00:12From Chicago with HousingWire, I'm Alison LaForgia, and today I am sitting with the
00:17senior director from JazzX, Mike Romano. Hi, Alison. Hi, Mike. How are you? I'm doing well.
00:24I'm excited to talk to you today. Absolutely. Thanks for having me. I'm excited to dive into
00:27a hot topic, which is AI in the mortgage industry. It is all we're talking about right now.
00:34I'm really curious, what is the biggest misconception mortgage lenders have about AI today?
00:40I think that the biggest misconception is that all AI is the same, when there's actually
00:47really been a distinct evolution already within AI because the space is moving so quickly.
00:52Right. But the assumption that, you know, plugging AI into a piece of the workflow or, you know,
01:00as part of the process that it's going to deliver equivalent results, regardless of what kind
01:04of, you know, the AI technical capabilities are. Mike, you set me up for my next question,
01:10which really is diving into the differences because all AI is not alike and there are very
01:17different things that mean different things. So what is the difference between generative AI,
01:21agentic AI, and EGI? And why should lenders care? Why should they care? That's a good question. I mean,
01:27we were all introduced to AI really not that long ago through generative AI. And generative AI is your,
01:33is, you know, ChatGPT, for example, is the example, right? It's very good at, you know,
01:37summarizing information for you or performing a simple task, but you're still, as the user,
01:42giving it that context of what to do, it's not going to, you know, independently act on its own
01:47or, or think of an answer or, or perform a task, right? And so now what a lot of people
01:51are starting
01:52to talk about is, is this, the second layer or second level of agentic AI. So agentic AI can
01:57absolutely like perform a task, like a waterfall set of, of rules to follow, but it's, you know,
02:03it typically lives within pretty narrow guardrails. And those guardrails become quite brittle,
02:08actually, if something that's outside of the expected path is introduced.
02:12And nothing ever happens in mortgage. Right. And so we see that all the time in mortgage,
02:17of course, right, is, you know, you might be heading down the path of a conventional loan,
02:20but then a document comes in and it changes, it changes the whole loan scenario or the pricing,
02:25right? And, and, you know, an agentic model might not be able to handle that because we know in
02:30mortgage, very few things follow the simple path, right? No, we got to keep us on our toes.
02:36So that's what we've introduced, or well, EGI, right? So that's enterprise general intelligence.
02:41The idea is that, that this next evolution of AI, this is where you can actually have co-workers
02:47that, that actually think and, and reason across an entire loan and understand the entire scenario
02:53across the whole, the whole loan, right? So including things like appraisal or title,
02:59not just, you know, income and assets, but to be able to think and reason and collaborate
03:03across the other assistance to actually perform meaningful amounts of the person's work. So again,
03:10not just following a rule-based task, but to then ingest all of the data that comes with the loan,
03:17all the documents and reason across the whole thing and make decisions from there.
03:21You mentioned two frontier models, which brings me to ask why many lenders who are potentially
03:30interested in building their own AI maybe should or shouldn't. How should they approach AI customization
03:37without taking on unnecessary risk? Well, that's the key, right? Is that risk? If you're a lender,
03:43typically you're, you're quite good at, right, of, of originating loans and managing operations and trying to
03:48get as much as you can out of your people and, and, and building and scaling your team. You know,
03:54introducing AI on, on top of the regulatory risks that we already have. Now you're asking, you know,
03:58now you're, now you're forcing to, you're trying to build a tool that, you know, can not only help
04:02your, your teams perform, but still manage all that risk. And that's, you know, it's quite difficult
04:08to do for the average lender, I would think. It is. And I think you have to really have in
04:14-depth
04:14knowledge on both sides of it. And that's a really hard formula to answer. Right. To build an AI model
04:21that can keep up with all those changes and, and, and, and the governance of it, and then still
04:27perform those tasks as those, those tasks in need are constantly changing too. Um, that's a really
04:32difficult task for, I would say for most lenders. Now we talked about how there's different types of AI
04:37and there's lots of different newcomers in the technology space to the mortgage industry.
