B (7:20)

Tim, I appreciate your structured response to my slightly provocatively framed question. So mortgage is not devoid of innovation, but definitely some challenges. So despite said innovation, a lot of the industry does run on software that was built decades ago. And so moving over maybe to Andrew first or Andrew and Mike, and we're going to get into the details of what you're building shortly. But just first off, why does the industry running on old software matter for homeowners and lenders? Andrew, maybe start with you.

C (7:50)

Maybe the first thing, the most obvious thing is that because it's inefficient, because there's all these different systems, the cost to originate goes up. And when the cost to originate goes up, that fundamentally gets passed through in the form of higher mortgage rates, which means, conversely, if you can lower that cost to originate, then you've also made homeownership a little bit more affordable in a similar sort of respect, because it's a really important part of every homeowner's journey, right? If you have a better experience, you're obviously decreasing the stress for them in something that they care a lot about. Just imagine if you're buying a home and you're trying to originate the mortgage or get a mortgage from a mortgage lender. You could be in a position where you know, you, because it's inefficient, don't know what's really going on. And then you feel really stressed about are you going to be able to close your sellers telling you, hey, this isn't going to work because you're taking too long. And so these things are obviously really highly stressful situations. A little bit less obvious of a reason is things like mortgage assistance, right? So mortgage assistance is the part of the equation where, you know, a homeowner who has a mortgage isn't able to pay their mortgage. And so just imagine that you're a homeowner who recently lost their job or you have a partner that's sick and so you have to step away from work. You apply for mortgage assistance, which is again, a pretty time intensive process because you have to send all this paperwork in, you have to get them to acknowledge. But it's a little bit like sending a lot of information into the void. And you don't really know what the process looks like. You don't really know the parts of the process and where you're really in the queue. And so you call in the next day, they say to you they're working on it. You call in a few days later, they say they're working on it still. And you like really don't know if they lost it. You don't know if something had happened and they're supposed to give you a response in 30 days. But you know, between when you submitted it and the actual 30 days, it could be full radio silence, it could be lost, you don't know. And, and then you start to wonder, am I going to lose my home? Am I going to have to move on all these like really negative thoughts versus if you had way better software, if all of this was more efficient, then you can kind of see every part of the process like a pizza tracker. Now we have since learned that Domino's does not actually give you all the steps in the pizza tracker, but Just imagine that the way that it should work, which is as all these things are happening, as you're going through the multiple parts of the process, you know exactly where you are, every part of the way. And you can call in and say, hey, I know I'm in this part of the process. How long does it take to get from part A to part B to part C? And that gives you a lot more comfort. And it's a way less stressful way of going about this. And maybe the last thing I'll sort of add on top of that. Right. Is there's a deep reason, which is ultimately, homeownership is a really, really key part of the American dream. And so if you improve the fundamental infrastructure that allows all of this to happen as a country, you're able to effectuate public policy faster and it actually improves societal outcomes. Simple one is obviously Covid always right. When Covid happened, there was a bunch of regulations that came into place. All of that affected both what people had to pay, the mortgage assistance that was available, and the infrastructure at the time just really wasn't available. So there was a lot of hectic maneuvering between all the different servicers, the regulators. Plenty of people were stressed. But there's definitely a better way to do all of it where everyone ends up in the right place a lot faster.

B (10:59)

We are going to title this podcast. Mortgage should be more like Pizza.

D (11:03)

Please don't. I've read the status tracker analogy enough times in my life.

B (11:07)

I hear you, Mike. Maybe over to you, like, where does the infrastructure show up either in the consumer or lender's life? Like, why should the smart person not deep in mortgage care.

D (11:15)

Yeah, I mean, I think Andrew covered a lot of it. A lot of it is cost. A lot of it is customer experience somehow, weirdly, like, Andrew talked about those things. And I'm going to talk about capital markets, which is like the inverse of, I think, what. What we usually do. But I think that there's obviously a huge amount of manual labor and structural cost that creates for lenders, which, to Andrew's point, like, the consumer is eventually paying for everything. I don't think anyone looks at mortgage lenders profits and goes, oh, my gosh, the mortgage lenders are like making tons of money here. We don't need them to lower costs. It's like, if they can lower costs, competition will probably mean that results in lower costs for consumers. The thing that maybe Andrew didn't touch on, which I think is also really important about this old infrastructure, tends to be the availability and transparency of the data. And so every time you talk to, like, a capital markets person or an investor or someone trying to enter the private label market so that they can create some of these innovative financial products that may be able to help consumers differently, the problem you always end up with is, hey, you've got a system that makes the loan. That system doesn't necessarily have all the data in a structured format to begin with, and you definitely can't get it all out. And then you can't figure out how to, like, intelligently kind of pass that all the way through the capital market to the end investor, to the person buying the bond, it's lossy even to transfer it to the servicer, which tends to, again, make a terrible customer experience, but also just create a lot of confusion and inefficiency. And so the lack of transparency, I think, in the data throughout the mortgage ecosystem is another one of those, like, hotly discussed topics that really, I think, contributes to potential unfair treatment. It contributes to a lot of regulatory overhead where, like, the regulators actually have to spend a lot of manual time and effort figuring out what's going on. It contributes to a lot of friction in the capital markets being able to develop new products or better understand the credit risk of the securities they're buying, which of course hurts pricing. And so there's of course, the consumer experience, which is opaque and confusing. There's the lender experience, which is expensive and manual. And then there's also this underlying data problem from the old infrastructure as well, which I think causes a lot of bigger downstream problems as well.

