Guide

Building compliant software in real estate

Real estate software has a compliance problem the other regulated industries do not. In finance and healthcare, the rules mostly govern how you handle data. In housing, the rules govern the decision your software makes, which means the model's output is itself the regulated act. That changes what you have to build and what you have to test.

Published August 22, 2026. Editorial.

Key takeaways

The interagency AVM rule, effective October 1, 2025, requires quality control standards including compliance with applicable nondiscrimination laws for covered automated valuation models.
HUD guidance issued in May 2024 makes clear the Fair Housing Act applies to tenant screening and housing advertising when algorithms and AI perform those functions.
Discriminatory outcomes do not require discriminatory inputs. A model with no protected attribute can still produce a disparate impact through correlated features.
Testing for disparate impact has to be part of the build and the ongoing check suite, not a one-time review, because model behaviour changes as data changes.
If a decision affects someone's housing, they will eventually ask why. Build the explanation into the system rather than reconstructing it later.

A real estate product usually begins as software that helps somebody make a decision, and then, gradually and without any single moment where it changes, becomes software that makes the decision. The screening tool that showed information starts returning a recommendation. The valuation model that gave an engineer a quick check starts providing values to a lending workflow. The ad platform that targeted interested buyers starts deciding who sees a listing.

That gradual change matters more in housing than in most industries, because housing is one of the areas where the law is concerned with outcomes rather than intentions. You can build a screening model with no protected characteristic anywhere in the training data, deploy it in good faith, and still produce a pattern of outcomes that creates real legal exposure and real harm to real people.

The distinguishing feature of this industry

In financial services, most obligations govern how you handle records. In healthcare, most govern how you handle data. In real estate, a large share govern the decision itself.

That is a different engineering problem. Encrypting a database is a solved problem you can verify once and assert forever. Establishing that a model does not produce a disparate impact across protected classes is a statistical question with no permanent answer, because the model's behaviour changes as its inputs change, and the population it acts on changes too.

Which means the compliance work in proptech is not primarily about controls around the system. It is about measurement of the system, continuously, and about being able to explain a specific decision to the specific person it affected.

Two developments made this concrete rather than theoretical.

The AVM rule. The OCC, Federal Reserve, FDIC, NCUA, CFPB, and FHFA adopted a final rule, implementing section 1473(q) of the Dodd-Frank Act, requiring institutions using automated valuation models in certain credit decisions or securitization determinations to adopt policies, practices, procedures, and control systems ensuring the models adhere to quality control standards. Those standards are designed to ensure a high level of confidence in the estimates, protect against manipulation of data, seek to avoid conflicts of interest, require random sample testing and reviews, and comply with applicable nondiscrimination laws [1]. It took effect on October 1, 2025 [1].

The final clause is the notable one. A nondiscrimination quality control factor attached to a valuation model is a requirement the engineering team has to meet directly, because you cannot satisfy it with a policy. You satisfy it with testing.

The HUD guidance. In May 2024, HUD issued guidance on the application of the Fair Housing Act to tenant screening and to the advertising of housing opportunities through online platforms, including when algorithms and artificial intelligence perform those functions [2]. The message is that using a model does not move the obligation somewhere else.

Why AI-native development helps, and where it does not

We build every product with a check suite derived from its requirements, and in this industry that matches what the regulation asks for closely, because the regulation asks for testing rather than for documentation. Random sample testing and reviews is written into the AVM standards [1]. A build process whose normal output is a continuously running measurement suite is producing exactly the artefact the rule describes.

The honest limit is that no amount of engineering rigour decides what fairness metric is appropriate for your product, and that choice is consequential. Different statistical definitions of fairness conflict mathematically: you generally cannot satisfy all of them at once, and choosing among them is a policy and legal decision informed by the specific context, not something an engineering team should settle by picking whichever one the library implements by default.

So the division of labour is: counsel and policy decide what must be measured and what thresholds mean, and the engineering process makes the measurement continuous, visible, and hard to disable without anyone noticing. Our page on testing a model for disparate impact covers the mechanics with that boundary respected.

What this guide covers

Which obligations reach a proptech product at all is on what regulation reaches a real estate product. The AVM standards in engineering terms are on AVM quality control standards in practice. Screening is on tenant screening and fair housing, and ad delivery, which is the least understood of these, is on housing advertising and audience targeting.

The building pages cover measurement and explanation, which are the two capabilities a housing product needs and usually lacks. If you are looking for the engagement side rather than the regulatory detail, our real estate software development page covers what we build and how.

None of this is legal advice, and fair housing determinations are genuinely fact-specific. What it is, is a description of what a system has to be able to do so that the legal questions can be answered at all. A product that cannot report its own outcome distribution or explain an individual decision cannot demonstrate compliance no matter how carefully it was built.

If you are building or reviewing a real estate product, our real estate work is the place to start. If what you actually need is to find out your current position first, get in touch and we will work through what your product actually needs to measure.

Explore the guide

Start here

What regulation reaches a real estate product

Proptech teams usually ask which regulations apply to their category. That is the wrong question, because the answer depends on what your software decides rather than on what kind of product it is. A listings site and a screening tool can be the same codebase with a different feature switched on.

