The Trailblazing New Uses of Generative AI in Home Finance

Fresh off of a tech sprint being held in July, Chief AI officer of the Federal Housing Finance Agency, Tracy Stephan and lender innovators who were a part of the competition come together to talk about the tech of tomorrow. This discussion exploring the creativity and ingenuity of the sprint's participants will reveal not simply what is possible today, but the exciting, cost and time-saving capabilities of the near future.


Transcription:

Maria Volkova (00:09):

I've just been introduced, but I'll introduce myself once more. My name is Maria Volkova. I'm a reporter for National Mortgage News. I'm excited for today's panel discussion where we will be discussing FHFA's recent tech sprint. It highlighted how conceptual generative AI applications can remove certain pain points in single family in the single family and multifamily space. Joining me for today's discussion is Leah Price, Senior Financial Technology and Innovation Specialist at the FHFA and Ziggy Jonsson, Senior Vice President of Engineering at Better. Thank you guys for joining me.

Ziggy Jonsson (00:55):

Thank you,

Maria Volkova (00:56):

Leah. So as an organizer of the Tech Sprint, I was hoping you can discuss it and maybe talk about what the goals were for the Year's Tech Sprint and why the focus was specifically on Gen AI.

Leah Price (01:12):

Sure. So let me set the stage a little bit. All three of us were at the tech sprint this year. So Maria was there covering for National Mortgage News. I recruited Ziggy from his very busy job to come spend a couple days with us and this was actually my second tech sprint last year. I participated as a lender. Alright, so actually did any of you watch the tech sprint? Okay, not that many. Alright, so let me start high level. So a tech sprint is kind of like the idea of a hackathon. You get a group of people together and you work on some kind of a problem. So sometimes that can mean you build a solution in over a certain period of time. For us at FHFA, the way we set it up is we had a four day program where people from the industry came together and worked to come up with a use case, not necessarily a product, and also some risks and control measures related to the specific type of technology, which this year we chose generative AI.

(02:21):

And so your question was why Gen AI? And the answer is because everyone is talking about generative AI, including the regulators, people are trying to figure out and wrap their heads around the risks. And we really wanted to be at the forefront. No other financial regulator has led a tech sprint over generative ai. So we wanted to be at the bleeding edge in that case and we wanted to engage our other fellow regulators as well. On the process, let me tell you a little bit more. So this year we recruited mostly lenders and this event was also oversubscribed. So we had hundreds of applicants for a limited number of slots and specifically we set it up. So we had 12 teams. Our industry tends to be very single family focused, but this year we wanted to make sure we had multifamily and servicing teams. So we had a very good distribution. And the problem statement is here in bold, there are a lot of words to it, but the idea was how might the responsible use of generative AI promote a transparent, fair, equitable and inclusive housing finance system while fostering sustainable home ownership and rental opportunities. So that was the inspiration.

(03:46):

And there were four buckets. So each team had to select an area to compete. The four buckets were consumer experience, credit worthiness, operations, and risk and compliance. And then we had a winner in each category that was selected by a panel of judges. Winners didn't win anything except bragging rights on LinkedIn. Okay, that's all I got.

Ziggy Jonsson (04:16):

Which means everybody won.

Leah Price (04:18):

Yes, everybody won. It was a great experience. The video is available on YouTube. FHFA has a YouTube channel and if you look for demo day, you can watch it there.

Maria Volkova (04:27):

Everybody was a winner. That means Ziggy was a winner too because he was one of the 12 teams that presented a product. And Ziggy, I was hoping you could maybe tell us a little bit about the product that you and your team built.

Ziggy Jonsson (04:43):

So when I got the call FHFA tax print on generative AI, I pinched myself a little bit. Is this actually happening? Talked to our contact at Fannie and he was very excited too. So yeah, this is happening actually, it's interesting. You don't know what's going to happen there. My experience of hackathons is usually you have a lot of engineers and they're programming in real time together. I think this is more like a lot of business people, but tech people as well. And I think you managed to get the mix very nicely. So there was a tech person in each team. It's interesting, you go there and you meet the team for the first time and day one this MS right day one is everybody has an idea and then we poke holes in this other idea because we want to win. This is a hackathon, it's a competition, right?

(05:32):

It's a competition, but at the same time it's a great venue to ideate with your colleagues that you're not working with day to day about what's actually important, what's relevant, what can help the consumer, what can help the industry went back and forward. But we ended up at a solution that actually surprised my general counsel a lot better. He was happily surprised when I came home and told her what is, basically what we came up with is a real time compliance layer available to the industry. And one of the things that is hindering broad adoption, particularly in customer communication, is LLMs in generative AI as non-deterministic, right? As opposed to the old models we know very well, it's like a decision tree. If you have an IVR, you know exactly what Ivar is going to do with LLM don't, it's a black box but you can get pretty confident.