B (13:05)

Sticking with you for a second, you were already in the mortgage industry before you started Vesta. Do you talk a bit about. You could have started many different companies in the mortgage industry. You chose to start an los. Why?

D (13:16)

Yeah, so an los. LOS stands for Loan Origination System. You can think of it as the system of record that a lender uses when they go to originate a loan. And so I had the fortune to work at Blend with Tim, actually. I sat next to him and we worked on basically the consumer front end of that experience. So you go on like Wells Fargo.com and you apply for a mortgage, and that's blend. But what happens actually as soon as you submit your application is all that data gets handed off to an internal system. And the internal system is really where the data is considered the source of truth. It's where the processors, the underwriters, the closers are all going to be logged into that system all day long, looking at loans, looking at documents, Understanding and underwriting your loan. It's the system that integrates to the internal system that integrates to every other vendor that these lenders are using to manufacture the loan. And that kind of back office system is what they call the loan origination system or the los. And these are historically pretty old systems. I won't say very old, like Andrew replaces very old systems. I replaced pretty old systems that contribute to a lot of kind of like manual process and inefficiency. I think that when we were starting the company, there were three key reasons that we looked at which said, you know, it was time to build new L O. S. The first, honestly, was that the market was just begging for it. Like, I had the great fortune of getting to know a lot of big clients at blend and executives would just call me. It felt like every month. And it was like, when's blend building an los? And I'm sure Tim got this question a lot too, where it was like, you know, do you know of any new lose or which LOS should I switch to? I don't like mine. And there were actually a couple particularly cheeky lenders who asked like, hey Mike, how about you just come and build us an losing? And so you get those questions a lot. And I think that reason number one obviously is the market is demanding it. There's a lot of lenders who are just saying, hey, we need some competition in the space, we need some fresh thinking. Reason number two was really around operational cost. Andrew and I touched on it a little bit at a higher level, but fundamentally, like mortgage lenders that are independent mortgage banks don't really make money per loan on average today. And the main reason for that is the cost to originate continues to balloon. There are all sorts of contributing factors, but a big one is actually that the manual process basically like monotonically gets longer and monotonically gets more expensive. There's like very little ability to actually reduce the cost of processing, underwriting and closing alone because they're extremely human driven processes. And basically people only ever add to those human driven processes. They never subtract from them. And so that was like a clear existential business problem that these lenders wanted a lot of help with. And that we basically determined, Devin and I, that the only way you could actually move the cost to originate down was to replace the core system that those people were in so that you can completely change their workflow, you can take work away from them and you can cut cost. And then the third big thing for us was this belief that lenders are becoming more and more technology driven in their strategy. And that means that they no longer want one vendor that sells them a monolith, and they want to kind of construct a constellation of services that they consider best in breed, a combination of build and buy to really build a differentiated technology stack and strategy. And that the underpinning of that had to be a new, different shape of platform that was much more about an open ecosystem and open access to the data and APIs, and less about some of the traditional architectures. And so between those three things, I think it was just Devin and I left London. It was very obvious that the next thing to do in this industry was for there to be a new los. And that's kind of how we ended up here.

B (16:18)

Excellent. All right. And so, Mike, if you're more upstream in the los, Andrew, you're at a different part of the ecosystem. So maybe same question back to you. You also were already in the mortgage industry, decided to start a servicer. Why a servicer? Maybe explain what is servicing, since it's something that many Americans use and don't think of all that often.

C (16:37)

Yep. So servicing is everything after origination. As Tim had mentioned, the way that this all works is at the end of the day, a lender makes a mortgage and that mortgage generally gets sold to a GSE or some sort of entity. And what that creates is a servicing. Right. Which is the right to service and collect payments on behalf of the mortgage holder, that is the government, and remit it to the government and maintain the relationship with the homeowner. The person who does that is the servicer. So you're kind of the person who's in the mixture between a payments business, a collections business, an accounting business, a compliance business, and you kind of do all these different things. And then you're also customer experience. My reason for looking at this whole thing, and it's now we're flip flopping. Mike talked a lot about experience, and then now I'm talking about my very, very silly reason for starting the company is I used to be a mortgage investor. And so what that meant was I would buy these mortgage loans and I would have to go find a servicer. But before I did that, I actually started off as buying only public mortgage bonds, which means you don't have to do anything like someone's handling in the back end. But when we started buying whole loans, like actual loans, you then have to deal with all of the stuff that comes along with that. Right. Like you might have a mortgage loan, like you might have bought a Mortgage loan, but someone's literally got to go collect that money for you. And so we employed a servicer, and I was like, okay, this is kind of annoying, but I'll get the servicer to contractually do this for me. And then what I ended up learning was the first firm that I worked at had no back office doing this work. And so they were like, hey, Andrew, if you made this investment, it's your responsibility to figure out how all of this data flows and how we get the money and how this all sort of like, reconciles. So my job went from doing investments to very quickly being 50% of my time reconciling the information that we got, the data that we got. And like, for me, very personally, when you make these investments, you're like saying, hey, I think it's going to make X return. That means I'll get Y cash, my portfolio and position will be X tomorrow.

