James Cook: I remember very clearly when chat GPT first came out. November 2022. I was with my two nephews for Thanksgiving, and we had all downloaded it to our phones. And we just could not believe what it was capable of. We would ask it to write a rap song about an epic battle between Luke Skywalker and iron man. And it would return it to you instantly. And it turns out we were not the only ones to employ this amazing new AI tool for such general silliness.

Daniel Fenton: My personal use case was creating songs to try to get my daughter to brush her teeth, which was partially successful. But then fast forward to February when GPT four came out and realized that it can write code, which is pretty cool. Like That's not a toy. then we soon after discovered that it could abstract leases.

James Cook: That's Daniel Fenton. Daniel leads. Our generative AI application called J L G P T, and the developer platform underneath it. He quickly realized there was a huge potential for this new tool in commercial real estate. Property data has been around for a long time. We use that data to manage our portfolios and build predictive models. But generative AI brings along with it. The ability to interpret that data in a matter of seconds. For example, you now have a smart sidekick that can effortlessly understand and categorize lease documents.

But here's my big question. How much can we trust all of this new AI magic compared to what we humans can do?

Ankit Kapoor: Depends on how much prompt engineering you really put into it.

James Cook: I see. So, it's all about how well you can create the prompts that is Ankit Kapoor. He works for McKinsey and company, helping their clients build digital solutions.

On this episode of trends and insights aren't get, and Daniel are ready to school me on the best ways to use these AI tools to supercharge real estate property data.

James Cool: This is Trends and Insights: The Future of Commercial Real Estate. My name is James Cook, and I am a researcher for JLL.

James Cook: So, Ankit. I want to start with you and let's talk big picture first. What's your point of view on the most immediate impacts that AI might have specifically within the commercial real estate industry.

Ankit Kapoor: There are lots of different directions that a company, might take or want to go. When it comes to how do you use A. I. And now Gen AI, that's a much more possible tool really use across gamut of different use cases that one might face.

We see folks, really starting to spend time on a couple different areas. Namely, I think data extraction is a really big one, then utilizing that data that you have extracted in order to better manage your portfolio of assets. And there's a few different elements and angles to that we can get into.

And then, lastly once you then have that data building predictive models on top of it. And so that's really the main, set of stages that a lot of the clients that we work with end up going through in order to really set up their, call it next three to four years, roadmap in terms of tech and data.

James Cook: It's interesting because this is like all stuff that most big companies are already doing, data extraction, creating prediction models, but hopefully we're doing it better, more accurately, more efficiently. Is that kind of the idea?

Ankit Kapoor: Yeah, that's exactly right. So, let's take data extraction instance. So, data extraction, that's something that historically, a technology called OCR, for instance, which is extracting information from PDFs has been around for quite some time. Where we've now gotten to in terms of, what a lot of these cloud players, such as AWS or Azure can offer out of the box, is really quick, cheap and efficient solutions to be able to actually ping the APT that service that allows you to operate at scale and do things much faster and much more efficient manner than you used to be able to. So that's just one example in terms of, a technology O. T. R. That's now become commoditized. GenAI would be another one, right? So, the ability to actually then interpret the text behind it has become a lot more mainstream before, you used to be able to do things like sentiment analysis, and that was a little, commoditized.

But now you can do far more powerful exercise and activities with that underlying text data that you couldn't really before.

James Cook: We’ll start off with the lease itself. These days, it's going to be a PDF, right? And maybe in the past we had people whose job it was to just look at it and type something or copy something into a spreadsheet or something.

Daniel Fenton: It's not in the past.

James Cook: In the, yeah, in the not-too-distant future, it'll be in the past. If that makes sense, it probably doesn't. So, tell me Ankit, what can we look forward to with the data from these leases? Is it just like flipping a switch?

Ankit Kapoor: It’s a pretty exciting set of analytics use cases that we can now do release extraction. So, you can do a couple of different things. So, one the automated process of OCR that I mentioned before, that's actually taking something like a PDF, running it through and converting it to a set of, basically characters.

Once you have that, you need to somehow interpret that. And so that's where actually Gen AI comes into play. So, what you can do is you can actually train LLMs in order to interpret that raw text as different components of how a human might actually read a lease. So, something like, lease start date, end date, break point, different price points how that's going to evolve through time, different sort of legal agreements that you might have.

