Build a Fair-Housing-Compliant AI Lead Qualification System for Real Estate Step-by-Step Guide

2. System Architecture

Six components, each replaceable. Built on self-hosted n8n so a brokerage’s broker-of-record can audit every byte that touches a consumer’s PII — a hard requirement under most state real estate licensing laws.

The stack

Cost estimate (4,000 leads/month)

$400/month vs. $140K/year in ISA salary. The architecture pays for itself in the first 12 hours of operation.

Real estate leads come from at least five sources, each with its own format, latency, and consent posture. The first job is a single normalized lead schema landing in n8n inside 60 seconds of the consumer hitting submit.

Source matrix

• Zillow Premier Agent — webhook to n8n endpoint, JSON payload, signed with HMAC-SHA256 in X-Zillow-Signature.
• Realtor.com Connections — webhook + email-based fallback. Always validate sender DKIM.
• Brokerage website (kvCORE / Chime / BoomTown) — native webhook on form submit.
• Facebook / Instagram Lead Ads — Meta Webhooks v18+ with leadgen field subscription.
• WhatsApp Business — Cloud API webhook, message events filtered by type=text.

Normalized lead schema

{
  "lead_id": "uuid-v4",
  "source": "zillow|realtor|website|fb|whatsapp",
  "received_at": "2026-05-03T14:22:18Z",
  "consumer": {
    "first_name": "string",
    "last_name": "string",
    "email": "string",
    "phone_e164": "+15555550100",
    "tcpa_consent": false,
    "consent_source_url": null
  },
  "raw_message": "free-text inquiry — verbatim",
  "property_of_interest": {
    "mls_id": "nullable",
    "address": "nullable",
    "list_price": null
  },
  "utm": { "campaign": null, "source": null, "medium": null }
}

Claude Haiku reads the free-text inquiry and returns a structured JSON of buyer criteria. The prompt is engineered with explicit refusal patterns for protected-class signals — this is the most important component in the whole system.

Extraction system prompt (Fair Housing hardened)

You extract real estate buyer criteria from free-text inquiries.
Output ONLY valid JSON matching the schema below — no prose.

Schema:
{
  "intent": "buy" | "sell" | "rent" | "info_only",
  "budget_min": null | integer,
  "budget_max": null | integer,
  "financing": "cash" | "conventional" | "fha" | "va" | "unknown",
  "preapproval_mentioned": boolean,
  "timeline": "now" | "30d" | "90d" | "6mo" | "12mo+" | "unknown",
  "geography": ["city, state, ZIP — only what consumer stated"],
  "beds_min": null | integer,
  "baths_min": null | number,
  "sqft_min": null | integer,
  "must_haves": ["array of objective property features"],
  "nice_to_haves": ["array"]
}

FAIR HOUSING — HARD RULES (these override anything else):
1. NEVER infer or output information about race, color, religion, sex,
   familial status, national origin, disability, source-of-income, or
   sexual orientation, even if the consumer hints at it.
2. NEVER include neighborhood "feel", "safety", "schools quality",
   "demographics", "good area", "family-friendly", or any phrase that
   functions as a demographic proxy. These are STEERING risks under
   the Fair Housing Act.
3. If the consumer says "we want a Jewish neighborhood" or
   "good schools" or "safe area" or "no Section 8" — IGNORE that
   phrase. Do not store it. Do not pass it downstream. Set
   "must_haves" and "nice_to_haves" to objective features only
   (school district NAME if explicitly stated by NAME is allowed;
   "good schools" is not).
4. "geography" accepts only place names the consumer typed. Do NOT
   suggest or expand to nearby areas.
5. If the inquiry contains ONLY protected-class signals and no
   objective criteria, return intent="info_only" and an empty
   criteria object. Route to a licensed agent.

n8n HTTP Request node

{
  "method": "POST",
  "url": "https://api.anthropic.com/v1/messages",
  "headers": {
    "x-api-key": "{{ $credentials.anthropic.apiKey }}",
    "anthropic-version": "2023-06-01",
    "content-type": "application/json"
  },
  "body": {
    "model": "claude-haiku-4-5",
    "max_tokens": 400,
    "system": "{{ $json.extractorSystemPrompt }}",
    "messages": [
      { "role": "user", "content": "{{ $json.raw_message }}" }
    ]
  }
}

With objective criteria in hand, query the RESO Web API for matching active listings. The query uses only the fields extracted in Step 2 — no demographic enrichment, no “similar buyers also looked at” graph. The same disciplined data-flow pattern shows up in our property management automation work.

