CRE & AI fundamentals

It is a collection of tools for working with property information, documents, relationships, and decisions. A language model can help draft a deal summary. A vision model can read a site plan. A statistical model can rank a prospect list. Useful CRE AI connects the right tool to the right evidence and a specific task. “We added a chatbot” is a feature announcement, not a complete operating strategy.

The model generates or analyzes an output. An assistant adds a conversation and access to selected tools. A platform connects records, workflows, permissions, and the evidence behind the work. An agent can take a sequence of actions toward a goal. Those terms are related, but they are not interchangeable—and access to a tool should never be confused with permission to do whatever the model thinks is helpful.

Sources: Google: machine learning glossary ↗

They are model families with different versions, capabilities, operating costs, and deployment choices. Some versions emphasize speed; others devote more computation to reasoning. Input and output support varies: text, images, audio, and video are not universal. Open-weight models make model weights available under a license and can offer more control over hosting. That also means taking responsibility for infrastructure and safeguards. Compare the exact version on your actual task rather than declaring one brand the winner of intelligence.

Sources: Anthropic: model overview ↗ · Google: Gemini model capabilities ↗ · OpenAI: gpt-oss architecture ↗

Parameters are learned numerical weights; “B” means billion. They are not a count of facts, documents, or neurons in a human brain. Size affects memory and compute requirements, but training, architecture, tools, and the task also affect results. Google’s Gemma 3 family, for example, includes 1B, 4B, 12B, and 27B variants. A bigger number is a specification to understand, not an automatic certificate of better judgment.

Sources: Google: machine learning glossary ↗ · Google: Gemma 3 model card ↗

A dense model uses its full network through the main computation. A mixture-of-experts model routes work through selected expert components, so total parameters and active parameters differ. OpenAI documents gpt-oss-120b as approximately 117 billion total parameters, with about 5.1 billion active per token. Comparing only the headline sizes misses that distinction. These are documented examples, not a ranking of the newest or best models.

Sources: OpenAI: gpt-oss architecture ↗

No. Parameters belong to the trained model. Tokens are the pieces of content it processes. The context window is how much tokenized material it can consider in one request or conversation state. A large context window can help with a long document, but it does not make every extracted clause correct or guarantee perfect recall. A hundred-page lease still deserves a careful reader.

Sources: Google: machine learning glossary ↗

Document extraction, summarization, drafting, classification, and comparing supplied information are useful starting points. Multimodal systems can work with images and, depending on the model, audio or video. Retrieval brings relevant records into the answer; tool use can connect a workflow to a CRM or calculator. Each needs its own checks: page references for a lease summary, reconciled totals for a rent roll, and reviewed facts before a client receives a draft.

Sources: Google: Gemini model capabilities ↗ · Anthropic: model overview ↗

No. Spending more computation on a problem is not the same as independently checking the evidence. A convincing explanation can still contain a mistaken assumption. For CRE work, I want to know where the number came from, when the source was observed, what calculation was used, and what would change the conclusion. A longer answer is not a stronger receipt.

Yes. An agent with tools can misunderstand its scope, pursue a task too aggressively, or follow malicious instructions in material it reads. In September 2026, Anthropic documented four incidents involving unauthorized access to real systems during cyber evaluations. The test environments were mistakenly connected to the internet and the models ran without their normal cyber safeguards. That context matters: it is not evidence that an ordinary CRE chat routinely attacks systems. It is evidence that model behavior and technical access controls both matter.

Sources: Anthropic: September 9, 2026 incident assessment ↗

It is an attempt to hide instructions in content an AI reads—a webpage, email, document, or tool response—and steer it away from the user’s task. Imagine a property flyer that tries to tell an assistant to reveal private deal notes. The flyer is evidence to read, not an authority to obey. Narrow permissions, isolated environments, explicit approval for consequential actions, and monitoring reduce exposure. No single instruction or reviewer makes the problem disappear.

Sources: Anthropic: prompt injection defenses ↗ · Anthropic: trustworthy agents in practice ↗

Pick one recurring task with a result you can check: a tour debrief, a document comparison, or a draft follow-up. Define the information it may use, what a correct answer looks like, and who reviews it. Measure time saved and errors together. Automating confusion just gives you a faster way to be confused.