It's able to actually intelligently consume all of that and compartmentalize it into those different components such that someone that is a human would have to spend, 10, 15 minutes actually reading it and entering all of that into some sort of a form. You don't need to do that anymore.

So that's where we are. I would say, not too distant future. You're going to start to see lots of commercial real estate firms starting to leverage this technology as more and more companies, start to build out products for it or companies build out their own version of tools and technology around it.

James Cook: This begs a question for me, cause I try to test as many different AI tools as I can. A lot of times I'll have them, summarize an audio file or like a report or something like that. And if it's something I'm already familiar with, sometimes I'll find errors, right? So how trustworthy is doing it this way, as opposed to doing it the human way?

Ankit Kapoor: It depends on how much prompt engineering you really put into it. Right off the bat, with all these models you have to train them, so right off the bat, you might not be getting the accuracy or insight that you might expect by asking certain questions of it.

Or by, training the model to structure the interpretation of lease in a certain manner, but by building, 100, 200, 500 different prompts around it, you can actually train the model to get better. And spending some time with the model as well as then also putting sort of a set of checks after the fact is really important.

And that's something that our clients sometimes forget, which is you have the out of the box LLM that you're going to be using, Chat GPT and open AI, in order to build out. But on top of that, you actually need to have a check layer, right? So, checking for certain hallucination points that exist and the combination of those two then all of a sudden becomes pretty powerful where you have a really solid foundational model that millions of people around the world are now using. But then have your own sort of custom layer that sits afterwards that checks for things that you know might be off.

James Cook: And I'll just point out that term hallucination points for listeners that aren't aware, Generative AI sometimes make stuff up and that's called an hallucination. I feel like the future of AI is like ironing out those hallucinations.

Ankit Kapoor: But it's really interesting in that, in our estimation, this kind of technology, once you have all the information, you can see at a blink of an eye how many break points do you have in a given month or two? That's something that you would have to put a ton of time into in order to actually get an accurate representation of that.

If you build your prompts, you can get that in seconds. So, we think this is going to provide up to, 20 percent potential time savings for asset managers, portfolio managers, it's going to boost revenue by, half a percent. So, it's going to be real savings and a revenue driver.

James Cook: Wow. Wow. That is real savings. So, Daniel, I'm going to turn it over to you. You oversee our AI platform at JLL a large commercial real estate brokerage.

So how did this all come about?

Daniel Fenton: JLL really has like technology at its core at this point. We started telling people we're a technology company five six years ago, and people laughed at us, but it was a statement of aspiration. And if you look at where we are today with JLL Technologies, we've brought in all this talent from Silicon Valley and other places to really build a culture of people who are always looking at the cutting edge of technology to figure out how to supply commercial real estate. So that was the backdrop for when we spotted this technology early on. And like many of us started exploring chat GPT when it came out November 2022, didn't actually know what it was useful for. My personal use case was creating songs to try to get my daughter to brush her teeth, which was partially successful.

But then fast forward to February when GPT 4 came out and realized that it can write code, which is, that's not a toy. Then we soon after discovered that it could abstract leases.

Early on, because of the, our culture, we had people who are discovering what this thing could do. So, we took a forward leaning approach, knowing that people were going to use this technology, but maybe not in the safest or in house way. And we said, okay, how can we harness this technology? How can we create an environment that enables distributed innovation, which is something that JLL is very good at, but do so with guardrails? And so, the first thing we did was come up with a policy which said here are the ways that you should use Generative AI responsibly. Here's how you should protect your JLL client IP and conflicts of interest. But we knew that we were going to need our own system to be able to make that happen at scale. That was the genesis, and it was like immediately adopted by a large part of the organization with tons of room to improve. And we've just been running from there.

James Cook: So, Ankit, when you're working with clients, how sophisticated are they around this stuff? Do you have to do a lot of education in the process?

Ankit Kapoor: The commercial real estate industry as a whole, is a little farther back than other industries when it comes to adopting that new technology, analytics, et cetera. So, if I had to guess an 80, 20 split is approximately, 80%, probably a little farther behind 20% quite sophisticated advanced know what they're talking about, how the tech in place.

For that 80%, I would say, it really starts with use case prioritization, right? The worst thing I think you can do is to go and tackle all of your data and just, focus, spend two, three, four years, on cleaning a bunch of data with no real use case in mind.

