Roof AI CEO on ChatGPT — and the need for guardrails 

Roof AI has been in the AI-powered chatbot business since 2016 — also known as 6 B.C. (before ChatGPT).

And it was one of the first players of its kind. Now, nearly two years after ChatGPT made conversational, instant generative AI available to all, Roof AI has plenty of company. It also faces challenges that are unique to the real estate industry.

"If you're developing a real estate AI assistant like ours, ensuring compliance with fair housing laws is essential," Roof AI CEO Pierre Sabbagh. "However, GPT lacks robust guardrails in this area and others. You'll need to establish your own safeguards, which is no small feat."

That's what Roof AI did for today's launch of Roof Pro, which launched today and will roll out over the coming weeks. Roof Pro uses ChatGPT and other advanced large language models (LLMs) to help brokerages engage with consumers in real time throughout the homebuying journey.

"This integration will allow us to answer buyer questions about the real estate process, neighborhoods, and properties instantly, keeping them engaged on our website and fostering trust," said Keyes Company chief information officer Wendi Iglesias.

Sabbagh provided written answers to our questions, which have been edited for brevity and clarity.

What has surprised/amazed/scared you about the rise of ChatGPT and other large language models, especially in the world of real estate?

Large language models have introduced some hallucinations in their output, requiring an extra step of due diligence on our part. In real estate, for example, we've seen instances where LLMs invented a MLS system that didn't exist, generated market trends and statistics out of its own imagination, and more. However, these are merely artifacts within the system that are gradually being addressed with each new version.

What should people know about the Roof Pro guardrails around ChatGPT and what did it take to make them?

You may not want to rely on GPT (yet) for tasks such as internet searches and gathering real-time data, like market statistics or property information. This is because (a) GPT is still prone to hallucinations and (b) it's unclear where it's really sourcing the information from (yet). It's important to maintain control over your data sources and dictate how the AI interprets and processes that data. Simply preventing GPT from handling these types of requests presents its own R&D challenge, one that we encountered at Roof AI.

Exposing any LLM directly to your customers is not ideal most of the time. So it's important to enforce guardrails and establish your own barriers between the LLMs and the customer, particularly concerning content safety. Implementing these measures took us over six months.

How do you balance moving fast vs. doing it right and maybe not being first every time?

AI moves quickly, and it's only going to accelerate. We can't afford to relax, as this "hype" isn't fading away like many others. So it's essential to continuously test and experiment with new models, which seem to emerge almost weekly at this point. But when it comes to developing scalable products for our customers, it's crucial to take the necessary time.

Some of these LLMs are black boxes. You must take the time to understand how to manipulate them while also dedicating time to learn about their individual strengths and weaknesses and how to navigate around them.

Take Roof AI, for example. Last year, we embarked on what seemed like a "small" project to integrate OpenAI's GPT into our Conversational AI. What was initially planned as a three-month project turned into a year-long journey filled with technical challenges and moments of head-scratching. Why? Because it took time to discern what GPT excelled at and where it fell short. So we went on a tour exploring a variety of other LLMs to complement GPT's capabilities. This led us to develop a hybrid stack of LLMs that could effectively communicate with one another after months of trial and error."