(06:30):

But what happens in the conversations where your LLM is actually saying something, it shouldn't really be saying. And what makes the market industry very special is that it's heavily regulated and it's like there are federal laws regulations, you have the most state level, you even have them locally. So the problem is very complicated. You could have the LLM that gives a perfectly right answer, but the borrower happens to be in Texas and actually nonetheless actually applies. So our idea was basically a compliance layer. Obviously speed tiers of importance. If you are a lender, the idea is you can work with any AI company, whether it's a chatbot or real time speech, whatever it is. But then right before we respond to the customer, you actually paste the conversation is actually you do an a PO call to this layer and instantaneously you get a red or red or a green flag.

(07:22):

Now obviously most conversations are just going to get a green flag, but in the cases where your LLM may be responding to the customer that he or she can pay four points on a HELOC in Texas, you get a red flag. So actually what you say, don't say what your LLM was about to say, you actually say we are going to bring in a human into the loop here and that can assist you further with your questions. I can't answer that. So that was the general pitch. And we actually did have, so thankfully this was long enough that we could actually do a working demo. So we actually had a working prototype, obviously not the full thing. We did connect it to a bunch of regulation Fannie Mae sellers guide via a technology called rack. Many of you know what it is. And we created 20 test cases where some were bad, some were good. And I'm happy to announce, which is why I think I won in reality that we actually passed all 20 of those cases. So obviously very much of a prototype, but it was very interesting and it's kind of interesting to work with a team in this sort of environment to come up with new solutions.

Leah Price (08:36):

Yeah, I'll say one thing that was interesting is that seven of the 12 teams were able to build a working prototype with Gen AI tools. So that was exciting to see.

Maria Volkova (08:46):

Yeah, I was hoping that we could maybe aside from Ziggy's project that we can discuss any other projects that stood out to you guys.

Leah Price (08:54):

Yeah. Alright, so what I have here is a slide that explains all 12 of the use cases. So I'll touch on the ones that were the winners. Winners considered the most promising in the different categories, but really they were all pretty fantastic. You'll see up at the top those were the single family lending use cases. You'll see the servicing use cases. One of the most promising was a homeowner assistance model, which would help detect when a borrower might be entering into a distressed scenario and help come up with an intervention plan for a servicer to approach the borrower to try to keep them in their homes longer in the multifamily. So it's interesting of the most promising three of the four were in the multifamily space and none of them were in the single family space. On the operations side, there was a Gen AI tool that would help detect multifamily fraud.

(09:55):

So that's a hot topic right now. We're really interested in this. At FHFA, there was a multimodal Gen AI tool that would be used to help multifamily owner tell whether or not a property was a compliant. So you could take a picture of a room or of a property and it would give you a compliance score with a DA, and then it would tell you what you would need to do in order to make it more compliant. And then the last one I'll mention is, was the winner for consumer experience. So that was what they described as an uber-like platform to match renters with landlords. A renter would fill out a common application, their data would populate this application that could then match them up with properties that they would be eligible for. Any stand out for you Ziggy?

Ziggy Jonsson (10:49):

So it's interesting. I actually, I loved all the projects here. I think the one that stood out for me, it was kind of strange. I think this is something we should do right now. I got that feeling with one of the projects and that had to do with down payment assistance being available in DU and LPA. And just realizing when you think about affordability, there are plenty of customers that got denied by the system that should have been able to get to a house. And we as financial professionals, we actually didn't really do the work to figure out how we can make that possible. So having that part of the response when you hit the agencies, it seems it's a great idea. Now, generative AI, the reason why generative AI comes into play is that the data around those down payment assistant program is not structured. There are a lot of them and they're changing and there's a lot of text around them. So obviously I don't think you can get to DU responding with this is available, but DU can say it looks like the following programs may be applicable if the borrower has issues down permit. So I thought that was very compelling

Leah Price (11:57):

And I would say all of these are excellent ideas. We would love nothing more for the industry to go out and build these. So these are considered a public good. If you build them, let us know. We would love to hear some success stories about how these projects drove down. Cost reduced complexity helped with risk management.

Maria Volkova (12:23):

So as part of developing these projects, teams are required to identify risks and implement controls to mitigate things like AI hallucinations. Could you guys discuss the risks associated with Gen AI and how to be cautious when bringing generative AI products to market? And I was hoping maybe Ziggy can first discuss the thought process behind the risks and controls of the project that you guys presented.