D (18:30)

And.

C (18:30)

And when it's all not that, like, you get yelled at by your team and the back office management saying, hey, your numbers just don't make any sense. Did you make a mistake on your investment? So you get really, really paranoid about the whole thing, because then you're like, I got to go dig in. I got to check every single line item. And so then you end up realizing the servicer's data has a lot of problems. Then I said, okay, this is such a problem. I need to go visit these servicers and go figure out what the heck is going on. Because this is just like an atrocity. This is ridiculous. So I go visit these servicers and then I realize there's this like, absolute Frankenstein of all these myriads of systems. Like, I think Wells Fargo at some point had 500 systems just to do servicing. And you're like, oh, this is part of my language, a clusterfuck. Like, this is ridiculous. And so you're like, okay, there's a technology problem here. This is like a technology problem. I mean, it's a people problem as well, but it's a technology problem. So my brilliant idea was, let me go find someone to go start a business, to go invest in this software. It'll make my life better and I can go back to doing my job. So I actually tried to convince my current CO to do it, and she told me, no, absolutely not. You're doing this first before I join. And so go try to get it started and maybe I'll join you thereafter. She ended up joining thereafter, but that's what got me started. And I said, you know what? I should Eat my own cooking. How hard could this be? Let me go start a company, let me start a servicer. This will be fun. Hint, it was not very fun.

B (19:49)

So maybe staying on that. Why haven't there been more new entrants in the servicing space?

C (19:54)

Yeah, so the truth is there's actually been a lot of attempts, more than people actually know. And a lot of the ideas that we have in terms of what we should be doing, they're not unique. Right. I think people are generally smart. It's a question of can you actually tackle the problem? And actually, incumbents have spent hundreds of millions of dollars funding spinoffs, and there have been startups who've been trying to do it. My deep belief here is there's a class of problems out there where there's this truly entrenched complexity that has suppressed competition long enough that there's this astronomically large opportunity. But you have to be patient, you have to be relentless as a team to actually take over that opportunity or take advantage of that opportunity. And in this case, like Mike said, what we really learned was you need to have a complete overhaul of the underlying systems in the business, which actually also means at the same time, you need to understand every single part of it. You need to deeply understand all the workflows, all the aspects, who cares about what all of those different things. A wedge product, which is, I think, fundamentally what a lot of venture strategies orient themselves around, doesn't actually fix the problem. You, like, find a specific process and you're like, I'm going to fix it. Well, it turns out you need to have all this context about all these different things. And there's 25 different systems you have to integrate with just to solve the wedge product or wedge problem. And so what are you actually doing? And so we took a very crazy approach, crazier now that we think about it, but it ended up working, was we said, let's go build a servicer, which is, by the way, a heavily regulated business. We got licensed in 50 states, including California and New York, which are extremely regulatory, difficult process. And we're like, hey, we don't know how to do this, but you should definitely trust us. It's going to be okay. We got multiple federal level certifications, approval, Fannie, Freddy, Ginny, rating agencies. And once you've done all of that, you've gotten the tickets to even be in the business. Now you have to go to these large financial institutions, right, who are also very regulated. And you should say, hey, those hundreds of billions of dollars that you guys have to collect. Let us do it. Let us learn how to do this and do it on your dime. Like, that's a lot of convincing you have to do. And we're not even talking about, like, the pain tolerance. You just live and build it yourself. Right. And so between all those different things, my honest answer is it's a matter of do you have the patience to execute over extremely long time frames, knowing that you can succeed? And the answer is, most people don't really like that type of business and that type of journey.

B (22:07)

Tim, with your former hat on as the CEO of Fannie Mae, right, Like you launched one of the best mortgage innovations, Day One Certainty, under your watch, how did you approach sort of either building new technology in house or partnering with new technology partners?

A (22:24)

So the way we thought about it at Fannie Mae was that there were some things that we could build ourselves, but mostly what we were trying to do was to facilitate the adoption of technologies that were being built by third parties. We recognized that we had certain proprietary things we needed to do. We needed to build the technology to do that, but that most of the industry didn't want to use Fannie Mae technology. It wanted to use its own technologies, and it wanted to promote competition and innovation in those technology providers. So what we tried to do was to create the conditions where those technology providers would see that there was a good opportunity to gain a significant market share. And so we basically pivoted towards trying to drive as much digital adoption as possible and to facilitate our consumption of that digital data. Virtually everything that relates to a mortgage loan is just a piece of digital data. It's about consumer credit characteristics, it's about property characteristics, and today, almost all that information is digital data. And so we put Fannie in a position to be able to consume all that through APIs and try to encourage originators, lenders and servicers to be able to communicate with us in that way. And we would facilitate that, and we'd actually try to make it easy for them to do it. And we create certain rewards for them to do that, like Day One Certainty, which gave them relief from certain kinds of order known as representations and warranties in the industry that created potential liability for them. So we said, if you deliver the data to us in a reliable way using high quality data, we'll consume that and we'll give you relief around that. The other thing is that from my point of view, everything about the consumer experience and everything for the lenders is like, slower, more complicated, more expensive than it ought to be. And so my view was that we should try to drive this digital adoption and make it faster, more convenient, cheaper for everybody, but not compromise any of the risk metrics involved in this business. Like we were not going to take on more risk in order to be able to do this. We actually thought that by using verified digital data, we could actually drive the risk down for the consumer, for the lender, for us, Fannie Mae, for the end investor in the mortgages. But we could actually make it also cheaper, faster, easier for everyone involved.

