How to Build a Fair-Housing-Safe AI Lead Qualifier in n8n

Most AI lead qualifiers I audit are one screenshot away from a Fair Housing complaint. Not because the agent using it is biased. Because nobody told the model what it isn't allowed to look at.

Here's the typical setup: a Zillow lead gets piped into GPT-4o with a prompt like "score this lead 1–10." The model sees a name, a ZIP code, and a message that mentions "good schools for my kids." It produces a score. Nobody ever sees the reasoning behind it. Now you have an automated system making routing decisions on proxies for familial status, national origin, and race — and no record of how it got there.

That's not a hypothetical risk. It's the default state of most AI lead-scoring setups running in real estate right now.

Here's how I build it instead. It takes about 90 minutes in n8n.

Step 1: Catch the Lead

A webhook receives the inbound lead from Zillow, Realtor.com, or your website's contact form — same as any standard intake flow.

Step 2: Redact Before the Model Ever Sees It

This is the step almost everyone skips, and it's the one that matters most. A Set node strips out the fields before the AI touches them: name, photo URL, raw ZIP code, and any free-text mention of family, religion, or national origin.

The model can't be biased by data it never receives. Redaction isn't a policy you write down and hope people follow — it's a node in the workflow that makes the biased path structurally impossible.

Step 3: Score on Fundamentals Only

What's left after redaction gets scored by Claude on four things, and only four things: budget stated, financing status, timeline, and property criteria. Classic BANT. The prompt forces structured JSON output with a required reasoning field on every single score, so there's always a written explanation for why a lead landed where it did.

Step 4: Route Automatically

A Switch node routes based on the score. Hot leads go straight to Follow Up Boss with a call task attached. Warm leads enter a 10-touch sequence. Cold leads get added to a monthly market update list.

Step 5: Log Everything

Every decision — the inputs the model saw, the score it produced, and its stated reasoning — appends to an Airtable audit log automatically. If anyone ever asks why a particular lead was routed the way it was, you have the receipt, not a shrug.

The Cost of Doing It Right

The redaction step and the audit log add maybe 15 minutes to the total build time. They're also the only two nodes in the entire system that actually matter if you're ever asked to explain how it makes decisions.

68% of agents are running some form of AI in their lead flow right now. Almost none of them could show you what their model actually looked at when it made a call. That's the gap this system closes.

Build It Before You Need It

The best time to add redaction and an audit trail is before a complaint, not after one. If you're running leads through any kind of AI scoring right now — homegrown or otherwise — it's worth an hour to check what fields your model is actually seeing.

We build these fair-housing-safe qualification systems for brokerages and property managers using n8n, Claude, and Airtable, wired directly into your existing CRM.