Every AI glossary for real estate agents explains what the words mean. Almost none of them tell you what to do about it, which is the only reason an agent would look one up.
So each entry here has three parts: what the term actually means, how it shows up in your work, and the question to ask the person who just used it on a sales call. That last part is the useful one. Most of these words are deployed to sound impressive, and one specific question usually reveals whether there is anything behind them.
The words behind the chatbot
Large language model (LLM). The engine behind ChatGPT, Claude, Gemini and most tools that write text. It predicts what words come next, very well, having read an enormous amount of writing. It does not look anything up unless it has been specifically wired to. Ask: “Is this your own model or are you reselling someone else’s?” Most real estate AI products are a friendly interface on top of a model built elsewhere, which is fine — but it changes what you are paying for.
Hallucination. When the model states something false with complete confidence. In your work it looks like a comparable that does not exist, an invented year built, or a county rule it has imagined. It is not a bug being fixed next quarter; it is a property of how these systems work. Ask: “Where did that figure come from and can you show me?” If the answer is that the model knows, it does not know.
Prompt. What you type in. The quality of the output tracks the quality of the prompt more than anything else, which is why the same tool works brilliantly for one agent and uselessly for another. Ask: “Can I see the actual prompt this sends?” Products that hide it are selling you a prompt you could have written.
Token. Roughly three quarters of a word. Models are billed by the token, so it is the unit behind almost every usage limit you will meet. Useful mainly for translating “500,000 tokens a month” into “about 375,000 words”.
Context window. How much text the model can hold in mind at once. Exceed it and the earliest part of the conversation falls out, silently. This is why a long chat starts contradicting itself and why a 60-page inspection report gets summarised badly. Ask: “How long a document can this handle in one go?”
Training data and cutoff. What the model read, and when it stopped reading. A model with a cutoff eighteen months ago has never heard of a rule that changed last spring. Ask: “Does this know about anything that happened this year?”
Fine-tuning. Taking an existing model and training it further on specific material so it adopts a style or a domain. Genuine fine-tuning is expensive and rare. Ask: “Fine-tuned on what, and how many examples?” Vague answers usually mean nobody fine-tuned anything.
RAG (retrieval-augmented generation). The model looks something up in a real database first, then writes using what it found. This is the honest version of “trained on MLS data” and it is what you actually want, because the answer traces back to a source. Ask: “Does it retrieve from live data or is it working from memory?”
Agent, agentic. Confusingly, not you. Software that takes multiple steps on its own — reads an email, checks a calendar, drafts a reply, sends it. The interesting question is never the capability, it is the leash. Ask: “What can it do without me approving it first?”
Temperature. A setting between predictable and creative. Low temperature for anything factual, higher for marketing copy. Rarely exposed to you, but it explains why the same prompt gives a different answer twice.
The words behind the pictures
Diffusion model. The technique behind almost all AI image generation. It starts with visual noise and refines it into a picture. Worth knowing only because it explains the characteristic failures: furniture floating a centimetre off the floor, a chair leg passing through a radiator, shadows falling where the window does not allow.
Inpainting, generative fill. Replacing part of an image while leaving the rest alone. This is the actual mechanism of virtual staging and of object removal. Ask: “Does this touch anything outside the area I selected?” The answer should be no.
Object removal. Taking things out rather than putting them in — the toaster, the dish rack, the bins. Usually priced separately from staging, and often what a cluttered kitchen actually needs. Ask: “Is removal included or is it a separate line?”
Upscaling. Increasing resolution by inventing detail that was never captured. Fine for a hero image, dangerous the moment the invented detail is a feature of the property. It does not recover what the camera missed; it guesses.
Deepfake. A synthetic image, video or voice of a real person. In this industry it turns up in seller impersonation: a forged licence, a face on a remote notarization call. Ask nothing of a vendor here. The defence is verifying facts through a channel the impostor does not control, not detecting the fake by eye.
Content Credentials, C2PA. An open standard for recording where an image came from and what was done to it — described by the coalition behind it as “like a nutrition label for digital content”. Its steering committee includes Adobe, Google, Microsoft, OpenAI, Amazon, Sony and the BBC, so it is likely to be the mechanism behind any future disclosure requirement for listing photos. Worth knowing before your MLS asks about it.
The words on the product page
“AI-powered”. Means nothing on its own. A spreadsheet formula qualifies if the marketing team is bold enough. Ask: “What specifically does the AI do that the previous version could not?” A vendor with a real answer gives it immediately.
“Trained on MLS data”. Usually false as stated, and if literally true it raises a licensing question the vendor may not want asked. Nearly always this means retrieval — the tool queries MLS data when you ask it something. Ask: “Trained, or querying? And under what agreement with the MLS?”