B (24:59)

And so maybe. Tim, do you have a couple examples of either data that's available that potentially the GSEs or others are not able to make use of, or things that consumers are historically used to paying for that maybe aren't necessary anymore, and evolutions that the industry could go through?

C (25:15)

Sure.

A (25:15)

So look, still, much of what passes for credit underwriting today is basically just like proxies for real data. So what do I mean by that? Today it's possible to actually deliver people's digital bank account statements to lenders, or let's say to Fannie Mae, and we could actually peer into that data and we can see exactly how much money is coming into someone's account every month and how much money is leaving their account every month and through. In that way, whether they're a regular W2 employee or they're a gig worker, or, you know, they're an independent contractor, or they own their own business, you can actually get a much clearer and more refined sense of what their true financial picture is than by relying on things like credit reports or credit rating scores or those sorts of things. So our view is like, we should be consuming the real data. Let's try to do real time, nuanced analysis of that and be able to underwrite people more effectively, but also more accurately and hopefully actually be able to expand the class of borrowers who could get access to credit because they might be non traditional borrowers. So that's one example. Another example is that I just look at how much money consumers spend in this process, the mortgage process, without getting a whole lot of real value out of it. So, for example, every mortgage loan that's underwritten or backed by the GSEs requires an estimate of value of the property. So what does that mean? Well, most consumers end up having to go out and pay some appraiser, typically a guy whose average age is 60 years old, going out with a tape measure and measuring how many square feet there are in a house and verifying the dimensions and the condition of the property. The reality is that there are Much faster, easier, more accurate ways to collect that data. And once the GSEs have that data, they have enormous databases of property values much bigger than any other private party has in the United States. And so at Fannie Mae, we already knew what the property was worth. What we really just wanted was somebody to go out and verify it was there and confirm that it hadn't burned down or had been struck by lightning. But there are much less expensive ways of doing that. Paying somebody a couple thousand dollars to go do an appraisal. And we actually really didn't care very much about the statement of value from the appraiser, which we found to be not as accurate as our own estimate and often was actually quite biased one way or the other and tended to be actually less reliable. The higher the loan to values, the more that you were actually looking for the property value to cover the risk. That's when the appraisal was actually the least reliable. So you think about what all that costs the consumer. And that's just one example. This is in effect, an indirect tax on all transactions that policymakers, regulators, the GSEs, lenders, consumers, we should all have an interest in trying to reduce that number to as close to zero as possible.

B (28:23)

So, Mike, I want to come back to building a loan origination system. And anytime you're trying to wedge in as a new company to an entrenched industry, there's generally two strategies. One is build a these days, interesting agents or other workflow on top of, and then use that as a wedge into becoming the system of record. The other is, all right, just take the pain at the beginning, build the entire system, and then sell a software that's, that's way better. You chose the much harder ladder strategy. Talk about why.

D (28:59)

Yeah, I think that there's a fintech has had like a debate over the last decade probably around whether you do like people call it like core replacement or core wrappers. And I do think that 10 years ago, everyone was in the core rappers camp. Like, my favorite is the company that I don't think was an Andreessen portfolio company that got acquired called Mantle. And it's literally called Mantle because the Mantle is the thing around the core, the earth. And I think to me that's like the pinnacle of the fintech strategy of that version of the strategy. And I think for us, I had the fortune again of working with Tim at Blend. And I think for us, like Blend in many ways was like the wrap around the los. And the question was, hey, can you actually drive a lot of the foundational change in the back office that you want to change without going in and replacing the whole los. And I think the answer basically was no. Like, I think that if we had tried to build something as a wedge around the whale that is Encompass, I think we would have probably ended up in a similar place where we could have added some value. But it would have been very hard to like, drive the fundamental change with lenders that you need to drive. The other thing that I've noticed, for example, when we're like implementing the LOS is like as we go into a customer and we go and deploy Vesta and we rip out whatever the existing LOS might be, a huge percentage of the project is actually not deploying the technology. And so it's not building the system, it's not setting up the system for them and configuring it. It's actually the operational transformation of changing the way that these people do work and the way that they think about the process. And you have to achieve that operational transformation kind of almost regardless. And so in some ways it's worked out really nicely for us because being that biggest project, being the fundamental platform and changing that is like the only way really to force a lender to understand that they're also going to have to do the operational transformation. And honestly, a huge percentage of the value, of course, comes not just from the fact that you're using a new piece of software, but that the software enables that operational transformation. That's what gives our customers really good results. And so I think ultimately it was just the blend experience. After having it made it kind of obvious, I think to Devin and I, that you also kind of get the curse of success where you have a big customer base. They're all very happy. Now you have to go take on as a competitor, your biggest partner. And I think that also is like a very awkward place for business to be. And so for us, it was like you have nothing to lose. Go after the big prize, day one.