Ziggy Jonsson (12:52):

Yeah, it's very interesting. In our use case because our solution was the guardian layer for the risks in play, then it's a question of who guards the guardians. So we actually, we put a lot of thought into this and actually kudos to FHFA in terms organizing this, making risks a big part of the scoring mechanism because honestly I think we shouldn't be afraid of the risk, but we should tackle 'em systematically with solutions that we can trust because we lose if you just say, Hey, this is too complicated and too risky, let's not do it with the way we win this. Here are the risks, here's how we tackle them, here's how we can move forward for the benefit of the customer. But from our perspective, we actually, it's interesting when you have short amount of time to spin off an entire project, probably a lot more work needs to go into this.

(13:44):

But our solutions had to do with samplings that are kept for 48 hours. We actually OVAs, which is a security term that some of you may know, they have a top 10 lists for LLMs, which is kind of a cool benchmark to start thinking about risks. So we actually addressed that list that went into that. I don't remember all the specifics, but the key thing is making sure that we are sampling and you have regression tests that are testing the output. What's actually interesting is most people who are working on the LLMs, you have a foundational model and then you may have a layer or prompting on top the foundational models, they can be changing. So if something worked last month, it may not be working this month. So systematic testing and sampling was something that we highlighted as an important step.

Leah Price (14:32):

So I would say it's interesting being on the regulator side because we're very biased towards use cases, including my colleagues. It's like, let's talk about what's going to get built and everybody gets excited about, okay, how's that going to change things? And nobody really gets excited to talk about the risks and I think that's because it's so hard to think about. So this slide here just shows the top risks that the sprinters mentioned, which is consistent with what all the consulting agencies say. So every single team mentioned hallucination and accuracy as a big risk, then privacy and bias. So I mean the one observation is just that the mortgage industry is perfectly well aware of all of the same things in terms of risk that the rest of the world is saying. What's particular about our industry though is that people really gravitate towards using these tools for risk and compliance itself. A lot of the media talks about fraud and bad actors using Gen AI, but there's less focus on how powerful these can be to help with risk management. So that was great to see. And actually I think almost every single team wanted to compete in the risk management and compliance bucket, but we coached them out.

Maria Volkova (15:56):

Okay, cool. So what role do government agencies like the FHFA play in advancing generative AI use cases while ensuring that their development and application adhere to ethical standards? And to Ziggy, what role do mortgage lenders have in fostering ethical innovation with generative AI? Leah?

Leah Price (16:20):

Alright, so I'll start. I mean FHFA is taking a pretty novel approach right now and we're really leading the way with our other financial regulators as well. I mean to be seen if this is the right way to go about things. But we're being pretty bold in leading these tech sprints related to this cutting edge technology. Other regulators, O-C-C-C-F-P-B, they were all in the room. We had them as subject matter experts there with us. We hope that they will also hold their own text prints for themes that are relevant to their organizations. So in terms of what is the role of the regulator, I think every regulator has a different take on that and some will be slower to move and be more cautious. And then some of us will really try to get ahead of any of the risks that are out there by actually, because you don't know what the risk is if you can't even talk about what the use case or what the tool could look like.

(17:23):

And what's really interesting is that, so I had to recruit hard to get lenders to give up this much time. And some of the companies that I spoke to didn't even want to send anyone to participate in the tech sprint because they didn't want other regulators to judge their institution for having participated in anything related to generative AI. So to me, that paralyzing force that seems like the opposite of the way to go. You want people to at least talk about and try to experiment in order to know what could go wrong because otherwise you're really in the dark. Somebody could just build something out of the blue and that takes us all down, Ziggy.

Ziggy Jonsson (18:10):

So I think for example, tax prints like this, it's not necessarily about particular products coming out the sprint. It's you get all the positive people from the industry, not the negative people, to your point, to a single location. The subject is generative AI. I think that construct FHFA as an industry, where do we want to be in a few years? And as you're aware, most other industries, when you think about customer service or anything at 2:00 AM you want to call and you have some questions, you want to talk to somebody and you're fine. If it's an AI, this is where other industries are going. So the question for the mortgage industry is where do we want to be? And if you look at the cost structure of producing a mortgage, I come from Iceland. Getting a mortgage is more like an afterthought. You find the house you negotiate, there's only one realtor, not two, only one be able to minimum commission.

(19:05):

And you go and get a mortgage and you're done here in the US for some reason it's very expensive. A lot of that expensiveness is due to the complications of the mortgage process itself titled insurance and all that stuff. But then there's an army of people answering questions and many times they can't pick up the phone and this is too complicated for customers to easily out themselves. So my hope is that generative AI is not only going to improve customer communication and accuracy, but also reduce the overall cost, make timelines faster and make the markets eventually an afterthought for the customer.

