramp with real limits on what it can safely handle.
Which of These Actually Matter
Thirty terms is a lot to hold at once, and you do not need to. If you remember five, remember these: prompt (the quality of your ask determines the quality of your result), hallucination (why nothing ships unchecked), training on your data (the confidentiality question to settle in writing), human-in-the-loop (the right design for high-stakes work), and accuracy (a number that means nothing without the test behind it). Those five carry most of the weight in a buying decision.
The rest are there for the moments a vendor reaches for a word to make a simple thing sound proprietary — “we use embeddings in a vector store for retrieval-augmented generation” is often just “the tool can search your documents.” Vocabulary is how you tell substance from fog. Comparing the general-purpose tools your team would use daily is the natural next step, covered in our explainer on ChatGPT, Claude, and Gemini for real estate teams, and the structured path to getting a whole team fluent is laid out in the CRE AI training playbook.
If you want a faster route to confidence than reading definitions, a free AI-readiness assessment is a short working session that maps your firm’s real workflows — the leases you read, the deals you screen, the write-ups you draft — to where AI would save the most time, and translates the vocabulary against your actual tools rather than in the abstract. Book a free AI-readiness assessment and you will leave able to read a proptech pitch for what it is.
Frequently Asked Questions
What does AI jargon like “LLM” and “RAG” actually mean for a real estate firm?
An LLM (large language model) is the engine behind tools like ChatGPT that read and write language — it is what drafts your market write-up or summarizes a lease. RAG (retrieval-augmented generation) is how an AI answers questions about your own documents by looking them up first, so a firm with twenty years of lease files can query them and get sourced answers. Stripped of the acronyms, most AI jargon describes two things: a tool that reads and writes for you, and a way to point that tool at your own data safely.
Do I need to understand AI jargon to buy AI tools for my CRE firm?
You do not need to become technical, but you do need enough vocabulary to judge a vendor claim. The risk of not knowing the terms is that you pay for a “proprietary” feature that is really a standard capability, or miss a confidentiality risk hidden in plain language. Roughly thirty terms is the difference between negotiating a proptech purchase and nodding along to one.
What is the most important AI term for a principal to understand?
Hallucination — when an AI states something false with total confidence. It is the reason no AI output touching money, legal terms, or a client-facing document should ship without a human check. Every other buyer-defense concept, from human-in-the-loop review to demanding a source for every figure, exists to manage it.
What does it mean when a vendor says their tool is “trained on your data”?
It can mean two very different things, so ask which. The safe version is retrieval-augmented generation — the tool looks your documents up when answering, without absorbing them into a shared model. The version to scrutinize is a vendor using your content to train a model other customers benefit from. The business tiers of the major AI products state they do not train on your content by default, but confirm it in writing for your specific plan and tier.
Is it safe to put confidential deal data into an AI tool?
It can be, with a clear rule. The training question is manageable: business-tier tools from the major providers state they do not train on your content by default. The separate question is disclosure — whether pasting NDA-bound material into any third-party tool is permitted by that NDA. Most firms classify documents so public and internal material flows freely while genuinely confidential deal data is redacted or kept out.
What is the difference between an AI assistant and an AI agent?
An assistant suggests and drafts while a person stays in control — Microsoft Copilot proposing an email you approve. An agent takes a goal and carries out multiple steps on its own — monitoring an inbox, screening each deal against your criteria, and filing the results. The practical distinction is autonomy. For high-stakes CRE decisions, most firms want assistants and keep a human in the loop before anything counts.
How accurate are AI tools at reading leases and CRE documents?
Accurate enough to be useful, never accurate enough to trust unchecked — and any percentage a vendor quotes is meaningless without the test behind it. “95% accurate” says nothing until you know what documents, which fields, and against whose answer key. Treat every AI extraction as a fast first draft a person verifies against the source, especially for figures and clauses that carry legal or financial weight.
Where should a CRE firm start if the whole vocabulary feels overwhelming?
Start with one general-purpose AI tool, one real task, and the five core terms above. Pick something your team does constantly — summarizing a lease, drafting a market write-up — and practice writing good prompts for it, checking every output for hallucinations. The vocabulary stops feeling abstract the moment it attaches to work you recognize. A free AI-readiness assessment can map your workflows to the fastest starting points.
Arthur Wandzel