Why removing the protected field is not a defence

The most common thing a proptech team says when asked about fairness is that the model does not use protected characteristics. It is true, it is well intentioned, and it does not answer the question, because a model does not need a protected attribute to reproduce a protected pattern.

The rules that apply to your code

AVM quality control standards in practice

The AVM rule is unusual in that it names testing as a requirement rather than leaving it implicit. For an engineering team, that makes it one of the easier regulations to build for, because most of what it asks for is measurement you can automate.

Tenant screening and fair housing

Screening software is used at the point where a person does or does not get housing. HUD's May 2024 guidance addressed how the Fair Housing Act applies to that function, including where algorithms and AI perform it, and the design implications are concrete.

Housing advertising and audience targeting

This is the obligation engineers find least intuitive, because audience targeting and delivery optimisation feel like technical concerns. They are not. Deciding who sees a housing listing is deciding who learns the opportunity exists, and that is advertising of a housing opportunity.

Building it right

Testing a model for disparate impact

Fairness measurement in housing products usually happens once, in a review, producing a document. The obligation is better served by measurement that runs continuously, like any other check, because a model's behaviour changes as its data changes and a good result in March says little about September.

Explainability and adverse action

Every product that makes consequential decisions eventually faces the same request: tell me why. The teams that handle it well captured the explanation when the decision was made. The teams that handle it badly try to reconstruct it months later from a model that has since been retrained.

Auditing what you have

Compliance gaps in proptech products

Real estate products accumulate a characteristic set of gaps, and they differ from the ones in finance and healthcare. The problem is rarely a missing control around the data. It is that nobody can say what the software decides, who it decided it about, or why.

Common questions

Do fair housing rules apply to an algorithm rather than a person?

HUD issued guidance in May 2024 addressing the application of the Fair Housing Act to tenant screening and housing advertising including when algorithms and artificial intelligence perform those functions. Using a model does not move the obligation to someone else, and the practical consequence is that the software's behaviour becomes the thing that has to be examined.

What does the AVM rule require of an automated valuation model?

Institutions using covered AVMs in certain credit or securitization determinations must adopt policies, practices, procedures, and control systems ensuring the models adhere to quality control standards designed to produce a high level of confidence in estimates, protect against data manipulation, avoid conflicts of interest, require random sample testing and reviews, and comply with applicable nondiscrimination laws. The rule took effect on October 1, 2025.

Can a model discriminate if it never sees race, sex, or familial status?

Yes, and this is the central technical point in housing. Features that correlate with protected characteristics, such as postal code, school district, commuting distance, or name-derived signals, can reproduce a disparate pattern without the protected attribute appearing anywhere, which is why removing the field is not a defence and measurement of outcomes is the only way to know.

Is a one-time fairness audit sufficient?

No, because a model's behaviour changes as its inputs and its population change, so a good result in March says little about September. The obligation is better served by continuous measurement that runs like any other check, with the outcome distribution tracked over time rather than sampled once.

Who decides which fairness metric a housing product should use?

Counsel and policy, informed by the specific context, rather than the engineering team. Different statistical definitions of fairness conflict mathematically and generally cannot all be satisfied at once, so defaulting to whichever metric a library implements is a consequential decision made by accident.

What should a proptech product be able to do to demonstrate compliance?

Report its own outcome distribution across groups, on demand, and explain any individual decision in terms a person can understand and challenge. A product that cannot do those two things cannot demonstrate compliance regardless of how carefully it was built, because both questions will eventually be asked about a specific applicant on a specific day.

Does using a third-party screening or scoring service transfer the risk?

Not in the way teams hope. HUD's guidance discusses screening companies helping to implement rather than effectively set a housing provider's policies, which indicates that the provider keeps responsibility for the criteria applied. Practically, a vendor whose scoring you cannot explain leaves you unable to answer the question an applicant will ask.

Where do proptech compliance gaps usually appear?

In the gradual change from decision support to decision making, which happens without a moment anyone notices. A tool that showed information starts returning a recommendation, the recommendation starts being followed automatically, and nothing in the process re-examines whether the obligations changed when the software's role did.

How is real estate compliance different from financial services or healthcare compliance?

Financial services obligations mostly govern how records are kept, and healthcare obligations mostly govern how data is handled. Real estate is different because a large share of what regulators care about is the decision itself, such as a valuation, a screening outcome, or who sees a housing ad. That makes measuring the decision's output the central engineering task rather than a secondary one.

How long does it take to build fairness measurement into a housing product?

There is no fixed timeline, because it depends on whether outcome data and group information already exist in a usable form. Products that recorded outcomes only for operational purposes typically need a data model change before any measurement can start, so the honest first step is assessing what is already capturable rather than assuming a measurement pipeline can be added quickly.

Is it safe to rely on a third-party AVM or screening vendor for compliance?

Not by itself. HUD's guidance describes screening companies as helping to implement rather than effectively set a housing provider's policies, which means the provider generally keeps responsibility for the criteria applied even when a vendor supplies the score. A vendor score that cannot be explained leaves the provider unable to answer the question an affected applicant will eventually ask.