AI Contact Search

Good data isn't verified data

We had solid data partners. They still got things wrong sometimes, and a wrong owner means a wasted call or a burned relationship. So brokers did what brokers do: opened a pile of tabs. The company's website, public records, news articles. They were validating our data by hand, every time.

The obvious fix was another data source. I didn't think that was the problem. The brokers had already figured out how to verify an owner. We just weren't in that flow.

I watched them work

I spent the research phase watching brokers do the job. Where do they look first? What makes them believe a contact? What makes them stop trusting one?

Brokers are hard to get ahold of. They're on the phone, in the field, or both. We landed about five interviews, and I took every one of them. Five is small. But they described the same loop, almost step for step.

That flow became the prompt. The agent searches the web the way a broker would, then compares what it finds against our vendor data and property records. Matches get ranked by how much evidence backs them up.

The research didn't inform the feature. It was the feature.

Two buttons, on purpose

We shipped In-Depth Owner Lookup next to the existing Quick Lookup, and we left both there for months. Some brokers already knew who the owner was and only needed a confirmation. Others wanted to see if the AI would do better. I didn't want to force the answer. I wanted to watch it.

Both cost one credit, whether or not the search finds anyone. We pay for the lookup either way, so a refund-on-empty model wasn't on the table.

The waiting state is quiet on purpose. The agent finishes fast, so a progress theatre of steps and spinners would have taken longer than the work. The status text changes as the lookup moves along, with a small animation. That's it.

Trust is receipts

A broker is going to pick up the phone based on this screen. "The AI said so" isn't good enough.

So we compare everything the agent finds against vendor data and property records. When a contact matches legal documents or assessor records, it earns a badge (AI Verified, Assessor Match, Legal Entity Officer) and floats into Highest Confidence. The evidence is the ranking.

The AI summary and the contacts show their sources on hover, so a broker can see exactly which links the agent found. The agent didn't search once and grab the first thing it saw. It checked several places, and the broker can see every one.

The popup also says, right at the bottom: "AI can make mistakes. Please review the information provided below thoroughly." I wanted that there. A tool that admits it can be wrong gets trusted more than one that doesn't.

Watching clicks instead of asking

Early on, brokers were hesitant, and they weren't going to tell us why. I'd planned on thumbs up and down to measure quality. Nobody clicked them.

So I stopped asking and started watching. In Pendo I tracked clicks per account. If someone tried the AI lookup and then went back to Quick Lookup, that was the answer: the prompt wasn't good enough yet. We tuned it with about 20 users at a time, a couple of rounds, until the trend flipped.

It flipped hard. Quick Lookup was getting about 70–100 clicks every 30 days. The AI button was getting around 2–3 thousand. In January 2026, we took Quick Lookup out.

What I got wrong

I assumed people would rate the results. They didn't. I should have planned a short survey for after the feature had been used for a while.

That said, the usage data got us to the same conclusion a survey would have. People vote with their clicks.

Next time I'd still build the survey. I just wouldn't wait on it.