Making home search less overwhelming - without cutting the agent out

OneHome is Cotality's real estate portal, and the thing that makes it different from the big public listing sites is that it connects a homebuyer to an actual agent. Their brief - Project Evolve - asked how AI should change search on OneHome without breaking that.

Spring quarter was research. Summer was design and evaluation, roughly ten weeks. We scoped it hard: the homebuyer search and comparison experience only. Sellers, commercial real estate, Cotality's other products, and production engineering were all explicitly out.

Four of us worked on it, and the work was genuinely shared - a teammate drove much of the spring research, and the reframe, the two feature concepts, and the principles we held ourselves to were argued out together.

The two primary features were shared across all four of us, with everyone weighing in as the concepts and designs evolved. My deepest ownership was the AI interaction layer inside both - Ask AI, Summarize, Suggested Questions - which I carried from concept through final design, with the rest of the team shaping it along the way like everything else. I also built alternate concepts for the Preference Summary and Comparison Summary; neither became the final layout, but pieces of that thinking made it into the delivered designs. Separately, I concepted a preference-overlay feature - letting buyers thumbs-up or thumbs-down individual insights instead of restating one blanket preference - that never made it past the sketch stage; it's on the list of what we'd build next if we had more time.

On the research side, I designed the Ecosystem Map and the Friction Map, two of the shared artifacts the team built on. They lay out where homebuyers and agents run into problems across the search process, drawn from the team's research, so Cotality's partners could see the story of that research at a glance.

The instinct I try to bring to a team like this is less about owning one polished screen and more about catching what everyone else is boxed into. Four people converging on the same kind of solution is a real risk on a capstone team, and I look for the question nobody's asked yet - the clearest example is in the AI chat layer below.

We started with a question that already assumed its own answer: how might we strengthen collaboration by incorporating AI search tools into OneHome? Research moved us off it. Buyers weren't asking for a smarter search box - they were overwhelmed by content-heavy listings with no sense of what mattered to them, and they mostly used AI to understand the homebuying process, not to find homes. Agents weren't asking for AI search either; they wanted the repetitive work back - paperwork, emails, scheduling, manually cross-referencing listings before sending options to a client.

So the question became: how might we strengthen homebuyer-agent collaboration by making the search process less overwhelming for buyers and reducing the repetitive workflows that consume agents' time? We handed Cotality three recommendations out of spring. They had already started building the first one - expanded filters plus conversational search - so summer went to the other two.

A listing gets a match score, and under it two expandable groups: what matches your top preferences, and what's missing from them. Color-coded, scannable, collapsed by default. The same information routes back to the buyer's agent, so the agent stays current on priorities that shift mid-search instead of finding out on a tour.

The unmoderated round settled the main question: every single tester correctly explained what the % match feature did on their own, unprompted, usually by pointing at a specific miss like hardwood floors. Asked whether they'd use it for real, all of them said yes. One pattern failed consistently, though - we'd put Suggested Questions in a horizontal scroll, and testers simply didn't find it.

Two listings side by side, one flagged as the best match, and a How They Compare panel that scores both against the buyer's own criteria - split into must-haves and nice-to-haves, with checks and X's per row. The point is that the criteria are the buyer's, not a fixed preset list. Agents do this comparison by hand today before sending a client options, so automating it buys back time they'd rather spend with the client.

In testing, every user found the must-have and nice-to-have tables intuitive and the check marks fast to scan. Agents read the feature differently than buyers did: less as a shortcut, more as a source of talking points and better questions to walk a client through.

Ask AI, Summarize, and Suggested Questions are one shared component embedded in both features rather than a separate chatbot destination - the piece where I had the most direct concept-to-design ownership, with the rest of the team weighing in as it evolved. Testing shaped it twice: users - first-time buyers especially - told us a blank prompt left them with no idea what to ask, and the label we'd used, "Ask Me Anything," left people unsure who they were even talking to. So the final version leads with suggested questions and states plainly that the responses are AI-generated.

One direction I explored in mid-fi concepts never made it into the final design or the report: having every AI answer link back to what actually produced it - the listing details behind a % match, the criteria behind a recommendation - instead of just asserting a confident summary. Buyers and agents were being asked to trust an AI's judgment on one of the biggest purchases of their life, and I didn't think that trust should ride on tone alone; it should be checkable. It didn't survive scope and time - the chatbot got less iteration than the other two features - but it's the same instinct behind the preference overlay: make the AI show its work, not just assert it.

Public listing platforms are capturing homebuyers earlier and earlier with their own AI features, which puts pressure on Cotality to keep pace without giving up the thing that makes OneHome different: the agent relationship at the center of it.

Both quarters of research pointed the same direction. Buyers want AI to save time and lower stress. Agents want it to clear repetitive work off their plate. Neither group wants AI to become the default experience - they want it available on their own terms. That's what these two features give Cotality a way to build: AI that deepens the buyer-agent relationship instead of commoditizing it.

The most useful next move is more testing: validate the final iteration against a larger, more diverse sample than 21 people from our networks and CloudResearch.

The honest gap is the agent's side. Both features send information back to the agent, but our prototype only ever covers the public-facing view. We never designed or tested how a buyer's activity actually surfaces to an agent in their own interface - that's a whole half of the flow we're asserting rather than showing.

Early conversations with Cotality also touched on improving the information flow between buyer and agent directly, which we couldn't test this phase. I sketched one concept for it: after a buyer favorites a listing, a quick overlay asks what specifically mattered about it, then lets them thumbs-up or thumbs-down individual insights - commute, HOA cost, yard, school zone - more of this, less of that, instead of restating one blanket preference. That signal would feed both their own future summaries and their agent's view. It's a sketch, not a finding.

On my own list: an actual working prototype of the Ask AI layer. The course never called for that level of fidelity, but I want to see the sourcing behavior and suggested questions actually run, not just click through static screens - that's the next thing I'd build for this portfolio, separate from the course deliverable.

We came in asking what AI should do for home search. The more useful question turned out to be what it shouldn't. Once we stopped trying to make the AI decide and let it do the narrower job - organize, prioritize, explain - the designs got clearer, and testing backed them.

The final direction for OneHome is a supportive layer, not a replacement: AI that hands a buyer the questions worth asking, and hands their agent the context to answer them.