How to build a Real Estate Agent AI Chatbot

We’ve just gone through the learning curve of building (and rebuilding a number of times) a hopeful “Turing Test” passable Real Estate Agent AI Chatbot. Was it easy? Hell no. The Turing test is a huge benchmark for a reason– the natural language we bumble through even as a toddler, is incredibly sophisticated. To get a computer to do that like a real Real Estate agent in a SciFi show (there may not be too many real estate agents in a sci-fi show, but you get the idea), is really light years ahead.

It’s live and working. Check it out on MotionProperty.com.au and their Facebook page (yes, it answers all live chats on Facebook as well). We have at best 6 months of programming to improve it.

The AI Part of any chatbot engine is really just a comparison tool. It looks for like-minded words or phrases, and comprehends if you replace one for another. For example, “Rent” could be “Lease” or “Renting”, and as soon as you give that example, is Rent as in “lease my place” or “I want a place to rent”, interchangeable again is “I want to live there”, which also means “live” as in performance. This shows how fast things can escalate outside the scope of AI or at best, form part of the training inside the AI platform.

We came up with a couple of processes to support these challenges– you can approach this from Top-down or Bottom-up.

Bottom-up is where you sort of go for it, and start with an open-ended “ask-me-anything” chat and try building in intended cornerstones to get to the bottom- the info you want to capture, the person’s name, location of interest, budget, etc.

Top-down is where you build out an elaborate data flow diagram (DFD pictured here), showing all the “pathways” you think how most conversations will go down. We did this from our second rebuild onwards for two reasons:

First, there were too many options to cover off in the Bottom-up approach. So while we did get something working, the potential conversation pathways were enormous, and conversations that got to the end were under a few percent.

Second, we could launch what was basically a multi-choice questionnaire from the DFD, so that 100% of the conversations reach a conclusion. It wasn’t anything better than a dumb form at this point, but it gave us a lightbulb moment. We could tackle the programming in a piecewise process, removing layer by layer of the multi-choice into smaller AI conversations. This approach worked superbly. As we pulled out the first multi-choice from the top of the DFD tree, it changed the entire feel of the conversation– far more human and realistic.

While Bottom-up chatbot programming can work, it seems only applicable when programming something simple, such as handling a single Question pathway (example below). This liner pathway is essentially a zoomed-in version of every box in the larger DFD.

For our “Be the Agent” AI chatbot with 5-6 key services and a range of sub-trees for every conversation point, the Top-Down approach has been a great solution so far, and is only getting better with every iteration.

For more info on our Chatbot strategy, design and development services, please see our section on Chatbots under Our Services.