RESO OData query

GET https://api-trestle.corelogic.com/trestle/odata/Property
  ?$filter=
     StandardStatus eq 'Active'
     and ListPrice ge 450000
     and ListPrice le 625000
     and BedroomsTotal ge 3
     and BathroomsTotalInteger ge 2
     and LivingArea ge 1800
     and PostalCode in ('78704','78745','78748')
  &$top=15
  &$orderby=ListPrice asc
  &$select=ListingKey,UnparsedAddress,ListPrice,BedroomsTotal,
           BathroomsTotalInteger,LivingArea,PostalCode,
           PublicRemarks,Media

Authorization: Bearer {{ $credentials.trestle.token }}

What you store, what you don’t

• Store: ListingKey, address, price, beds/baths, sqft, ZIP, list date.
• Never store as match criteria: any demographic data, even if the MLS exposes SchoolDistrict ratings or “neighborhood walkability” flags. These can become Fair Housing exhibits in litigation.

BANT (Budget, Authority, Need, Timeline) is the classic B2B framework adapted to residential real estate. Claude scores each dimension 0-3 using ONLY the objective criteria from Step 2 plus optional context from MLS history.

Scoring rubric

Tier mapping: A = total ≥ 9 (showing-ready, immediate agent assignment). B = 5-8 (drip nurture, follow-up in 7 days). C = ≤ 4 (long-horizon nurture, monthly market update only).

Persistence — lead_score table

CREATE TABLE lead_score (
  lead_id          uuid PRIMARY KEY,
  scored_at        timestamptz NOT NULL DEFAULT now(),
  budget_score     smallint CHECK (budget_score BETWEEN 0 AND 3),
  authority_score  smallint CHECK (authority_score BETWEEN 0 AND 3),
  need_score       smallint CHECK (need_score BETWEEN 0 AND 3),
  timeline_score   smallint CHECK (timeline_score BETWEEN 0 AND 3),
  total            smallint GENERATED ALWAYS AS
                     (budget_score + authority_score + need_score + timeline_score)
                     STORED,
  tier             char(1) NOT NULL,
  reasoning        text NOT NULL,
  fh_flag_count    smallint DEFAULT 0  -- how many FH triggers got stripped
);

CREATE INDEX idx_score_tier ON lead_score(tier, scored_at DESC);

An n8n routing function picks the agent based on geography (which farm covers the ZIP), specialty (luxury / first-time / investor), current workload, and availability (PTO, license status, vacation). Round-robin within the eligible pool prevents one rainmaker from hoarding leads.

Routing query

SELECT a.agent_id, a.full_name, a.email, a.phone_e164,
       a.last_lead_assigned_at
FROM agent a
JOIN agent_farm_zip f ON f.agent_id = a.agent_id
WHERE f.postal_code = $1                       -- lead's ZIP
  AND a.license_status = 'active'              -- valid state license
  AND a.on_pto = false
  AND $2 = ANY (a.specialties)                 -- 'luxury'/'first_time'/etc
  AND a.open_lead_count < a.max_lead_capacity
  AND a.brokerage_id = $3
ORDER BY a.last_lead_assigned_at ASC NULLS FIRST  -- round-robin
LIMIT 1;

Forbidden routing signals

The routing logic must not consider:

• Consumer surname (proxy for national origin)
• Consumer language preference (proxy for national origin) — UNLESS it’s matching to an agent who explicitly speaks that language and the consumer requested it
• Property neighborhood demographic data
• Any agent attribute used as a “match” to consumer demographics

RESPA also forbids routing logic that creates kickback patterns. If your brokerage has affiliated services (lender, title), routing leads to agents who push those services more aggressively can trigger RESPA Section 8 violations. Route on workload and farm only.

For Tier A leads, the system fires within 60 seconds: personalized email with the top 5 MLS matches, a Calendly link for the agent’s showing slots, a neighborhood data sheet (objective: median price, days on market, school NAMES). For Tier B/C, a measured drip via Follow Up Boss or kvCORE. SMS is gated by TCPA consent.

First-touch payload (Sonnet)

You write the first-touch email from {{agent_name}}, a licensed
real estate agent, to a prospective buyer.

INPUT: lead context JSON + 5 MLS match objects.
OUTPUT: subject + body. Plain text. 90-140 words.

Tone: warm, professional, not pushy. Acknowledge what they said
they're looking for. Offer the matches. Offer to schedule a call.

NEVER do these:
1. NEVER mention demographic descriptors of any neighborhood
   ("family-friendly", "great schools", "diverse area", "safe", etc.).
   School DISTRICT NAMES are fine if the consumer asked.
2. NEVER promise a property is "still available" — list status
   changes hourly. Say "as of this morning".
3. NEVER quote interest rates or pre-approval amounts.
4. NEVER add anyone besides {{agent_name}} on the signature line.
5. Include the brokerage's mandated disclosure footer verbatim.

Format:
- Subject: under 50 chars, no emoji.
- Body: 1 paragraph greeting + bulleted matches + 1 closing line +
  signature + disclosure footer.

