The Questions That Don't Make It Into the Demo

Every AI vendor demo in real estate follows a script. The data looks clean. The model answers fast. Someone asks about occupancy trend in Q3 and the answer comes back correctly. Everyone nods. The meeting ends.

What does not make it into the demo is what happens when something goes wrong.

This is not a knock on demos. Demos are supposed to show the best case. But if you are considering connecting an AI model to your portfolio data through MCP, you are not buying a demo. You are buying something that will run on live data, answer questions your analysts and asset managers are going to act on, and handle information that your investors, your lenders, and your tenants trust you to keep controlled.

Before you do that, you should have answers to some uncomfortable questions.

Where is the data coming from, and how current is it?

A model answering from yesterday's snapshot is a different product than a model answering from live data. Both are valid. They are not the same thing, and the difference matters in ways that are easy to miss until they are not. Ask what the timestamp is on the data backing any answer you receive in a demo. If the vendor does not know without checking, they have not thought carefully about it.

Is my data being used to train someone else's model?

In most responsible implementations, no. But "most responsible implementations" is doing a lot of work in that sentence. Ask the question directly. Get the answer in writing. If the vendor seems annoyed that you asked, that is information.

How is sensitive information handled?

Financials. PII. Tenant data. Lease terms. Can sensitive fields be redacted before they reach the model? Is access permissioned by user and role, so a site-level property manager cannot inadvertently surface portfolio financials by asking a question? If the honest answer is "we're working on it," that is fine — these problems are not trivial to solve. But you deserve to know the current state before you connect anything.

Can I trace an answer back to the source?

If the model tells you occupancy dropped at Property C, you should be able to click through to the underlying transactions that explain why. This is not a feature. It is the baseline requirement for trusting the answer enough to act on it. A system that gives you conclusions without a clear path to the evidence is not a tool you can use in a meeting with your partners or lenders. No trace, no trust.

What happens when the model is wrong?

Because it will be. Sometimes the data is inconsistent. Sometimes the question is ambiguous. Sometimes the model just misreads. None of this is disqualifying — it is the normal operating condition of any complex system. What matters is whether you have a workflow for catching and correcting errors before they become decisions. Ask the vendor how a wrong answer surfaces, and who is responsible for fixing it. If they tell you it rarely happens, the conversation is over.

None of these are reasons not to use MCP. They are reasons to use it carefully, and to pick vendors who have already worked through the answers.

The vendors who have not will tell you they're working on it. The vendors who have will walk you through their approach without waiting to be asked. That distinction is worth more than any feature in the demo.

The question you are actually trying to answer is not "which AI should I use." It is: what does the data layer underneath it look like, and can it be trusted? Before you get excited about an AI rollout, ask your team:

• What is our source of truth for each metric that matters?
• Are those metrics defined the same way across every property, fund, and region?
• When someone asks why something happened, can we answer it today without three people in a spreadsheet?

If the answers are shaky, no protocol is going to save you. Fix that first, and MCP becomes genuinely useful. Skip it, and you are going to get confident, fluent, and wrong.

We will write more specifically soon about what we are doing with MCP at symmetRE, and how we are thinking about security, redaction, and the workflows that make AI useful in a real operational context — not just in a demo.
Explore more at https://symmetre.com/ai_ready_layer.