Originally published at parvejshah.com/blog/before-building-real-estate-voice-agent-understanding-calls by Parvej Shah.
I've been spending more time thinking about AI voice agents for real estate.
At first, the product seemed fairly straightforward. A lead calls about a property. The AI answers, asks a few questions, qualifies the lead, checks availability, and hopefully books a showing.
Technically, we can build that. But the more I looked at it, the more I realized I was starting with the wrong question.
I was asking: "How do we build the voice agent?"
Before answering a much more important one: "What is a good real estate agent actually trying to accomplish during this call?"
That distinction changes almost everything.
The technology is becoming the easy part
Building a voice agent today is surprisingly accessible. We can connect a phone number, speech models, an LLM, a CRM, calendars and other tools. We can make the system answer questions, collect information, and trigger actions.
So it becomes tempting to think of the workflow as:
Call → AI → Appointment
Nice and clean. Except the actual conversation between those three boxes contains most of the business logic.
Someone might call about a specific property. Someone else might only know their budget and preferred neighborhood. One caller might be ready to tour this weekend. Another might not have financing sorted out. Another might be selling their current home first. Someone else may simply have a question that requires a human agent.
Those aren't the same conversation.
And the data reinforces how important this early interaction is. Zillow's 2025 research found that 52% of successful buyers said contacting a real estate agent was their first home-buying activity, and 80% had contacted one within their first three activities. A private in-person tour was also a common next step. [1]
So the conversation isn't some minor administrative step sitting on the edge of the buying process. It can be near the beginning of the relationship.
A booked showing isn't always the goal
This was another assumption I had to challenge.
It would be convenient to measure a voice agent like this:
More calls → more booked showings → better agent.
But that's too simplistic. A successful call might end with a showing booked, a qualified lead captured, a question answered, a follow-up scheduled, or a handoff to the right human. The system needs to understand which outcome makes sense.
That's especially important because people don't all want to communicate the same way. Zillow's 2025 survey of prospective U.S. buyers found phone calls were the preferred method for 33%, while SMS was preferred by 34% — and preferences varied considerably by age. [2]
A voice agent therefore shouldn't necessarily try to keep someone on the phone until it achieves its own definition of "conversion." Sometimes the best outcome of the call might be "I'll text you those details." Sometimes it's "let me connect you with the agent." And sometimes it's "I can schedule that showing."
That's a much more interesting system to design.
The ??? is becoming the important part
The mental model I started with looked like this:
Incoming call → AI → Showing
Now I think about it more like:
Incoming call → understand intent → gather relevant context → determine next action → execute or escalate
The AI is only one component inside that flow. And suddenly the questions I'm interested in aren't primarily about models.
What information does an agent need before arranging a showing? Which questions should be asked, and in what order? When should the AI stop asking questions? What should happen when someone doesn't know their financing situation? What happens when the caller asks something the system shouldn't answer confidently? When does a human need to take over?
Those questions require understanding real estate, not just understanding LLMs. That matters because human agents still do much more than provide property information. The National Association of Realtors' 2025 U.S. buyer and seller research found that 88% of buyers purchased through an agent or broker. Buyers particularly sought help finding homes, negotiating terms, and navigating paperwork. [3]
AI doesn't make that complexity disappear. It has to operate inside it.
Maybe "human-like" is the wrong benchmark
A lot of voice AI demos focus on how natural the system sounds. Latency is tiny. It interrupts naturally. It says "um." You can barely tell it's AI.
Those are impressive engineering achievements. But for a business, I think there's another benchmark that matters more:
Did the conversation move the customer toward the correct next step?
An incredibly human-sounding agent that qualifies the wrong person, invents an answer, or pushes every caller toward an appointment isn't a good system. Meanwhile, a slightly less magical voice experience that correctly understands intent, retrieves reliable information, captures context, and knows when to hand over to a human may create much more value.
That changes what "good" means.
Domain knowledge becomes part of the engineering
There's a broader lesson here that goes beyond real estate. As AI makes implementation easier, understanding what should be implemented becomes more valuable.
A developer can learn a voice platform, its APIs, how to connect tools, how to improve prompts. But none of those automatically teach us how a brokerage operates.
That's why, before trying to build the perfect real estate voice agent, I'm becoming more interested in the calls themselves. What do great agents ask? What do buyers ask repeatedly? Where does a conversation change direction? What information actually affects the next action? Where do leads disappear? Where should automation stop?
I'm still researching those questions. And I think that's exactly the point.
The goal isn't to build an AI that sounds like a real estate agent. It's to understand what a good real estate conversation accomplishes, then decide which parts AI should handle.
The technology can answer the phone. The real work starts after hello.
Sources: Zillow Consumer Housing Trends Report 2025 — Buyers, Zillow Consumer Housing Trends Report 2025 — Prospective Buyers, NAR 2025 Profile of Home Buyers and Sellers
Parvej Shah is a Lead Full-Stack Web Developer & Platform Architect based in Dhaka, Bangladesh. Explore full architecture case studies and production code at parvejshah.com.
