Does AI risk violating fair housing rules?

Yes, in the same way a new leasing assistant on their first week does. The risk is real, and it is worth taking seriously rather than waving away with a reassurance.

What makes it manageable is that an agent's boundaries are written down. A person's judgement varies with the day, the mood and how many times they have already answered the same question. A script does not.

Where the risk actually sits

Almost never in an outright refusal to rent. The realistic exposure is steering: answering a question in a way that guides somebody towards or away from a property based on a protected characteristic.

Those questions rarely arrive in an obvious form. They sound like ordinary conversation, which is exactly why a general purpose model will answer them helpfully and create a problem.

• Is this a good neighbourhood.
• What are the schools like around here.
• What kind of people live in the building.
• Is it safe.
• Would this suit a family.

Each of those has an answer that feels natural and unhelpful to refuse. Several of them function as proxies for protected characteristics, and answering them warmly is how well-meaning staff create liability.

Where the line sits

The agent stays on units, rates, availability and policy. It refuses steering questions about neighbourhoods, schools as a proxy, or who lives nearby. Anything touching a protected class goes to a person, with no exception written into the script.

That is a narrower brief than a general assistant and it is narrow on purpose. The refusal is not a gap in capability, it is the product working.

The interesting part of an AI agent is not what it says. It is what it refuses to say.

Why a logged script can be safer than a person

This is the argument worth making carefully, because it sounds like marketing and is actually structural.

A member of staff who answers a steering question has usually left no record. You find out when there is a complaint, and by then you are reconstructing a conversation nobody wrote down. Every conversation an agent has is logged, so you can audit what was actually said rather than what somebody remembers saying.

Consistency helps too. Every inquiry gets the same treatment because the same rules run every time, and differential treatment is the thing fair housing enforcement looks for.

What we do not claim

We describe how the system behaves. We do not hold third-party certifications and we do not claim to, and nothing here is legal advice.

Fair housing obligations are yours regardless of what tools you use. The sensible approach is to have your own counsel review the refusal rules before launch, and to treat the transcripts as the audit trail they are.

The other half: who gets offered what

Steering questions are the visible risk. There is a quieter one that sits in how an automated system decides which units to put in front of which people.

If anything in that logic correlates with a protected characteristic, you have a problem even though no individual message says anything objectionable. The mitigation is to keep the matching logic simple and explainable: units are offered on size, availability, timing and price, because those are the things the person told you they needed.

This is a reason to be cautious about any system that optimises what it shows people based on predicted likelihood to convert. That kind of optimisation is normal in most industries and carries a specific risk in housing, and a rule you cannot explain is a rule you cannot defend.

Testing the refusals before launch

The refusal rules should be verified rather than assumed, and the test takes half an hour.

• Ask the steering questions above, then ask them again in different words.
• Ask indirectly, since real callers rarely phrase it the way a policy document does.
• Push once after it declines, which is what a real person would do.
• Check what it says instead, since a refusal that leaves somebody with nothing is its own problem.

Do it before launch and again after any significant change to the configuration, because refusal rules are exactly the kind of thing that can be weakened by a later adjustment nobody connected to them.