B (31:03)

I think one of the other points you made, like, for instance, if you notice as a consumer as you apply for a mortgage, it's a very linear type process. And one of the innovations of Vesta is you've completely rethought the workflows, but in order to do that, you also needed to change and adjust the data structures.

D (31:21)

Yeah. And I was on, you know, a demo today with a prospect where they were asking basically, hey, this process is so dynamic, right? Everyone talks about task based workflow, mortgage origination. And one of the problems that happens is you get like halfway through the loan and then something happens. Their example was like, then the borrower gets married. I was like, okay, I don't think that many borrowers like get married in the middle of their mortgage, but it does happen sometimes. But you know, other examples are like, all right, the appraisal comes back low and that means your LTV is above 80 and now you want to switch from a conventional to an FHA loan. Like that kind of stuff happens all the time. And so for us, to your point, it was like a maniacal focus on getting the right data model and then making sure the entire process was going to be data driven so that you don't have this like weird path dependent workflow that you're trying to sketch out that is super complicated and has to satisfy thousands of pages of rules. And it's much more. If you model the data correctly, you can figure out all of the steps that have to be taken looking only at the data, ignoring all of the steps that you've already done. In a way. I don't know if that's really an innovation or that's just like an approach that we've taken that's been very successful. But like a very maniacal focus on data model and everything being purely data driven, I would say is a big part of why the architecture works at all.

B (32:21)

Yeah, and I think one of the most surprising facts to me is with many of the existing systems, only one person can be in the loan at a time.

D (32:30)

That is a surprising fact to pretty much everyone we hire at besta. Yes, a lot of these things were built on not like SQL databases in many cases, but actually even older, like the flat file databases. And so you have to atomically save the entire loan file in a save button. And so actually some people have hacked their legacy loses where you can have two people in the file at the same time, but if you actually do that, their saves will like overwrite each other to the tune of like hundreds of fields and there's no merging mechanism. So yeah, there are like database level constraints that certainly prevent good dynamic parallel processing and existing systems.

B (33:04)

Hence the rebuild the whole thing. Okay, Andrew, you've also chosen not to wrap existing servicers. So you started by not only building your own software, but also being the user of the software, or a subservicer in industry parlance, which is one, enabled you to scale very quickly. And then two, maybe talk about how the business of servicing is usually a 5% margin business. You've taken it up to a 50% margin business. So what did it take to build that? And where are you going to go from here?

C (33:34)

Yeah, what did it take to build that? A lot of blood, sweat and tears. So we started the company six years ago and we went through the process of like, okay, when you start the company you first have to get the licenses. It turns out when you get the licenses, sometimes they say you actually have to have a certain number of years servicing before you can get the next set of licenses. So we like went through that entire sort of, okay, let's do one step at a time, let's get this first set of customers, let's do a subset of states, et cetera, et cetera, et cetera. And then there's a lot of like stepwise function growth, right, because once you like know how to service a loan, you can probably service 10 loans, 100 loans, but you also like mess up on a whole bunch of different things and then you stall for a while, make sure you fix your core system. Because like Mike, we really, really believe that you have to fix the core to do a good job. And then once you've done that, you can say, okay, let me 10x again. And so six years into that journey, we're a top 10 player. We have 1% of US market share and we're probably the most profitable servicer to ever exist. We are putting all of that behind us and we're turning 100% of our attention now and focus to providing software to other services in this space. And that's how we actually get enough market share to have our product affect other people's lives. Right. And what's really incredible is that the amount of receptivity we've had because of the work that we've done on the operating side, many of the United States largest and most prominent servicers have signed up with Valen and that means we have pipeline of hundreds of millions in ARR. And what more, we now see this really, really large opportunity to not just go after the mortgage servicing business, but really build an OS for servicing in general or regulated financial type businesses. And so it's a really incredible moment from going from a place where we were trying to get a servicer stood up and we just literally wanted to make sure we survived and it worked to a place where we're now in this place where these whales, for lack of a better way to put it, are now in a position where they're saying, hey, we've seen what you've done, it's amazing, it's honestly transformative. And now we're really excited to use your software as the way we power our businesses going forward.

B (35:28)

So clearly we all think mortgage is exciting. We are in the middle of one of the largest product technology waves of all time, which is is artificial intelligence, which is going to apply everywhere. Maybe starting with you, Mike, where are the most realistic, immediate applications of AI and mortgage? What could work today in production?

