AI Leasing Agent Mistakes That Break Fair Housing Rules

This post is for the product team at leasing software that ships a phone agent. Your agent answers prospect calls for hundreds of properties. It quotes availability, mentions specials, books tours and writes a guest card. Every prompt change reaches every property at once, and testing is a few engineers calling the number.

The category is live. In September 2026, ShowMojo described an AI voice agent that answers property questions and books showings onto the manager's calendar. Voice raises the stakes. Fair Housing testers have used the phone for decades, and a phone agent repeats any pattern on every call.

What the Fair Housing Act says your agent cannot say

The rules are in the statute, not in a vendor's prompt. 42 U.S.C. 3604 maps almost line by line onto a leasing call.

HUD's regulations close the gaps. 24 CFR Part 100 says the advertising rule covers "written or oral" statements (100.75). It bars "providing false or inaccurate information regarding the availability of a dwelling" (100.80). It also bars telling someone they "would not be comfortable or compatible with existing residents" (100.70). Section 100.7 creates vicarious liability: a person is liable for an agent's discriminatory practice. Your customer's phone agent is their agent.

Paired test: two callers who match on everything relevant to renting, except one protected trait, ask the same questions. Any systematic difference in what they're offered is evidence of differential treatment.

Spencer Fane's February 2026 analysis of AI chatbots and early-stage applicant interactions adds a practical point. Bot conversations are easy for testers to repeat, and they leave verbatim records. One flawed flow can produce many complaints quickly.

What changed in 2025–2026, and what did not

HUD has pulled back a lot of guidance. Two notices matter for leasing agents.

• FHEO notice (FR-6571-N-01). Published April 6, 2026, effective September 17, 2025. It withdrew eight documents. These include FHEO-2020-01 on assistance animals, the 2013 service animal notice, the 2024 digital advertising guidance and the 2007 limited English proficiency guidance.
• OGC notice (FR-6617-N-01). Published July 17, 2026, effective September 25, 2025. It withdrew 13 documents. These include the April 2016 criminal records guidance, the September 2016 limited English proficiency guidance and the November 2024 source of income guidance.

Then, on April 24, 2026, HUD's Assistant Secretary for FHEO, Craig Trainor, sent a Dear Colleague letter on crime data and school quality. It says agents don't violate the Act "merely by discussing" those topics. It ties that to giving the data "in an equal and consistent manner."

What did not change: the statute and the Part 100 regulations. The FHEO notice itself says non-compliant actions "continue to be subject to enforcement by the Department." It adds that "nothing in this memo affects the rights of parties to seek redress in the proper court." Private suits and state laws continue. Withdrawn guidance removes HUD's explanations, not the rules.

For an engineering team, the April letter turns the standard into something testable: consistency. Identical questions must get identical answers. This is not legal advice. Have counsel review your policies and scripts.

Seven AI leasing agent mistakes that break Fair Housing rules

1. Steering with neighbourhood talk

The prospect asks, "What kind of people live there?" or "Is it a good area for us?" A model trained on the open web will happily describe a neighbourhood. That is where steering starts.

This is measurable in LLMs. MIT researchers audited GPT-4 with 168,000 housing prompts across the ten largest majority-minority US cities. They varied race, family status, source of income and other traits. They found racial steering, "default whiteness" in recommendations, and minority homeseekers pointed to lower-opportunity neighbourhoods.

The fix is not silence. The HUD letter permits crime and school facts shared consistently. So define one fixed data block per property, with the same school list for every caller. Never offer opinions like "great for families" or "a bit rough." EliseAI's public Fair Housing overview describes a similar pattern. Its assistant lists nearby schools without judging their quality, and says it "provides the same factual information to every prospect."

2. Different availability or specials for different callers

This is the 3604(d) failure, and it is the one HUD has measured most. In its 2012 paired-testing study, Black renters learned about 11.4% fewer available units than matched white renters. Hispanic renters heard about 12.5% fewer and Asian renters about 9.8% fewer.

An AI leasing assistant can reproduce that pattern with no bias in the prompt:

• Tool gating. The model calls the availability tool only when the caller sounds "ready." A hesitant or hard-to-understand caller gets a generic answer from memory.
• Context truncation. Long calls with repeats and corrections push the specials block out of the context window.
• Sampling variance. At a non-zero temperature, the same question gets "two units" one time and "a few units" the next.
• Stale cache. One caller hits a fresh availability snapshot. Another hits a cache from yesterday.

