Why ChatGPT Won't Fix Your Property Business

Two executives try the same experiment. Each opens a chat window, types a real question about their portfolio, and reads the answer.

The first gets something generic, a competent-sounding list of considerations that could apply to any property company anywhere, containing nothing they did not already know. They close the tab and conclude that AI is overhyped, at least for a business like theirs.

The second asks a broader question and gets back something articulate, structured, and impressive. They conclude that this changes everything, and walk into their next leadership meeting with considerably more conviction than information.

Both have just formed a durable judgment about what AI can do for their business by using a tool that has no access to their business. It cannot see their rent roll, their maintenance history, their entity structure, their vendor performance, their lease terms, or a single number in their general ledger. It answered about property companies in the abstract, because that is the only thing it was in a position to answer about. And both executives, the skeptic and the enthusiast, have now miscalibrated in opposite directions from the same flawed test.

What a general model has and does not have

It is worth being concrete about the information condition, because the whole confusion follows from it.

A general-purpose model has read an enormous amount of public text. It knows what a lease is, how NOI is calculated, what typically causes turn delays, and what property managers generally worry about. It can reason, explain, draft, and structure. That is real capability and it is not nothing.

What it does not have is any access to your specifics. It does not know that your Riverside property has the boiler that fails every February, that one owner insists on variance explanations before the report lands, that your intercompany allocations were set up in a way that makes the close painful, or that your turn times are twelve days worse than they should be. It does not know your numbers. So when you ask it a question about your business, it cannot answer about your business. It answers about the category, and the answer sounds like advice from an intelligent consultant who has never visited.

This is not a limitation of the model's intelligence. A brilliant consultant with no access to your data would produce the same generic output, and for the same reason.

This is how most enterprise judgment is being formed

The reason this matters at a leadership level is that the ungrounded consumer experience is not a marginal way people encounter AI at work. It is the dominant one.

Verizon's 2026 Data Breach Investigations Report found that 45% of employees are now regular AI users on corporate devices, up from 15% the year before, and that 67% of users accessing AI services on corporate devices are doing so through non-corporate accounts. Two-thirds of workplace AI use is happening through personal accounts on consumer tools with no connection to the organization's systems.

So the organization's collective impression of what AI can do is being assembled, at scale, from experiences with a tool that is structurally incapable of engaging with the organization's actual work. That impression then shapes budgets, strategy, and executive conviction. It is a large-scale calibration error, and almost nobody has named it as one.

The two errors, and why they have the same cause

Dismissal - An executive tests AI on a hard, specific problem, receives a generic answer, and concludes the technology is not ready for their industry. The reasoning feels rigorous, since they tested it on real work. But they tested it under conditions where it had no possibility of succeeding, then generalized from the failure. It is like judging an analyst's ability by asking them questions about accounts they have never been given access to. The answer tells you about the access, not the analyst.

Overestimation - The opposite error is subtler and probably more expensive. An executive asks something they cannot easily verify, receives a fluent and well-organized answer, and reads the fluency as competence. This is the trap covered in the AI that makes you look bad: polished output disarms scrutiny precisely when scrutiny is most needed. Ungrounded models are at their most convincing exactly where they are least reliable, because with nothing specific to be wrong about, nothing contradicts them.

Both errors come from the same missing ingredient. Without grounding in real data, the output is untethered from your situation, and an untethered answer can be dismissed as useless or admired as brilliant with equal ease, because there is nothing in it to check.

The honest part: general AI is genuinely useful

It would be a serious overcorrection to conclude from this that consumer AI tools have no place, and the sneering version of this argument is wrong in a way that costs organizations real value.

There is a large class of work where the general model is exactly the right tool, because the task does not require your specifics. Drafting a first version of a policy. Explaining an unfamiliar accounting treatment. Summarizing a long document you paste in. Rewriting something for a different audience. Thinking through the structure of a problem out loud. Translating. Producing a first draft of anything that a human will then make specific. For all of these, general knowledge and language ability are the whole requirement, and the tool performs well.

There is also a legitimate case where generic advice is what you actually want. If you have never run a formal turn process and you want to know how one is typically structured, general knowledge is the correct source, and asking for it is sensible.

