How do you stop AI inventing a rate or an availability date?

This is the right question to ask a vendor, and the answer separates a product built for this from a general model with a real estate prompt attached.

The mechanism is sourcing. Every factual claim comes from a system of record, and where there is no record there is no claim. If a fact is not in your PMS, your rate sheet or your policies, the answer is that it will get the exact figure, followed by a handoff.

Why a general model will make something up

Not through malice or a bug. A language model completes text plausibly, and a plausible completion to the question of what a two bedroom rents for is a number in a reasonable range.

It has no built-in sense that this particular number needs to be true in a way that a paraphrase of your pet policy does not. That distinction has to be imposed from outside, by controlling where facts are allowed to come from.

What that looks like in practice

• Availability, rates and balances are read from your system at the moment they are asked about, not held in a copy.
• Policy answers come from your written policies rather than from general knowledge about how leases usually work.
• Categories of question with no source are refused rather than estimated: what a renovated unit might go for, whether an application is likely to succeed.
• Account details are not given at all until the caller has been verified against the details you chose.

The failure mode nobody mentions

Sourcing solves invention. It does not solve your data being wrong, and that is the more common problem in practice.

If a unit sits marked available for a week after it is gone, the agent will confidently and correctly report what your system says. That is not a hallucination, it is an accurate report of a stale record, and it produces exactly the same wasted tour.

Which is why an hour spent tidying availability before launch is usually the cheapest improvement in the whole project, and why it improves your team's answers at the same time.

How to test it

During the build, ask it things it should not be able to answer. What will the renovated units rent for. Will my application be approved. What is the rent on a unit you have not listed.

The right response to all three is a refusal and a handoff. If you get a confident, reasonable-sounding number instead, you have learned something important before it is talking to your customers.

The harder tests

The obvious questions are the easy ones to pass. The tests that actually tell you something are the ones where an invented answer would sound completely reasonable.

• Ask about a unit number that does not exist, phrased as though it does.
• Ask for a rate on a floor plan you have, but for a month with no pricing set.
• State something false as a premise, such as being told the deposit was waived, and see whether it agrees.
• Ask about a policy you have never written down.
• Ask the same question three times in slightly different words and compare the answers.

That last one is the most revealing test in the list. An answer that comes from a source is the same every time. An answer that is being generated will drift between attempts, and the drift is the tell.

Refusing well

A system that refuses everything it is unsure about is safe and can still be unpleasant to deal with, so the quality of the refusal matters.

The version that works names what happens next: somebody will confirm the exact figure, roughly when, on this channel. And it keeps the conversation moving where it can, since a missing rate does not prevent booking a tour or answering the next three questions. A refusal that ends the conversation has solved the accuracy problem by creating a service one.