So, if you are a bit farther behind, first thing you have to do is assemble all of your use cases and prioritize them. Once you can do that, you can then see, okay, for those use cases, what are the different data elements that are going to be required and necessary? What are the different tools of technology that are going to be necessary, for those use cases?

And then only do you embark on that journey of, okay, let's build the 1st, one or two proof of concepts and go from there. I think some commercial real estate firms actually struggle with being able to contextualize all these different things that they can possibly do, right? So that tends to be one of the first steps that we take to really set them on the right path.

James Cook: And I'm not sure who this is the right person for but let me put it out there. So oftentimes when we talk about technology and software, there's this big question of, do you buy something off the shelf, or do you build it yourself? But is it fair to say you can't really build this stuff from scratch? You've gotta use somebody's existing model because, these are so expensive to like, build and train. Is that fair to say?

Daniel Fenton: Yeah, I think it's fair to say for the commercial real estate industry that we're not training foundation models, right? That is a very expensive proposition. Not just the compute, but like the staff that you need to do it. There are only a few companies in the world that are doing a great job of this. I think perhaps a cautionary tale is like Bloomberg GPT. It was very exciting. Bloomberg trained their own GPT model on finance data and showed that it was better in some range of queries. And then, like shortly thereafter, it was shown that just prompting GPT4 correctly did a better job. And so it was like a cautionary tale for training or even fine tuning models. So, I don't think the commercial real estate industry at this point is going to be doing a lot of building financial models. And then there are products that you can buy that are like the application on top of the large language models and some of them are better than others.

But when you look at the scale of a company like JLL, which has a hundred thousand employees worldwide, at that point, it really makes sense for us to build our own. And then because we're building our own, we really can tailor it for commercial real estate with these skills, knowledge and tools that you can't get elsewhere. And so, I think it really depends on your scale and what you should be building versus what you should be buying. Yeah. If you're small, maybe you're going to focus on just building a few tools, curating your knowledge. If you're big, like us, you may be building your own application.

James Cook: So, when you're building like your proprietary whatever tool, you're taking a foundational model and then you're training it to do something specific for you. Like Daniel, what was that process like? Do you just start throwing proprietary data at it? How does that work?

Daniel Fenton: What matters most is first, how powerful is the model that you're using? And then it matters how good are the instructions that you're giving it? So that's really what it's mostly about is creating really good instructions that go into the model. And then you look at JLL on our scale, we have people within the company who have become experts at this, who have figured out like the right prompts, the right instructions for their line of work, their business and redistribute those to the company at scale. So that's mostly what it looks like.

It's about the instructions versus this concept of in technical terms of like fine tuning or training models.

James Cook: Ankit, in your experience, has the trick been in figuring out, this prompt engineering?

Ankit Kapoor: I would say that's probably, 80, 90 percent of it. The other 10 percent is probably a little bit more data engineering, just to have the right data being sent in and out of the model. In the case of that lease example that I provided before, there's the underlying LLM that we're, we're using open AI for, there's then a bunch of data engineering that we have that actually takes underlying data that might be external to the concept of that lease and pulling it in terms of other details, joining that on.

So, to make the experience more holistic and powerful for the user, you might want to provide more than just the results or outputs of LLM. It's just like any other product or tool that you're building and an LLM is just one sort of arrow in the quiver. But in getting like the basic sort of functionality, it's all about the prompt engine.

Daniel Fenton: And I was going to call it like boring old engineering,

Ankit Kapoor: This is what we've been doing for 10, 15 years, right?

Daniel Fenton: Yeah, totally, right? Did you read the right set of documents and load them in to your work? Were you able to pull your data out of a database to include all of this together? That's classical engineering. And then people also underestimate, I think like, building the application around this stuff to do a particular job. So, in the case of lease abstraction found that we're really far away, if ever, from a totally hands off of a lease abstraction system, especially if you like really need it to be accurate. It just doesn't work. It's unlikely to work for some time. And key is building something that makes the humans go faster and helps them be more accurate. And part of that is about how you use a large language model, but a big part of it is the application that wraps around the large language model that helps the person with their workflow. And that's just, classical engineering.

Ankit Kapoor: Yeah, Daniel, it's almost using it in order to flag or alert things that require more human review, right? That's something that even, before this whole GENAI wave and traditionally AI, like in the banking sector, for instance, there's the project I worked on a while ago where basically we were flagging companies that were at risk of basically defaulting, right?

