Anything it cannot source: a rate, an availability date, a balance. And anything touching a protected class.
Published April 20, 2026
The list of things an agent must never say is shorter than people expect and more important than everything it is good at. It is the part worth agreeing before launch rather than discovering afterwards.
This is the category with real legal weight. Questions about the neighbourhood, the demographics of a building, whether an area is good for families, or schools used as a proxy for the same thing, all sit close to fair housing lines.
The right behaviour is not a careful answer. It is a refusal and a handoff. The agent stays on units, rates, availability and policy, and anything touching a protected class goes to a person with no exception written into the script.
If a fact is not in your PMS, your rate sheet or your policies, it does not get said. Not approximated, not hedged, not offered as a rough idea.
The correct response is to say it will get the exact figure and bring in a person. A confident wrong answer costs more than a missed call, because it creates an expectation you then have to take back.
An agent should not negotiate, waive a fee, promise an exception or agree to hold a unit outside your normal rules. Those are decisions with money attached and they belong to somebody who can own them.
This is also the boundary that protects the agent's usefulness. Once it can make exceptions, every conversation becomes a negotiation and you have lost the consistency you were buying.
If somebody asks directly whether they are speaking to a person, it tells them. Claiming otherwise is a bad trade: it buys a slightly smoother conversation now in exchange for a much worse one when the truth becomes obvious.
Every item here is a script decision made during the build, not a judgement the agent makes in the moment. That is the point. A boundary that depends on good judgement under pressure is not a boundary.
The interesting part of an AI agent is not what it says. It is what it refuses to say.
A boundary is only useful if the way it is enforced does not itself damage the conversation, and there is a real difference between a good refusal and a bad one.
A bad refusal is a flat statement that it cannot help, which leaves the person with nothing. A good one names what happens next: that somebody will come back with the exact figure, roughly when, and on which channel. The person still does not have their answer, and they do have a reason to wait rather than to call the next listing.
These rules should be verified rather than assumed, and the test is straightforward: try to get the agent to break them.
Half an hour of that is worth more than any assurance a vendor can give you, including ours. Do it before launch, and do it again after any significant change to the configuration.
On the site
Keep reading