D (35:49)

Yeah, I think the obvious use case that people have been trying to do for ages in origination, that we've deployed in production is reading documents. It's kind of like a very unfortunately basic use case. But the reality is if you think about the mortgage origination process, it's basically you've got like a pile of data that describes the borrower situation and a pile of rules that you get from the GSEs for the most part. And a lot of that data lives in documents today. And so a huge percentage of what you're actually paying human beings to do is read the document, extract the data from the document, key that into a system, compare that document to other documents. It's all really like document management. And so for decades the industry has been trying to figure out how do I take the data out of the document and just put it in my los or put it in a schema somewhere that I can manage it. And people have spent tens of millions, hundreds of millions of dollars like training small models, like training convolutional neural networks, even before there were convolutional neural nets, like training other stuff to try and read documents. And for the most part the only thing you could really read super successfully was like W2s. And of course a lot of the advancements in language models have meant that reading things like purchase contracts or title insurance policies or closing protection letters have become problems that are frankly actually really simple to now hook up because you just plug them into a large language model. Of course there's a bunch of like orchestration you have to do. You have to do a certain amount of reasoning and consensus across multiple LLMs and prompting and whatnot. But for the most part it's a much lighter weight effort to basically say, hey, I have an API that can read and I have a bunch of documents that need to be read. And passing that document there and getting the data out I would say is a place where we've seen a lot of returns for customers already and it's a pretty easy straightforward case. There's lots of other more interesting and exciting things you can do as you start to unlock like computer use and can help copilot Your configuration and there's all sorts of things we're thinking about. But the obvious low hanging fruit is the problem that has been there for years and years and years. And it's just some labs in San Francisco gave us an easy way to solve it.

C (37:37)

We use a lot of the same sort of use cases as Mike. We use LLMs to do summarizations, we have voice agents, we parse documents, all type of stuff. I'm going to say some things that will probably drive some amount of industry ferment. What we find super interesting, where we're going to really get to right, is personalized AI agents. And so if you think about what we do as Valen, we basically have big sort of like workflow system and task system and all these different primitives and mortgage servicing. And we basically allow our users to stitch together and put together their processes on our workflow system. And that's great, that's awesome. They can sort of orchestrate, they can control all of their business processes and have a lot of visibility and auditability, all that type of stuff. What's really cool though is that we also have the tasks that humans do. When you need a human in the loop in that workflow, which happens to be a really amazing place to get a lot of training data. Whenever the workflow gets done, right? Think about it as like there's a bunch of things that have to happen, a lot of context. There's a task that's the next step, which is human in the loop. They figure out the answer, they make a decision, the decision gets reported. And now you basically start to create these training sets. And what's really great about that is that then you can automatically start training AI agents based on that specific task. Which basically means if people put their entire business process on our workflow engine, they then are able to build automatic AI agents for that specific task. And so you have this amazing thing where you're able to then get people to not only put their business process and be automated, whatever else, but they can have their best people work on specific workflows. And you can learn from that, you can sample the information that you need, you can reinforcement learn and you can effectively clone your best people. And so what that basically means in our minds is that you can create a world where by building on Valen, you're building for an AI world by default, right? And you're not only being able to automate a lot of your work, but you're able to clone your best people. And that to us is like an extremely exciting thing. And also, honestly, at this point, again, to Mike's point, about some really, really great researchers in San Francisco doing a lot of great work for us, that's a world that is like, actually possible today. And we see that as a near future reality.

B (39:43)

Yeah, I think one of the most exciting things going on here is the software you can now sell in. Like the attempts before might have been 2x better. This is now 10x better in that all of the things you wanted it to do, AKA document ingestion, now actually work. It can incorporate and scale some of your best employees. And then I think, Mike, to your earlier point of the why now? Financial services used to be slow at buying and now there is a tremendous urgency from boards, from executive teams that have very much embraced the Jensen quote that AI is not your competitor, it's your competitor that adopts AI before you. And so bringing that into your business is a very, very good way to prepare for the future. Tim, I want to ask about AI in regulation. And there is a, like every industry, an increasing amount of regulation in this industry. Even just the Fannie Mae selling guide, I think is 1200 plus pages. How do you think about how I could potentially both help regulation but also make it more efficient?

A (40:41)

Interestingly, one of the first things that we did in terms of using language models at Fannie Mae when I was still there and I left in 2018, so it's been a while since I've been gone, but we actually took the entire servicing guide and applied machine learning to it and created an inquiry system that allowed lenders and servicers to submit questions using regular language and get answers without sifting their way through the 1200 pages of material, which today doesn't seem like an amazing thing to do, but seven, eight, nine years ago was really something quite an important step forward. And so I think it's an example of the kinds of things that Mike and Andrew are talking about. And the technology has advanced so much so quickly that I really do think that virtually everything relating to the regulatory aspects of mortgage will end up being subsumed by AI. This is a way to essentially take all of the rules and processes and procedures required by the GSEs and make them into software code. And not only can lenders deploy it that way, but at the same time, the GSEs and the regulators could actually enforce and then monitor compliance with that using those same tools. So much of what happens in the mortgage business is really just a function of human error. There's some fraud, but it's relatively small these days. But there's a lot of human errors that occur. And if you can take the human error out of all of this and replace it with code and be able to monitor it with another set of code, I think it makes the life of the lender much easier. And in the end, from my point of view, the goal here is to drive down the cost of the consumer. So if you can replace all those people who are doing their best to try as good a job as possible, but nonetheless committing errors and replace that with these technologies, it's better for the lender. They'll be able to improve their cost base, but it's also better for the consumer because they'll end up paying less for it. And in the end it doesn't increase risk, it actually reduces risk. It should make regulators happy. It should be consistent with what policymakers want to see.