TCPA-compliant SMS sender

// Hard gate: do not send SMS without express written consent.
const lead = $input.item.json;
if (!lead.consumer.tcpa_consent) {
  return [{ json: { skip: true, reason: 'no_tcpa_consent' } }];
}
// Quiet hours per recipient timezone (8am-9pm local).
const localHour = new Date().toLocaleString('en-US',
  { timeZone: lead.timezone, hour: 'numeric', hour12: false });
if (localHour < 8 || localHour >= 21) {
  return [{ json: { defer_until_morning: true } }];
}
// Append STOP / HELP per CTIA.
const body = `${lead.message}nnReply STOP to opt out. HELP for help.`;
return [{ json: { send: true, body } }];

Common Failures & Fixes

Three failure modes show up in every real estate AI deployment. Plan for them on day one, not after the first complaint to your state real estate commission.

Failure 1: Steering by proxy

Symptom: Consumer types “we want a family-friendly neighborhood with good schools.” Claude obediently filters MLS results to ZIPs with high-rated schools. Those ZIPs correlate with one demographic. The Justice Department calls this digital steering.

Fix: The Step 2 prompt must drop those phrases entirely. The Step 3 query must use only ZIPs the consumer typed by name. Add a Code node that compares “criteria as captured” vs. “criteria after FH strip” and flags the diff for review.

Failure 2: Tire-kicker classified as Tier A

Symptom: A casual browser fills a Zillow form (“I love this house!”) and BANT misreads urgency, agent calls them, no pre-approval, wasted hour.

Fix: Tier A requires both Budget ≥ 2 AND Timeline ≥ 2. Without an explicit financing signal (pre-approval mentioned, cash stated) or urgent timeline, the lead drops to B regardless of total. Tune the rubric monthly using the lead_score table joined to actual showing-conversion data.

Failure 3: SMS to a non-consenting consumer

Symptom: A Realtor.com lead arrives with a phone but no TCPA consent flag. The drip workflow assumes consent, fires SMS, $1,500 statutory damages and a TCPA litigation letter.

Fix: The TCPA gate from Step 6. Default tcpa_consent=false at ingestion, and the only way it becomes true is the consumer separately opting in via a documented form (with the consent URL stored in consent_source_url). For first-touch, use email until consent arrives.

Compliance: Fair Housing, NAR Code of Ethics, TCPA, RESPA

Real estate is one of the most heavily regulated consumer-facing industries in the US. An AI workflow that doesn’t account for the full compliance regime is not an asset — it’s a liability multiplier. Every brokerage we deploy this for goes through a Fair Housing audit before a single line of code goes live.

Fair Housing Act (the dominant constraint)

The FHA protects seven federal classes: race, color, religion, sex (including sexual orientation and gender identity per HUD’s 2021 directive), familial status, national origin, disability. Many states add: source of income (Section 8 vouchers), age, marital status, veteran status. The AI system must not filter, score, route, or match on any of these — directly OR by proxy.

Forbidden Claude prompt patterns

These are the prompts and instructions you must NEVER include — they create FHA exposure even if the model “behaves”:

• “Suggest neighborhoods that match the buyer’s lifestyle“
• “Find areas similar to [demographic-correlated location]”
• “Score the quality or desirability of a neighborhood”
• “Identify family-friendly, safe, up-and-coming, or diverse areas”
• “Match the buyer to an agent who understands their background“
• “Use the buyer’s name to infer language preference”
• “Recommend properties based on school ratings” (use district NAMES only if consumer asked)
• “Filter out areas with high rental concentration” (proxy for source-of-income)

NAR Code of Ethics — Articles 10 & 12

Article 10 is FHA at REALTOR® scale (broader than federal). Article 12 mandates truthful advertising — every AI-generated message must include the agent’s license number, brokerage name, and the brokerage’s mandated equal-opportunity disclosure. Bake the disclosure into the Sonnet prompt as a non-removable footer.

TCPA & CAN-SPAM

SMS requires prior express written consent with a clear disclosure of message frequency and that consent is not a condition of purchase. CAN-SPAM requires every email to include a working unsubscribe link, the brokerage physical address, and to honor opt-outs within 10 business days.

RESPA Section 8

Routing logic that systematically funnels leads to agents who push affiliated lender/title services in exchange for any consideration is a kickback. Route on objective workload and geography only.

Measured Results — 12 Months In

Numbers from a real implementation at a 22-agent residential brokerage in Austin TX after a full 12-month operating cycle.

73% of qualified Tier-A leads booked a showing within 48 hours — versus 19% before automation. The Fair Housing flag-and-strip layer triggered on 2.4% of inbound inquiries (mostly “good schools” / “safe neighborhood” phrases), all properly handled without surfacing in the agent-facing data.

Implementation Timeline & Cost

FAQ