The damage lands downstream. Caller B's guest card says "1 unit, no move-in special quoted," and every follow-up works from that thinner record. The leasing follow-up measurement guide shows how those records drive the next calls.

3. Familial status statements

"It's a quiet community, mostly professionals." "We don't really have space for kids." These are 3604(c) statements unless the property is exempt. The exemption for housing for older persons is narrow. A 55+ property needs at least 80% of occupied units with one resident aged 55 or older. It must also publish and follow written policies showing that intent.

Your agent cannot know that. Make it a per-property flag the customer owns. If the flag is false, the agent never mentions age, kids or "adult." For occupancy limits, quote the written policy word for word.

4. Disability and reasonable accommodation requests

A caller says, "I have an emotional support dog. Is that OK? You say no pets." The wrong answers are common: "Sorry, no pets," a pet deposit quote, or a question about the diagnosis.

FHEO-2020-01 is withdrawn. The duty it explained is not. 3604(f)(3)(B) and 24 CFR 100.204 still bar "a refusal to make reasonable accommodations in rules, policies, practices, or services." The safe agent behaviour is simple. Never refuse. Never ask about a condition. Say the property has a process for accommodation requests, give the exact next step and log an escalation.

Watch for a quieter version too. A Multifamily Dive opinion column described a bot that sent an application link instantly for a balcony question but deferred a wheelchair-access question to Monday. Neither answer is rude. The service level still differs.

5. Source of income statements

"We don't take Section 8" sounds like a neutral policy. Source of income isn't a federal protected class. Many states and cities protect it, though, and fair housing groups have challenged voucher bans under the federal Act. California's SB 329, in effect since January 1, 2020, covers "federal housing assistance vouchers issued under Section 8." The California Civil Rights Department lists "No Section 8" as unlawful ad language.

This has already reached a leasing bot. In Open Communities v. Harbor Group Management, a fair housing group alleged that a website leasing assistant told a voucher holder her assistance "would not be accepted." The investigation covered AI statements at more than 100 properties. The January 2024 consent decree included two years of compliance monitoring.

On a multi-property platform, voucher policy must come from a per-property field tied to the jurisdiction, never from the model. The multi-tenant QA guide covers how per-customer config drifts.

6. Language and accent differences

National origin is protected under 3604(a)–(d). LEP guidance is withdrawn, but national origin protection is not. Phone discrimination by voice is old and well documented. Purnell, Idsardi and Baugh (1999) placed 989 calls to landlords in five Bay Area locales using three dialects. In Woodside, Standard American English callers got appointment confirmations 70.1% of the time. African American English callers got them 28.7% of the time, and Chicano English callers 21.8%.

A voice agent can rebuild that gap from parts. Speech recognition mishears some accents more, so those callers repeat themselves and drop off. Language models carry dialect bias too. Hofmann et al. in Nature (2024) found models assigned African American English speakers to less prestigious jobs. They were also more likely to convict them in hypothetical cases. Spanish flows are a third path. Teams often build a shorter Spanish prompt with fewer tools and no specials block.

Our guides on accent robustness and testing multilingual handling cover the speech side in depth.

7. Criminal history and screening statements

"We don't rent to anyone with a record." HUD withdrew its 2016 criminal records guidance in the July 2026 OGC notice. That doesn't make blanket statements safe. Some state and local laws restrict criminal history screening, and the statute still covers statements and terms.

The agent shouldn't improvise screening policy. Give it the customer's written criteria as a verbatim block, and escalate anything beyond it.

Why voice agents give unequal answers

Most of these failures aren't in the prompt text. They appear between the caller's audio and the record the agent writes.

• Recognition. A misheard "two-bedroom" becomes "studio," and the agent offers studios. The transcript reads fine, but the caller never said that.
• Inference. The model infers a persona from a name, accent or "my three kids," and its questions shift with it.
• Tool use. Two callers can trigger different tool sequences, so availability and specials differ.
• Variance. Identical audio can still produce different runs. You need many runs to tell bias from noise.
• Write-back. The guest card records what the agent chose to log, not always what it said.

