Short answer: under the Fair Housing Act, the decisions that matter — who gets housing, on what terms, and how you treat people who ask for accommodations — are yours and your staff's. Software can gather facts and prepare drafts. It must never screen, score, decide, or steer. Design it that way and you get the efficiency without the exposure.
Most property managers have asked themselves some version of this in the last year: can I let AI answer leasing calls, review applications, handle resident messages — without creating a Fair Housing problem? The question is the right one, because the risk is real: a well-meaning automation can produce different outcomes for different protected classes without anyone intending it, and "the software did it" is not a defense. 42 U.S.C. § 3604 and its state cousins apply to what happens, not to who — or what — did it.
The bright lines
Never decided by software:
- Screening and application outcomes. Approve, deny, conditions — human decisions, on human-reviewed criteria, applied consistently.
- Adverse action. Any notice that changes someone's housing prospects gets a person behind it, with the true, specific reasons documented.
- Reasonable accommodations and modifications. Requests from residents with disabilities deserve a human conversation, full stop. An AI that negotiates, delays, or deflects these is a liability generator.
- Steering. No system should describe neighborhoods, buildings, or units in terms that guide people toward or away from anything based on protected characteristics — including accidentally, through learned patterns in your own historical data.
- Terms, pricing, and promotions. Same terms, same information, same courtesy — for everyone.
Fine for software, with human review:
- Gathering facts: unit availability, work-order status, office hours, appointment scheduling by neutral rules.
- Drafting communications that a human approves before sending.
- Assembling history — what happened, when, with whom — so staff answer from the record, not from memory.
Why "the AI is neutral" is not a defense
Consistency is not the same as lawfulness. A system trained on past behavior learns past patterns, including the ones you'd never defend in a deposition. And automated systems make different mistakes than people do — subtler, harder to spot, applied at scale. This is why the design principle at Normal is blunt: the assistant prepares; a human decides everything that affects a person's housing. The approval step isn't bureaucracy. It's the control that makes the rest of it safe, and it leaves the paper trail that shows you took the duty seriously.
The questions to ask any vendor
- Can the system approve, deny, score, or prioritize applicants — even "just as a suggestion"? (You want no.)
- Do humans approve every communication to residents and prospects before it sends?
- How are accommodation requests routed — to a person, immediately, every time?
- What records does it keep of who decided what, and when?
- Where does applicant and resident data live, and can it stay on your own hardware if that's your policy?
A realistic first step
Audit anything automated against one question: "did a human make this decision?" Where the answer is no, add the human. If you want help drawing those lines for your operation — or a deployment built human-first from the ground up — talk with us. We'd rather lose a sale than design around a boundary.
Related: AI maintenance triage your residents won't hate · our approach for property management.