04:44How can lenders tell the difference between basic automation and AI that can actually reason
04:49through mortgage guidelines? I think what you need to do is, is maybe just test or ask some
04:55questions of the vendors, right? I mean, we, you know, we know there are, you know, the, the conventional
05:00Fannie Freddie guides or, or FHA, but then, you know, ask in a scenario that where something changes,
05:06where something comes in and it, and it doesn't, or I need to reprice it at a, uh, as a
05:10non-QM,
05:11like how does the system handle that? How quickly, if my needs as a lender, I want to take on
05:15a new
05:16investor because I'm, I'm doing some non-QM, how quickly can the system adapt and allow me,
05:21you know, uh, to, to introduce those new loan products. Right. And so, um, typically agentic model,
05:28that's a lot more challenging, right? Cause again, they're built to handle specific scenarios.
05:32Um, so you need to be able to ask, you know, what's it reasoning across? Are we including,
05:37you know, what risk categories are included in your analysis? Because my underwriter needs to
05:41look at a bunch of different things and understand how that all relates to the loan. And that's really
05:46the, the whole loan reasoning part, right? Like, you know, an underwriter is not only going to look
05:50at income assets and employment, or, uh, they're, they're going to consider the appraisal and title
05:55and property tax and all that stuff plays together in the approval of that loan. Right. And so you
06:03need to be able to ask those models, what happens when the appraisal comes in low and that changes
06:06the LTV, how does the system respond? Right. So, uh, a model that can reason will understand the
06:13downstream effects when something else comes in and when something changes it. So, you know, or, uh,
06:18let's say a bank statement comes in and on that bank statement, you know, that could materially
06:21change what the LTV is, right? Because of the down payment, how does the system respond to that?
06:27Uh, you know, is it flexible? Will it understand the downstream impact of that loan? Or is it going
06:31to check a box? Yup. You know, you, you introduced a bank statement and that clears that condition.
06:35Does it actually read what's on that statement and understand it and how it applies to the loan?
06:40Now you gave a couple of very specific use cases in that answer, and it talks about how much
06:47potential there is for AI to take on more and more mortgage tasks. Where does human oversight
06:54remain essential, especially in, to your point, underwriting and compliance?
06:58Well, absolutely. So, I mean, you know, as of at least today, a human has to touch it and ultimately
07:05make that decision on the underwrite, right? The decision can't be left entirely to, uh, an AI
07:11assistant or AI coworker. So a human most definitely is involved in making that final decision, but a
07:17human should also be involved in overseeing how these assistants are continuing to perform,
07:22right? Like as they're continuing to learn your business and, and the guidelines and policies that
07:28you have in place, like, you know, are you watching for, you know, any hallucinations or anything that
07:35that changes in the wrong direction, right? You want to make sure that you've got a human involved
07:40that's going to be able to say, you know, to, to make sure things are still pointing in the right
07:43direction. And, uh, an underwriter still need to ultimately needs to, to make that final decision.
07:48They're there. The, the definition of their role might be changing a little bit,
07:53right? But they're still ultimately involved in making that decision.
07:56Now, looking into your crystal ball into the future, how do you see AI changing the way that
08:03mortgage teams work and how should lenders be approaching that future and preparing for it
08:12now? Yeah. So there's absolutely no reason to fear AI that's coming. I think it's important to
08:18understand that, you know, we should all learn to work with it. The, what we're finding in deployments
08:24is, you know, the, typically the job, the job, or the job role tasks change faster than the job title
08:31changes. So like, you know, now instead of, you asked about the future, you know, instead of a,
08:36you know, a processor or underwriter managing, you know, 20 to 40 loans in their pipeline at any given,
08:41at any time of the month, you know, you might be now overseeing a team of assistants that is managing
08:484x that, right? But it's, it's still, so, you know, your role as a processor or a lead processor or
08:53underwriter is, is more in managing those teams of assistants instead of just only focusing on your
08:58day-to-day tasks that are, you know, that, that are, that are occurring on a monthly basis. It's,
09:03it's being able to work with the AI that's coming, uh, and I think it'll make us a lot more
09:09productive.
09:10Mike. Yes. Thank you for sitting down with me. It was my pleasure, Alison. I appreciate you giving
09:16our audience guidance about AI that actually moves the needle, the difference between different types
09:21of AI and what ChazX AI is up to. Great. I appreciate your time. I can't wait to see what's
09:25next. Okay.