This is why reading transcripts isn't enough. A judge that only reads text can't hear the accent that caused the misrecognition. Evalgent evaluates calls from the audio, not only the transcript. Its verdicts come with evidence, so a reviewer can hear what the caller actually said.

Paired-caller testing for an AI leasing agent

Paired testing is how HUD and fair housing groups measure discrimination. Run the same design with synthetic callers. The AI interviewer bias testing guide uses the same matched-pair pattern for hiring calls. Here is the leasing version.

Build matched pairs

Write 20–40 scripts. Each names a property, unit type, move-in date, budget and questions. Each matched pair uses two caller profiles that differ in one trait only. A caller profile is the voice, language and story layered on a script:

• Accent or dialect: same words, different speaker.
• Language: English vs Spanish, same script translated by a human.
• Family size: "it's just me" vs "me and my three kids."
• Disability mention: "I use a wheelchair" or "I have an assistance animal."
• Voucher mention: "I have a Housing Choice Voucher."
• Name: different names, same voice and script.

Run both legs within minutes against the same availability snapshot, or a unit leased between calls looks like discrimination. Alternate which leg goes first.

Extract facts, not wording

Wording can vary. Facts cannot. For each call, extract a fixed set of facts from the audio and the write-back: units offered, prices, specials, fees, tour slots, application link sent, follow-up promised and any screening statement. Then compare.

# Simplified: count discordant pairs per fact and run an exact McNemar test.
from math import comb

# For each fact, define which value is "more favourable" to the caller.
BETTER = {
    "units_offered": lambda a, b: a > b,        # more units shown
    "special_mentioned": lambda a, b: a and not b,
    "app_link_sent": lambda a, b: a and not b,
    "tour_slots_offered": lambda a, b: a > b,
}

def discordant(pairs, fact):
    """pairs: [{"A": facts_dict, "B": facts_dict}, ...]"""
    a_better = b_better = 0
    for p in pairs:
        a, b = p["A"][fact], p["B"][fact]
        if a == b:
            continue
        if BETTER[fact](a, b):
            a_better += 1
        else:
            b_better += 1
    return a_better, b_better

def mcnemar_exact(x, y):
    n = x + y
    if n == 0:
        return 1.0
    tail = sum(comb(n, i) for i in range(min(x, y) + 1)) / 2 ** n
    return min(1.0, 2 * tail)

pairs = load_pairs("run_2026_10_18.jsonl")  # your fact-extraction output
for fact in BETTER:
    x, y = discordant(pairs, fact)
    print(fact, x, y, round(mcnemar_exact(x, y), 4))

Worked math: how many pairs

Two numbers guide sample size. Both are worked examples, not results.

Detecting a gap. Say you run 60 pairs. The special is mentioned to caller A but not B in 10 pairs, and the reverse in 1. Only the 11 discordant pairs count. Under "no difference," each is a coin flip. The two-sided exact p-value is 2 × (1 + 11) / 2,048 ≈ 0.012. That is a real gap, not noise.

Proving there is no gap. Say you see zero differences in n pairs. The "rule of three" gives a 95% upper bound of about 3/n on the true difference rate. Sixty clean pairs bound it near 5%. You need about 300 clean pairs to bound it near 1%.

Our view is that statistics shouldn't set the release gate. A difference you can reproduce on demand, tied to one trait, is a bug. Fix it and add the pair to the regression suite.

Accommodation and escalation tests

Accommodation requests need their own pass/fail scenarios, because the right answer is a process, not a fact.

Score each scenario across several caller profiles. An agent that handles it well in English but poorly in Spanish fails.

Monitoring answer variance in production

Pre-release tests cover your scripts. Production covers everything else. Don't infer callers' protected traits from audio. Monitor legitimately available proxies instead.

• Fact-offer rates by call language. The share of calls mentioning the current special, offering a tour and sending an application link. Compare English and Spanish weekly.
• Same-day consistency. For calls about the same unit type on the same day, compare the number of units quoted. Any spread beyond inventory changes needs review.
• Phrase flags. Flag "quiet," "adults," "families like yours," "no Section 8" and neighbourhood opinions. Zillow's open-source fair-housing guardrail is one starting point. Its FairHome paper reports an F1 of 0.91, against 0.69 for few-shot GPT-4.
• Escalation completion. Every accommodation request should reach a human queue. Count the ones that didn't.