Maria Volkova (19:45):

I feel like you've answered part of a question that I was going to ask now.

Ziggy Jonsson (19:49):

Okay. Sorry. I'm excited.

Maria Volkova (19:51):

Yeah. Okay. So the question is, what impact do you guys expect Gen AI to actually have on the mortgage lending industry as a whole? So this is a fun question, hypothetical question.

Ziggy Jonsson (20:06):

So here I'm talking about my personal opinions and not necessarily those of the company just to be clear. But I honestly think if we collectively get the use of Gen AI correct, I think I almost answered the question before, but I think given the complexity of the underlying products, if Gen AI can help simplify for customers who may not be financial derivatives expert, which is basically in as what you have to be to decide whether you should pay points on a loan or not, if you can use that to build a bridge where you are focused on your home, you don't need to think too hard when it's clearly in your interest when to refinance and how that process goes. That in combined with cost reductions is going to make mortgages faster, easier, and cheaper for everybody.

Maria Volkova (20:58):

Yes, from interviewing some servicers, there's a lot of hope that Gen AI will help with the call center operations, where maybe in some years you'll talk to an artificial intelligence to get your questions answered.

Ziggy Jonsson (21:16):

So it's interesting because you have all experienced talking to a corporation on the phone when you help with something, eventually you find this one person who went understand your problem and is going to do something about it and then they say, oh, I'll call you right back. And you're like, no, don't hang up, don't hang up. And then the person hangs up and now you're like, okay, I have to call in again. And the dance starts again. You get to people who don't have a clue what you need. So imagine if you have generative AI where we have tackle the risks and the machine is actually going to have less error rates than a human will be. And at any time you want to pick up the conversation, it's right there. Buying a home is one of the hardest, biggest financial decisions that people face. And as an industry, I think the customer deserves to get the best customer support possible along the way. And I would be very surprised if AI is not going to be a big part of that.

Leah Price (22:19):

So with my FHFA hat, I would say that I hope generative AI leads to improved risk management, reduction in costs and that consumers benefit. And from the Leah price perspective over the 5 to 10 year horizon, I feel like just as human beings, our lives are going to be dramatically altered by this new technology. So it's almost unimaginable what that means for something like the mortgage industry that I don't believe is going to be the leader in terms of adopting this technology. I think as human beings we're going to exist in a different way and have different expectations about how we interact with technology going forward. So I think it's going to be changed everything.

Maria Volkova (23:07):

And you've also taken my next question. So what do you guys expect, what do you anticipate emerging in terms of AI use cases in the next five to 10 years in the mortgage space?

Leah Price (23:24):

Alright, well I'm speaking on a panel in 10 minutes where I'm going to talk about that, that'll be moderated, but I'll tee it up to say, AI agents I think are really exciting, but I'll wait for that. But Ziggy, maybe you want to talk about it.

Ziggy Jonsson (23:39):

Honestly, I think the biggest alpha or the biggest game changer I think is going to be in customer communication and just understanding the customer circumstances, however complicated they might be. I think the devil is in the detail in terms of how models are trained and how they're implemented, but I fundamentally believe that that's going to be the game changer. And obviously generative AI takes unstructured data and takes it into structured data and I thought most of that stuff is cost reduction to a certain extent, but then also speed to another extent. But I think the biggest game changer is comps.

Maria Volkova (24:24):

Cool. Would you guys like to add any other closing thoughts about Gen AI?

Leah Price (24:30):

Well, I think one question that you didn't get to ask me was about multifamily and Yep. Alright, well I

Maria Volkova (24:39):

Multifamily highlighted.

Leah Price (24:40):

Yeah, so I'll point out Sandra Thompson specifically requested that we go out and recruit multifamily. And as somebody who's coming from the single family space, I got schooled by the multifamily industry on even how biased my problem statements were for the single family segment. I made sure that when we brought in people from the multifamily, I had four teams that were entirely made up of people from the multifamily industry. Unfortunately, I feel like on the technology side we're really focused on origination tech, so they don't get as much focus. I think you'll continue to see us trying to focus on that space going forward.

Maria Volkova (25:21):

Okay. Any final thoughts about Gen AI?

Ziggy Jonsson (25:26):

One thing is talking about Gen AI, the other thing is plowing ahead, you have to start somewhere. So we have been testing generative AI in customer communication. And I mean if you think about it, sales cost of a loan can be 25 to 30%. So if you can achieve both or reduce cost and increase the customer support, I think that the future of the industry is quite bright.

Maria Volkova (25:59):

Okay. Thank you so much guys. Thank you.