AI Basics: The Terms Real Estate Teams Need

LLM (Large Language Model)

A Large Language Model, or LLM, is the foundation behind many modern AI systems. For real estate teams, the point is not the acronym. It is that LLMs can work with the material the industry actually runs on: PDFs, emails, rent rolls, leases, investment memos, reports and fragmented notes. The value does not come from the model alone. It comes from context, sources, permissions, review and integration into real deal workflows.

An LLM is a model that processes language. It can follow instructions, summarize text, extract information, classify documents and generate new text or structured outputs. In real estate, that might mean pulling key facts from a teaser, flagging lease clauses or preparing a first IC memo draft from scattered notes.

The boundary matters. An LLM is not a database, not a search engine and not an automatic truth layer. It works with the context it receives. If that context is incomplete, outdated or wrong, the answer can be polished and still unusable.

Used professionally, an LLM is closer to a language processing engine than an autonomous expert. It speeds up reading, structuring and drafting. It does not replace data management, source control or investment judgment.

Real estate work is document-heavy and context-heavy. An acquisition decision rarely lives in one clean system. It emerges from a teaser, rent roll, leases, technical reports, emails, broker notes, underwriting model, legal comments and IC material.

LLMs can make that material easier to read, structure and reuse. They can prepare lease abstracts, normalize tenant data, group risks by theme, draft due diligence questions, create management summaries or surface inconsistencies across documents.

The operational value is not chatting with a model. It is helping a team prepare the same decision faster, more consistently and with better traceability.

RAG

RAG stands for Retrieval-Augmented Generation. The simple version: before the AI answers, it retrieves relevant information from documents or systems and uses that context to produce a response.

RAG combines search and generation. The retrieval part finds relevant passages, files or records. The generation part uses an LLM to answer in natural language or produce a structured output.

Instead of asking a model to rely on general training, RAG gives it context from the company material that matters for the task.

A leasing manager might ask which lease contains a break option. An investment manager might ask which rent roll line explains a vacancy risk. A portfolio team might ask which assets have upcoming capex exposure.

RAG helps the AI retrieve the relevant documents first, then answer with the right context. That is far more useful than a generic response.

MCP

MCP stands for Model Context Protocol. It is a standard for connecting AI systems to tools, data sources and applications in a more structured way. The term sounds technical, but the business question is simple: how does AI get controlled access to the systems it needs?

MCP gives AI applications a common way to access external context and tools. Instead of every integration being custom, an MCP server can expose a system, file store, database or tool to an AI client in a structured format.

In plain language: MCP is one way AI moves from a chat window toward connected work.

Real estate work lives across data rooms, CRMs, email, SharePoint, Excel, BI tools, reporting folders and asset-management systems. AI becomes more useful when it can connect to those places under clear rules.

For example, an AI workflow could read documents, check a deal record, update a task list or retrieve a rent roll if the integration layer allows it.

AI agent

An AI agent is a system that can pursue a task through multiple steps. It can use tools, read context, make intermediate decisions and produce outputs that a human can review.

A chatbot answers a message. An agent works through a process. That might mean reading documents, extracting fields, checking missing information, creating a draft and escalating exceptions.

The agent still needs boundaries. It should know what it may access, what it may do and where a human must approve.

A deal intake agent could monitor incoming opportunities, read a teaser, extract key facts, compare them with an investment profile, flag missing data and prepare a first screening view.

That is more valuable than a chat answer because it sits inside the workflow.

Multi-agent system

A multi-agent system uses multiple specialized agents that work together on a larger task. In real estate, that makes intuitive sense: transactions already involve specialized work across documents, leases, underwriting, legal, financing and IC preparation.

Instead of asking one AI system to do everything, a multi-agent setup splits work into roles. One agent might extract rent-roll data, another reviews lease risks, another prepares due-diligence questions and another drafts an IC summary.

The value is specialization. The risk is coordination.

Real estate transactions are not one task. They are a chain of specialized judgments. A single general assistant can help, but specialized agents map better to the actual process.

A strong real estate AI system can separate document extraction, lease interpretation, underwriting support, risk review and memo generation.

AI hallucinations

An AI hallucination is an output that sounds plausible but is not supported by the underlying facts. For institutional real estate teams, hallucinations are not an abstract AI problem. They can affect underwriting, due diligence, IC material and reporting.

LLMs generate language based on patterns and context. They do not automatically know what is true. If the context is missing, ambiguous or badly retrieved, the model may still produce a confident answer.

That answer can be wrong even when it sounds professional.

A hallucinated lease expiry, rent number, covenant or vacancy assumption can create real risk. The danger is not only that the answer is wrong. It is that the answer can look polished enough to pass casual review.

That is why AI output should not enter investment materials without sources and human approval.

Evals

Evals are evaluations: structured tests that measure whether an AI system performs a task well enough. For real estate teams, they are the difference between a vendor demo and production confidence.

An eval defines inputs, expected outputs and scoring criteria. Instead of asking whether AI feels good, it asks whether the system extracted the right fields, cited the right sources or produced the right workflow artifact.

In short, evals are quality assurance for AI workflows.

A real estate eval might test 100 rent rolls, 50 lease abstracts or 30 IC memo drafts. The team can measure extraction accuracy, missing fields, source quality, formatting fit and reviewer time.

That is much stronger than deciding based on one polished demo.

Source grounding

Source grounding means AI outputs are connected to the evidence behind them. The system should not only answer. It should show where the answer comes from.

A source-grounded answer links material claims back to documents, rows, clauses, pages or approved assumptions. If the AI says a lease expires in 2029, the reviewer should be able to open the source.

This turns AI from a fluent assistant into a reviewable workflow component.

Investment decisions depend on evidence. Rent, vacancy, lease terms, capex, debt terms and risk notes cannot be accepted because a model wrote them well.

Source grounding lets professionals review faster because they do not have to hunt for the evidence manually.

Human-in-the-loop

Human-in-the-loop means a person remains part of the workflow at the points where judgment, accountability or approval matter. In real estate, that is not a limitation. It is how AI becomes usable in high-stakes work.

AI can extract, summarize, classify, draft and flag issues. A human reviews, corrects, approves and decides.

The point is not to slow the system down. The point is to put human judgment where it has the highest value.

An AI system can prepare due-diligence questions, flag lease risks or draft an IC memo. But the investment team owns the decision.

Human-in-the-loop design makes that ownership explicit and auditable.

AI operating layer

An AI operating layer is the system layer that connects AI to the way a company actually works. It brings together documents, data, workflows, permissions, review, outputs and integrations.

An AI feature might summarize a PDF. An AI operating layer connects that summary to the deal, the source documents, the underwriting workflow, the IC memo, the task list, the portfolio view and the audit trail.

It is not only about what the model can answer. It is about how AI moves through the organization.

Real estate work is fragmented across files, emails, models, presentations, data rooms and reporting tools. A single AI assistant can help, but it does not solve continuity.

An operating layer preserves context from deal intake through underwriting, due diligence, IC, asset management and reporting.