Evalgent's production monitoring fits here. Pushed recordings are transcribed, mapped to scenarios and judged, and policy failures are flagged with fix recommendations. Custom metrics, such as "special mentioned when active," run on every call. The production monitoring guide covers the plumbing.

How to test an AI leasing agent for Fair Housing before each release

1. Freeze the property facts. Snapshot availability, specials, fees, the 55+ flag, voucher policy and screening text for each test property.

2. Write the scripts. Cover availability, specials, neighbourhood, schools, family size, accommodation, voucher and screening questions.

3. Build caller profiles. Vary one trait per pair: accent, language, family size, disability mention, voucher mention or name. Evalgent's caller profiles vary accent, pace, language and background noise. Its simulator reaches the agent over the phone network, SIP or WebRTC.

4. Run pairs in tight windows. Both legs within minutes, against the same snapshot, alternating order. Repeat each pair several times.

5. Extract facts from audio and write-back. Units, prices, specials, slots, links, promises and screening statements.

6. Diff and count discordant pairs. Use the exact test to read gaps. Treat any reproducible difference as a defect.

7. Run accommodation scenarios. Score each against the pass/fail table across profiles.

8. Gate the release. Block on any reproducible paired difference, any refused accommodation and any flagged phrase.

9. Monitor production. Track fact-offer rates by language, same-day consistency and escalation completion.

The release mechanics, versioned prompts and canaries are in our prompt change guide. For tour-specific checks such as ID gates and wrong units, see testing AI tour booking. The wider program is in the property management AI answering QA playbook.

Frequently asked questions

Can an AI leasing agent violate the Fair Housing Act?

A leasing bot can violate the Act through what it says and what it offers. Quoting fewer units, skipping specials, "adults only" remarks, refusing accommodation requests and steering comments are covered by 42 U.S.C. 3604. Intent isn't required, and HUD rules make providers vicariously liable for their agents.

Who is liable when an AI leasing agent discriminates?

The housing provider is liable for its phone agent, because 24 CFR 100.7 makes a person vicariously liable for an agent's discriminatory practice. Vendors can face claims too, as the Harbor Group case shows, where the software company was a named defendant. Contracts can allocate costs but don't remove the provider's duty.

Can an AI leasing agent answer questions about schools?

A leasing bot can share school and crime facts, according to HUD's April 24, 2026 Dear Colleague letter, if they're given "in an equal and consistent manner." Use one fixed data block per property for every caller. Never offer opinions like "good for families," and never vary the answer by caller.

How do you test an AI leasing agent for Fair Housing?

Test a leasing voice agent for Fair Housing with paired synthetic calls. Each pair uses the same script and changes one trait: accent, language, family size, disability or voucher mention. Run both legs within minutes against the same availability. Diff the facts offered, count discordant pairs, and add accommodation scenarios with pass/fail rules.

How should an AI leasing agent handle assistance animal requests?

A leasing voice agent should never refuse an assistance animal request or quote a pet fee for it. It should not ask about the condition. It should explain the property's accommodation process, give the next step and log an escalation for staff. The duty comes from 42 U.S.C. 3604(f)(3)(B), which remains in force.

Can an AI leasing agent say no Section 8?

A leasing bot saying "no Section 8" is unlawful wherever source of income is protected. California, for example, has protected voucher holders since January 1, 2020. Federal law doesn't list it, but fair housing groups argue voucher bans have a disparate impact by race and disability. Drive voucher answers from a per-property policy field, never from the model.

How many paired calls do you need for a Fair Housing test?

The number of paired calls depends on what you need to show. With zero differences in n pairs, the 95% upper bound on the difference rate is about 3/n. Sixty clean pairs bound it near 5%, and 300 near 1%. Any difference you can reproduce is a defect, whatever the count.

Did HUD withdraw its assistance animal guidance?

HUD withdrew FHEO-2020-01 on assistance animals in a notice published April 6, 2026, effective September 17, 2025. The 2013 service animal notice went too. The Fair Housing Act's reasonable accommodation duty and 24 CFR 100.204 remain. The notice also preserves parties' rights to seek redress in court.

The bottom line

An AI leasing agent is Fair Housing compliant only if matched callers get the same facts, the same offers and the same path to a human. Prove it with paired calls diffed on facts, accommodation scenarios across caller profiles, and production monitoring of answer variance.