And when looking at credit downgrades, for instance, and so you don't want to necessarily assign a credit rating to a company, but what you can do is develop a model that says this company is X percent more likely to have a credit downgrade than every other company in its cohort, or that's like it therefore, hey, credit analysts, you should spend a little extra time on this.

And so, you can take that concept and really apply it to anything and anywhere there is high risk associated with making an automated decision, based on the output of a model, you don't want to do that. So, you want to have the human intervention. You want it to basically be flagging and alerting things, but really just making your day-to-day analysts more efficient versus, really replacing the actual, job function that they do.

Daniel Fenton: Exactly. And over the long run, if you do this right, it creates a data flywheel where what you flagged for the humans gets used to create better training data over time, you gain more and more confidence, and you automate more. But the stage that we're at today, it's really about building tools for people to help them go faster. And there's a lot of room for benefit there.

James Cook: You guys you're blowing my mind here. I went into today with this thesis that we were going to talk about how AI could make your data better, but what I'm realizing is that what it really does, it's turbocharging the people who are interacting with the data and giving them What seemed to me like superpowers to deal more efficiently and quickly with like large amounts of information and be more accurate and thoughtful and efficient their work.

Daniel Fenton: It definitely is that, but there's also this. We can unlock data from things like images and now videos and audio. Like we talked about, massive text documents at the beginning of this, like leases, that’s from last year's version of this technology. You could upload a lease and turn it into structured data and find the bits of it that were interesting. The frontier now has moved. What can you do with video? Take a drone inspection or a fly through of a property. There is data that you can pull out of that, that then you can bring into your classic machine learning models as features that help you do better predictions. And yes, it's going to turbocharge people. And I think we're really focused on that, we're getting a lot of value out of it today. But where the frontier is what data can we unlock that was always sitting there but was impractical to pull out at scale.

James Cook: Interesting. So last year was numbers and words. This year, it's like visual information. Just for fun, we're going to do this podcast again in five years. Say. What are we talking about? What is going to be cutting edge in five years on get any guesses?

Ankit Kapoor: I'll let Daniel go first on this one. He has an answer ready to go.

Daniel Fenton: A year ago, I didn't know what a year ahead of us was going to look like. Because there was just so much to explore. You had the AI companies as part of their marketing strategy say that superhuman AI was just around the corner. And we really need to think about what jobs people are going to do because the current ones are gonna be gone. I think that was a marketing strategy. It was very successful. And it was not true. But things were really blurry. I think at this point, the next two years are easier to figure out if you follow the research and you look at what startups are doing. The next five years are really blurry. But I do think that this concept of agents, or as I like to think of them employees, is going to be a thing where you could train a virtual assistant as if it were an employee and imbue it with these skills, knowledge, and tools and teach it, let it ask you questions. It will remember, it will update itself over time and be able to get to work and pop back up to ask you questions when it's needed but work over like extended period of time. I think that's going to be a thing that exists. These virtual employees. What gets interesting is when you have like tons of them and they're working together, they're working on your behalf with other people.

Ankit Kapoor: I think in five years’ time, every single employee at a commercial real estate company just because that's the podcast, is going to know how to code to an extent. We're already starting to see it with prompt engineering is a version of coding, I think in five years’ time, it's going to be just like being able to use Excel.

James Cook: This has been a fascinating conversation. Ankit, Daniel, thank you so much for joining me today.

Daniel Fenton: Thanks for having me.

Ankit Kapoor: Thank you for having me.

James Cook: And oh, in five years, my automated assistant will contact your automated assistants and we'll schedule a time to catch up. How's that sound?

Ankit Kapoor: Beautiful.

James Cook: If you liked this podcast, do me a favor and go into the app that you're listening to right now and give us a rating. Even better. Give us a little review, just write a sentence about one thing that you liked about the show. Of course, you need to be subscribed to trends and insights, the future of commercial real estate in that same app to get a new episode.

Every time we publish. Or you can find us on the web anytime at jll.com/podcast. We'd love to hear from you. Send us a message, a note, an idea for a new episode, whatever. Email us@trendspodcastatjll.com.

This episode of trends and insights was produced by Bianca Montes.

Our theme music was written and performed by Joel Karachi.