The DBIR finding does carry a real warning alongside the calibration point, which is worth stating plainly: the most common data type employees upload to unauthorized AI tools is source code, followed by internal documents and technical documentation. People attempting to supply the missing context are pasting proprietary material into systems the organization cannot see or govern. That is a genuine exposure, and the answer to it is providing sanctioned tools rather than pretending the behavior will stop.

What actually changes the answer

The variable that separates a generic response from a useful one is not model quality. It is whether the model has access to your actual records at the moment you ask.

The same question, asked of the same model, produces categorically different output depending on whether it can see your data. Ask about turn times in the abstract and you get a list of general causes. Ask with access to your actual work orders, unit history, and vendor records, and the question becomes answerable: which units, which stage, which vendor, how many days, compared to what. One is a search result. The other is analysis.

This is why evaluating AI through an ungrounded chat interface tells you so little about its enterprise potential, in either direction. You have been testing the wrong variable. The model was never the constraint. The information condition was.

How to actually find out

If you want a real answer to what AI can do for your business, the test has to meet two conditions, and neither is difficult.

Give it your real data - Not a description of your situation, the actual records. Until the tool can see what you see, you are testing its general knowledge, which you can already predict.

Pick a task where you know the right answer - Choose something you have already done manually and can check. If you know the correct output, you can evaluate the tool's version honestly instead of being impressed by fluency you cannot verify. Testing on questions you cannot check is how the overestimation error happens.

That is a small, cheap test, and it is a different thing from a formal pilot, which has its own design requirements covered in what a real pilot looks like. The point here is narrower: before you form any strategic view about AI in your business, make sure the experience your view is based on involved your business at all.

Most executive conviction about AI right now, in both directions, rests on a test that could not have produced a meaningful result. The confident skeptics and the confident enthusiasts are drawing opposite conclusions from the same non-experiment. The useful position is neither, and it is available cheaply: stop asking a stranger about your portfolio and go find out what happens when it can actually see it.

FAQs

Q1. Why does ChatGPT give generic answers about my business?
Because it has no access to your business. A general-purpose model has read a great deal of public text and can reason well, but it cannot see your rent roll, maintenance history, entity structure, or ledger. When asked about your situation, it answers about the category, since that is the only information available to it. The genericness reflects the access, not the capability.

Q2. So is AI overhyped for property companies?
That conclusion is not supported by the experiment most people ran to reach it. Testing an ungrounded tool on a question requiring your specific data guarantees a generic answer regardless of how capable the technology is. The honest position is that you have not yet run a test that could tell you either way.

Q3. What is the opposite error?
Reading fluency as competence. When you ask an ungrounded model something you cannot verify, it produces an articulate, well-structured answer with nothing specific to be wrong about, and confidence in the output rises without any corresponding rise in reliability. This is generally more expensive than dismissal, because it drives spending rather than inaction.

Q4. How common is ungrounded consumer AI use at work?
Dominant. Verizon's 2026 DBIR found 45% of employees are regular AI users on corporate devices, up from 15% the prior year, with 67% of that access happening through non-corporate accounts. Most organizational experience of AI is therefore occurring on consumer tools with no connection to company systems.

Q5. Is there a security dimension to this?
Yes. The same DBIR found that the most common data type uploaded to unauthorized AI tools is source code, followed by internal documents and technical documentation. People are attempting to supply the missing context by pasting proprietary material into systems the organization cannot see. Providing sanctioned, connected tools addresses both the calibration problem and the exposure.

Q6. What is general AI genuinely good at?
Anything where your specifics are not required: drafting first versions, explaining unfamiliar concepts, summarizing documents you provide, rewriting for a different audience, structuring a problem, translating. For this class of work, general knowledge and language ability are the entire requirement and the tools perform well. Dismissing them wholesale forfeits real value.

Q7. What actually makes AI useful on business-specific questions?
Access to the relevant records at the moment of the question. The same model, asked the same thing, produces categorically different output depending on whether it can see your data. Without it you get a list of general considerations. With it you get analysis of your actual situation. The differentiator is the information condition, not the model.

Q8. What is the cheapest way to find out what AI could do for us?
Run a small test that meets two conditions: give the tool your real data rather than a description of it, and pick a task where you already know the correct answer so you can evaluate the output rather than being impressed by it. That takes very little effort and produces a genuine signal, unlike the ungrounded chat experiment most strategic views are currently based on.