“Proprietary model”. Occasionally true. More often it means proprietary prompts wrapped around somebody else’s model. Not a problem — just do not pay a proprietary-model price for an interface. Ask: “Who trained it, and on what hardware?”
Automation versus AI. Automation follows rules you set: if a lead does not reply in three days, send email two. AI generates something new each time. Automation is more reliable, cheaper, and covers most of what agents actually want. A lot of what is sold as AI is automation with better branding, and that is not necessarily a bad deal — but it should not carry an AI price.
Integration, API. Whether the tool talks to what you already use. The gap between “integrates with your CRM” and “integrates with your CRM” is where a lot of wasted subscriptions live. Ask: “Show me it working with mine, not with the demo account.”
Seat. One user. Team pricing is usually per seat per month, and it is where a cheap-looking tool becomes expensive at brokerage scale. Ask: “What happens to the price at ten seats?”
Credits. Prepaid units of usage. The question that matters is whether they expire. Credits that never expire are a genuinely different product from a monthly allowance that resets, and the per-unit price is not comparable between the two. Ask: “Do unused credits roll over?”
The words that carry legal weight
Disclosure. Telling people that something was AI-generated or altered. For listing images this is a legal requirement in a growing number of states, and portals expect it regardless. It is the term on this list most likely to cost you money if you ignore it.
Human in the loop. A person reviews the output before it goes anywhere. This is the entire distinction between a tool that saves you time and a tool that publishes your mistakes. Ask: “At what points does a person have to approve something?”
Bias, disparate impact. A system that produces different outcomes for protected groups, whether or not anyone intended it. This is live in real estate: tenant screening, lead scoring, and any AI touching who sees which listing. Fair housing law cares about the effect, not the intent. Ask: “Has this been tested for disparate impact, and can I see the result?”
Training on your data. Whether what you type gets used to improve the vendor’s model. Matters the moment you paste a client’s financial position into a chat box. Ask: “Is my input used for training, and can I turn that off?” Business tiers usually say no; free tiers frequently say yes, in the terms nobody reads.
Data retention. How long they keep what you sent. Separate from training, and separate again from who can see it internally. Ask: “How long do you store my prompts, and who has access?”
Prompt injection. Hidden instructions inside a document or web page that hijack an AI reading it. Relevant the moment you let a tool read incoming emails or listing documents from strangers. Ask: “What stops a document from giving this thing instructions?”
Audit trail. A record of what the system did and when. Boring until a transaction goes wrong, at which point it is the only thing that helps you. Ask: “Can I export a log of what this sent on my behalf?”
The words specific to this industry
AVM (automated valuation model). A statistical estimate of a property’s value from public and market data. A Zestimate is one. It is a starting point with a published error rate, not an appraisal, and the error rate is the part sellers never see. Ask: “What is the median error rate, and on what sample?”
Lead scoring. Ranking leads by how likely they are to transact. Often just a rules engine. Its real risk is quiet: if the model learned from your historical conversions, it may be reproducing patterns you would not defend if someone asked you to explain them.
Conversational AI. A chatbot that talks to leads. Fine for qualifying and scheduling. The line to watch is where answering questions about a property becomes giving advice about it, and that line is drawn by your state’s licensing law rather than by the vendor.
Voice cloning, synthetic voice. A generated voice, sometimes a copy of a specific person’s. If it makes calls on your behalf, federal rules treat it as an artificial voice with everything that follows. Ask: “What does this disclose at the start of a call?”
Transcription. Speech to text, for showings and buyer consultations. Genuinely good now. The consent question is separate and it is a state matter: some states require every party to a recorded conversation to agree.
Sentiment analysis. Scoring a message as positive or negative. Sold as reading buyer intent. It reads punctuation and word choice, and it does not know that your client types in curt sentences because they are always on a job site.
The three questions underneath all of them
If you remember nothing else from this page, these cover most of it:
- Where did that come from? Applied to any number, fact or comparable a tool produces. It separates retrieval from invention.
- What can it do without asking me? Applied to anything that sends, posts, calls or signs. It is the difference between a tool and a liability.
- What happens to what I type? Applied before any client information goes near a text box.
None of these require you to understand how the technology works. They require the vendor to demonstrate that they do.
Sources
- Coalition for Content Provenance and Authenticity, for the Content Credentials description and steering committee membership.
Checked September 2026. Definitions here describe how these terms are used in practice by vendors selling to agents, which is occasionally looser than how a computer scientist would use them. Nothing here is legal advice.