A buyer asks an AI assistant a reasonable question about a property:
The answer arrives immediately. It lists property taxes, an HOA fee, utilities and insurance. The numbers add up. The explanation is clear.
It is wrong in the most dangerous way: it sounds finished.
The HOA figure came from an old listing. The tax amount came from a portal record that has not caught up with the latest assessment. The arithmetic is correct; the answer is not.
Clear explanation. Correct arithmetic.
Polished output · two invalid inputs
The buyer now has more information, but not better information. The agent inherits a new job: find the error, explain it and rebuild confidence.
Generative AI can turn information into a convincing response. It cannot, by itself, decide which source is authoritative, whether a record is current, which calculation applies or who should approve the result.
AI has spread faster than its grounding
The tools are already in agents' hands, but the first wave has been dominated by general-purpose assistants. What is largely missing is reliable connection to current, approved property information.
ChatGPT was the most-used named AI tool in NAR's 2025 survey at 58%, followed by Gemini at 20% and Microsoft Copilot at 15%. Half of respondents reported a positive impact from AI, while 46% reported neutral or no noticeable impact.
The same survey shows where AI actually landed in the work: content generation leads by a wide margin, and the capabilities tied to a workflow trail well behind it.
Leading named assistants
The top three were all general-purpose
Where AI appears in the workflow
Writing leads; integrated applications remain much smaller
The survey is best read as a picture of the first wave of AI adoption — one dominated by general assistants and content generation. The potential impact of property-aware systems remains largely unmeasured.
Delta Media found the same content-first pattern: 82% of brokerage leaders reported agents using AI for property descriptions in 2024, and descriptions remained the leading application in 2025 even as adoption began expanding into client communications, data analysis and administrative work.
A general assistant can discuss real estate. A grounded system can show what is true about this property — and why.
A grounded system works from current, approved property information. It can show where an answer came from, recognize missing or conflicting information and involve the agent instead of filling the gap with something plausible. This makes the agent's knowledge more valuable, not less: AI can make approved facts, professional context and judgment available consistently — to more visitors, across more questions and at the moment they are needed.
Five layers sit behind a dependable answer
-
Approved evidence
The system needs sources it may treat as evidence: current disclosures, assessor records, leases, inspection reports, utility histories, MLS fields or professional observations. The useful questions are who supplied the fact, when it was current and whether something newer supersedes it.
-
Property context
A correct fact can still be attached to the wrong property, entity or period. Addresses vary. Documents expire. Asking rent is not contracted rent. Context establishes what a fact belongs to.
-
The appropriate method
Some questions require retrieval. Others require extraction, calculation or prediction. Generative AI may explain the result, but it should not silently substitute for the method that produces it.
-
Rules and professional judgment
The workflow must define what the system may answer, which evidence it must show, when it must say "I don't know" and who owns the next decision.
-
Action and feedback
An answer matters only when it improves an action. The outcome must then travel back: was the source wrong, did a professional override the output, and did the estimate match reality?
JLL's 2025 corporate and institutional real-estate research reaches a similar conclusion: successful programs depend less on acquiring a model than on quality data, integration, change management and connection to core workflows.
Not every AI question is a writing task
Calling every capability "AI" hides the most important design decision: which method is fit for the question?
| The job | Appropriate method | Common failure |
|---|---|---|
| Find the current HOA fee | Retrieve from an approved, dated source | Generate a plausible figure |
| Identify a lease obligation | Extract the clause and preserve its citation | Summarize away an exception |
| Estimate ownership cost | Calculate from stated inputs and assumptions | Mix current and historical values |
| Suggest a price range | Predict from comparables and uncertainty | Present one precise number as fact |
| Draft a listing description | Generate from approved property facts | Add unsupported claims |
| Prioritize maintenance | Detect patterns and apply operational rules | Treat correlation as diagnosis |
A property record is not a folder of files
Real-estate organizations often have plenty of information, but it is trapped in PDFs, portals, spreadsheets, inboxes and individual memory. Putting those files into one repository does not create a decision-ready property record.
Ten duplicated records do not become more reliable by volume.
A usable record distinguishes fact from claim, current value from historical value, observation from inference, and approved evidence from marketing copy.
That structure lets a system answer three apparently similar questions differently:
- "What is the HOA fee?" requires retrieval.
- "How current is that figure?" requires a visible source and date.
- "What if it rises by 10%?" requires calculation.
A language model can phrase all three responses. The underlying jobs are different.
Giving AI the documents is only the start
Connecting an AI assistant to approved sources is often called grounding. It is essential, but retrieval does not resolve every problem. A source may be stale, two documents may conflict or a calculation may depend on an unstated assumption.
NIST recommends the same basic discipline: check outputs against known facts, document limitations, test systems under realistic conditions and monitor how they perform in actual use. Fluency is not evidence of reliability.
Three questions for the next AI demonstration
The real shift is from content to decisions
Generative AI has made real-estate information easier to query, summarize and communicate. But that is the beginning of the workflow, not the end.
The durable opportunity is to connect evidence to context, context to the right method, the method to professional judgment, and the resulting action back to an outcome.
Real estate does not need AI that always has an answer. It needs AI that can show what the answer rests on, recognize when the evidence is insufficient and help the right person make the next decision.
Sources
- National Association of REALTORS®, 2025 REALTORS® Technology Survey, September 2025.
- Delta Media Group, 2024 Real Estate Leadership Survey, January 2024, and 2025 Real Estate Leadership AI Survey, January 2025; both surveys of brokerage leaders.
- JLL, "Real estate's AI reality check," October 2025.
- National Institute of Standards and Technology, Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, July 2024; particularly MAP 2.3 and MEASURE 1.1, 2.3, 2.5 and 4.2.
Reactions
No reaction from you