B (43:02)

Mike, you talk to probably the heads of mortgage across most of the major institutions in the country. They're all wondering how should I prepare for the future with AI in it? What is the advice that you have?

D (43:14)

I would say there are a handful of things that really matter here. One is that most of these institutions are probably not going to get there on their own. Of course that's also a little bit self serving of a statement, but I do often use as a mental model we've seen like Meta throwing around like a hundred million bucks or whatever for researchers. And so the question I often ask mortgage lenders in general is would someone who could make a hundred million dollars from Meta really be in the mortgage industry at all, let alone working at your lender? And so I certainly think there's like a diffusion that will have to happen over time. But I certainly think that people that use partners and that find the right technology partners to work with, whether those be existing mortgage vendors, other partners that are more general purpose are going to be the ones who are able to get ahead first. I often advise people that what you want to make sure you're doing is you're figuring out how to benefit from all of these really smart people, mostly in San Francisco, who are working on making intelligence via API better and better as opposed to competing with them. And so there are still lenders out there who have told me they want to train their own models or they want to build their own agents. And for the most part I think that is probably not a great idea. And then it's really about how do you make sure that you are positioning your technology in a way that you have partners who are going to get better as the models continue to get better and that your business will basically allow you be able to take advantage of those models. So how are you changing your processes, changing your staffing, changing the profile of people that you're hiring so that when the technology is much better, one, two or three years from now, you're really like oriented towards that. And I think a lot of people see this AI thing and you know, people always ask like, where will the value accrue in the stack? And they're like, oh, it accrues to the models or oh, it accrues to the app company. So I have to go build my own apps, I have to go train my own models. What I always want to remind people is, yes, value will accrue to those two places, but most of the value in this AI revolution is going to accrue to the people who use the AI. And so how can you make sure that as a mortgage lender you are really well positioned to be one of the people who use the AI and that you're working with the partners who will bring you the actual innovations and AI as opposed to the people who are kind of like using AI to go fish for customers as a buzzword, which there are a lot of them. And so that definitely, I think is the hardest part is partner selection.

B (45:15)

Andrew, same question to you.

C (45:17)

Yeah, I think there's a, there's a lot of truth to like, we pursue certain strategies because we think that is true and obviously it's a little bit self serving because it is the strategy that we pursue. There's two things that I would think about. One is, look, at the end of the day, when you think about what does AI need, it needs a lot of context. It's really, really great with more context. It needs a lot of training data, it needs access to tools. And so like Mike said, don't try to compete in the sort of middle of the equation, right? Compete on really having the right platform to utilize or you have to decide you want to be a foundational model company. That sounds like a bad idea though. And so the question for me is like, more so how do you centralize all the context you have? How do you have clean trade data? How do you put as much of the foundational parts of your business in a way that AI can leverage and you can utilize it with the stuff that a lot of people are building? Right? And that will make you a big beneficiary in this outcome. Because ultimately to me, who are the biggest beneficiaries of AI, right? It's like the customers. And it's basically the operating systems that they run off of. It makes it like a really, really great outcome. And so I'd say that's one big thing to think about. And then the second thing about it is you should really think about mixture between the right partners and the right platform that you're working with. Right. Because we work in extraordinarily regulated space. And so you need the right guardrails. You know, people thinking about agentic behavior. Agentic behavior is kind of troubling, right? When it comes to a lot of the things that we need to do, there's a specific set of processes, a specific set of steps that you want to be able to follow so that you can explain it. One of my favorite stories is when I talked to the folks at FICO back in the day about like, why do they still use logistic regression to effectively come up with FICO scores? And the answer was it's just explainability. Like I need explainability because I need to be able to say this parameter did y in a very linearly explainable way, because anything but that, right? And this is 10 years ago when gradient boosting and whatever else was all the rage. That's not explainable. And then that means when I get in front of the regulators and I get asked all these different questions, I won't have a really great answer. And so I think working with the right platforms, where they've structured their solutions and their infrastructure in a way where you can kind of crawl, walk, run sprint, right? So you can do all of it in the right sort of steps and processes will be the right way and also the fastest way you'll be able to leverage AI for your business.

B (47:29)

Looking ahead, fast forward in 10 years, what does the best instantiation of the mortgage system look like?

D (47:34)

Well, 10 years is a very long time, especially in today's technology world. I think that in 10 years, consumers, assuming that the capital market structure doesn't change, which I think is like a debatable assumption, in a decade, I think that you'll have consumers really with near instant, fully automated one touch experiences and you'll still have compliance and disclosure and a whole bunch of things that you have to, you have to kind of work through and kind of like opening a deposit account, you still have to sign like a million disclosures to get a deposit account, even though it's like a one session experience. But I certainly think that the fulfillment will be automated. It'll all be source data and the execution will be much better on the Backend.

B (48:07)

Let's end on vision for a utopic future. Fast forward X number of years. It's a more modern, more efficient, more transparent. Pick your favorite adjective. What does that look like? And what are one or two things that are needed to get there?

