Writing
What being a real estate agent taught me about designing for AI
Before I was a designer, I sold homes. Before that, I sang for my supper in Colorado ski towns and on boats circling Manhattan. None of this was a plan. But when I found myself the only designer on Zillow's first LLM product, the strangest thing happened: the real estate license turned out to be the most relevant credential in the room.
Here's what nobody tells you about being an agent: a meaningful part of the job is not answering questions. People at an open house will ask you, warmly and without a trace of malice, what kind of people live in the neighborhood. Whether the schools are "good," with a look that tells you exactly what they're really asking. Whether they'd "fit in" here. These feel like reasonable questions to the person asking. They are also, very often, invitations to break federal law.
The Fair Housing Act prohibits steering: influencing where someone buys based on race, religion, disability, familial status, and other protected characteristics. Agents learn this the way you learn anything that can end your career: thoroughly. And what you actually learn isn't just the rules. You learn the craft of the graceful no. How to decline a question without shaming the person who asked it. How to redirect to something genuinely useful: crime data they can look up themselves, school ratings from official sources, a walk around the block at different hours. How to hold a boundary so warmly that the client trusts you more for having held it.
I did that for years. Then I became a designer, and eventually the AI era arrived at my desk.
The model gets asked everything an agent gets asked
When we started building AskZ, Zillow's first LLM chatbot, I sat in rooms where VPs and directors worked through why we needed fair housing guardrails in an AI product at all. For most of the room, this was a new frontier. For me, it was Tuesday. Of course people would ask the model what kind of people live in a neighborhood. People had asked me that, in person, while holding a cup of my open-house coffee.
So when it came time to test the model, I knew exactly what to do: I asked it inappropriate questions. Lots of them. All the coded phrasings, the innocent-sounding versions, the sideways approaches: every way I'd ever heard a steering question dressed up in my agent years. I was probing for the moments where the model would helpfully, cheerfully break the law. I later learned the industry has a name for this: red-teaming. I just called it remembering.
The engineers on my team ultimately built a Fair Housing Classifier to detect these moments, good enough that Zillow open-sourced it for the whole industry. But a classifier only decides that the system won't answer. Somebody still has to decide how it won't answer. And that's where the agent training paid off a second time, because I'd spent years performing exactly that interaction: the decline that doesn't shame, the explanation that isn't a lecture, the redirect that's actually useful. We designed the product's refusals the way a good agent delivers them, and I connected our content strategist with the legal and fair-housing folks so the words were owned by the people who needed to own them.
What the face-to-face years actually taught me
It goes deeper than fair housing. Every job I had before design was, in hindsight, a study of humans under real conditions. On stage, you learn to read a room in seconds and adjust in real time, which is more or less what a conversational product has to do. In hospitality, you learn that people don't remember what you got right; they remember how you handled what went wrong. And in real estate, you learn what a person looks like making the biggest financial decision of their life: hopeful, terrified, pretending to be neither.
I didn't study humans through books and documentaries. I studied them across kitchen counters and cocktail tables, thousands of spontaneous, unrepeatable, occasionally beautiful interactions. That's the dataset I bring to design. And it's why I hold AI products to a standard that sounds simple and isn't: the person on the other end should walk away trusting the system exactly as much as it deserves: no less and, crucially, no more.
The case for weird résumés
There's a version of this essay that's just a pep talk for career changers, and honestly, I'd stand behind it; I made my leap in 2018 and I've been obsessed ever since. But I think there's a sharper point for the AI moment specifically.
Designing AI products means designing for every strange, coded, human thing people will actually say to a system, and the people best equipped to anticipate that are the ones who've heard it all before, in person, with the social stakes live. The field is so new that nobody is traditionally qualified for it. Which means the qualification that matters is the one that never appears in a job description: years spent earning strangers' trust in real time, and knowing precisely what it costs to lose it.
Turns out I'd been training for this job my whole life. I just thought I was singing.