C (48:23)

Two angles. I think that two interesting things that I think about are capital markets. And then I think about what the experience is going to be for a consumer. And for the consumer I think it's more dystopian. The capital market side I think is more just like interesting to think about. Right. This is a very interesting, very niche but important topic. But one of the great inventions of basically the US mortgage market is this 30 year fixed rate mortgage that every big financial sovereign wealth fund institution, whatever else, buys from American homeowners and basically earns extremely low spread and returns on relative to the risk that they're taking on in terms of prepayments and whatever else. Right? Because of this amazing sort of liquid ecosystem that they've all built. And so as things get more efficient, fundamentally what ends up happening is that people should immediately refinance. And so basically rates keep dropping and so this instrument becomes less and less like valuable. Right. Today the way it works is there's a little bit of a spread above treasuries. You basically buy it because some people won't refinance when rates go down. So that's sort of like why there's a excess return and there's like a call option built into it. And so all this financial mathematics works because there's truly inefficiencies built into the system. When the inefficiencies disappear, that instrument looks very, very different. And so there's like a reckoning at some point in time in the future that will happen because of that. That's more like an interesting like financial mathematics exercise that financial geeks will geek out on. The second thing is this again more consumer oriented view, which is kind of dystopian. And I just like personally have started to believe that this will happen. If you think LLMs are just going to be able to eat up more and more tokens and have bigger context windows, whatever else, I think what's going to end up happening is that a mixture between neuralink or people having, if you've ever watched a movie circle, there's these things that follow around, record you everything. Basically the idea is, look at the end of the day today, even today, I have so much information about consumers that I can embody and really encapsulate all their information. A little Bit stale, but in the form of like tokens that can be fed to any originator. Right. Like, I have all the context about their loan documents, I have their contacts, about their work history, all that type of stuff, and also their pay history. And if you get more and more of that and that gets constantly fed into a machine, you can basically supply any super app or any sort of, like, customer. All of this embedded context about that individual, which means that they can take that immediately, get a new loan, you know exactly what their mindset is, exactly what's going on, and everything becomes instantaneous. It's more dystopian than it is, like, utopian. But I just think it's going to end up going there because if you look at the sequential steps that take place, that's what's going to end up happening. It's kind of cool to think about, kind of scary to think about at the same time. But, you know, it is what it is.

B (50:55)

It's like push button get mortgage. But you don't even need to push the button.

C (50:58)

Yes, exactly.

B (51:00)

Tim, what about you?

A (51:01)

Well, I don't want to promote a dystopian view of the world. It may come to pass as Andrew suggests, but I think there'll be a lot of good that comes out of it as well, and there'll clearly be a lot of change. One thing that I think is necessary in order to kind of drive all this forward is because the government and the GSEs play such a big role in this system. Policymakers, regulators, the management of the GSEs need to be involved in this set of changes. And those conditions don't really exist today to promote this kind of innovation. Fannie Mae and Freddie Mac have been in conservatorship, government conservatorship, for the last 17 years. When I joined Fannie Mae, I was told that the board of directors had been assured by the highest levels of the United States government that the future of Fannie and Freddie would be recharted in 12, at most, 18 months. That conversation was in December 2008. Okay, it's now 2025 and they're still in conservatorship. And in addition, in that time, regulations have been passed that prohibit them from engaging in any new products or any new activities without an elaborate regulatory approval process, which includes public comment from all the entrenched vested interests of the current mortgage industry. This is all set up to slow innovation. This is all set up to keep the existing power players in their current positions. And if what you really want to do is to drive all this innovation with the ultimate goal of better protecting the taxpayer, making the system safer and better driving down costs for consumers. That set of conditions is not helpful to that. So what I would say is the conservator and regulator of the GSEs, the Federal Housing Finance Agency, ought to be seriously thinking about not only how to make it easier for innovation to happen, but actually mandating that this innovation happen, even if it means that certain current entrenched interests get disintermediated or their role is reduced. Because there's no path to innovation where everybody who currently makes money makes the same amount of money and nobody else makes any money. Innovation comes with disruption, and if you want the innovation, you have to be willing to and actually promote the disruption. And our current regulatory structure, especially conservatorship and the current regulatory regime discourages all of that. So I think we're finally at a point where, given all the financial pressures on the mortgage industry and all the technological innovations and really the brilliance of entrepreneurs like Andrew and Mike to actually drive this change, but we need to make sure that the policy environment, including the regulatory environment, adapts to it as well.

B (53:50)

That is just the absolute perfect note to end on, and extremely well said. Thanks, Tim. We could have done a whole episode just on that.

D (53:59)

Thanks for listening to the A16Z podcast. If you enjoyed the episode, let us know by leaving a review@ratethispodcast.com a16z we've got more great conversations coming your way. See you next time. As a reminder, the content here is for informational purposes only, should not be taken as legal, business, tax or investment advice, or be used to evaluate any investment or security, and is not directed at any investors or potential investors in any A16Z fund. Please note that A16Z and its affiliates may also maintain investments in the companies discussed in this podcast. For more details, including a link to our investments, please see a16z.com disclosures.