How Fast Will AI Bring Real ROI In Originations and Servicing?

The integration of artificial intelligence is fundamentally changing the mortgage industry, with a clear focus on demonstrating a tangible return on investment. As originators increasingly look to servicing to build a customer-for-life strategy, the session explores the pace at which AI is moving from experimental use cases to proven, scalable solutions that deliver real value across both functions. The discussion will cover the current state of AI adoption, from automating routine processes like document reverification and underwriting to using generative AI for customer-facing tools and internal efficiency gains. The session will also examine how innovative platforms are leading to measurable cost savings and increased application volume.

Transcription: 
Transcripts are generated using a combination of speech recognition software and human transcribers, and may contain errors. Please check the corresponding audio for the authoritative record. 

Bailey Reutzel (00:08):
All right. Hey everyone, we are back. Thanks for the packed house. Very nice right after lunch; love that. I feel like every session that I've heard so far has talked about AI, so they're really stealing our thunder here. We were just talking about the Gaslamp District, so maybe we could talk about that if you guys are tired of AI—way more fun. But these are our specific sessions. We've got two in a row about AI, so let's sort of jump into it here. If you don't know me, I'm Bailey Reutzel. I'm the senior director of Live Media for National Mortgage News and American Banker. We are talking today about how fast AI will bring real ROI in originations and servicing. My panel here: I have Jay Promisco, the president of Sierra Pacific Mortgage; Dawn Svedberg, executive manager of Lending and FinTech at CI&T; and Mark Hansen, executive VP of product at CMG Mortgage. All right, so we're doing a speed round. It's only one question, so it's really not a round. I just want to know when you guys think that AI will bring ROI to this industry. I want a number and then the category, like days, weeks, months, or years. Jay, we'll start with you.

Mark Hansen (01:24):
Six months ago.

Bailey Reutzel (01:25):
Minus six months ago. Great. Dawn.

Dawn Svedberg (01:27):
Minus a year ago. Okay.

Mark Hansen (01:29):
Yesterday. And counting.

Bailey Reutzel (01:31):
And counting, okay. Right. So it means you all should be adopting this technology. I kind of want to start setting the scene in terms of where AI has been integrated into the mortgage industry right now. I know you all have heard both from startups and industry providers, but just big picture, where is this having the most impact right now? Jay, I'll start with you.

Jay Promisco (01:59):
AI is an interesting term to use because there are probably 4,000 or 5,000 different ways it's being utilized—from generative to decisioning AI, from ChatGPT to Grok, to trying to figure out how to put a good LinkedIn profile together today in the mortgage world. It's being used in some really good ways and some really bad ways. I already offended everybody yesterday, so I'm going to try not to offend everybody today.

Bailey Reutzel (02:38):
No, I think we should still continue with that. It makes it spicy.

Jay Promisco (02:42):
Okay, I'll be offensive today. It's used in the manufacturing process. I'm seeing it used in the CRM space and with voice AI agents. I'm seeing it used for product guidelines in the mortgage space in a very good way. What I think the biggest impact is is some of the agentic AI that is being utilized at some of the firms.

Bailey Reutzel (03:05):
Now I'm interested in the bad use cases. I don't know if you want to talk about that a little bit.

Jay Promisco (03:11):
A bad use case—here's where I get offensive—is a vendor creating a single-spot solution for a loan officer or a company that can be replicated just by going on ChatGPT and doing it yourself. If there's one piece of advice I can give to everybody: go to prompt school—you can go online, it takes about two hours—and learn how to prompt. You could do a lot of this work yourself. The bad use case is paying for a service that's only fixing one thing.

Bailey Reutzel (03:43):
Fair enough. Dawn and Mark, I know both of you are working in this space and have some use cases that have already been implemented. Dawn, tell us where you're seeing the most impact for AI right now.

Dawn Svedberg (03:57):
From the CI&T side—CI&T is relatively new in the US in the mortgage and FinTech space. However, we're a 30-year-old company with 7,500 people out of Brazil, so we're nearshore. Two years ago we acquired Coreip and Inter-S, which were mortgage tech providers with folks that had built Mello and Sagent and those types of companies. Where we've had the most experience at CI&T is really in mainframe code transformation. We've done 2.9 billion lines of code for major banks around the world, transforming those. We use Gen AI—and have for the last four to five years since the advent of that—to modernize that code into a modern codebase. That's human-assisted AI. That is where we have seen massive transformation and huge ROI around the world with PepsiCo, Anheuser-Busch, and other companies. We're doing that now for lenders and banks in the US. But on the mortgage side, I'll transition to my dear friend here at CMG and some of the really innovative things they're doing in partnership with us.

Mark Hansen (05:27):
Sure, absolutely. Coming from more of a technical and product perspective, a lot of the demos you see are forward-facing products for loan officers, underwriters, or borrowers. That's great and important. But what we're also focused on is a completely other area, which is your internal software developers, product people, and business systems analysts. We're trying to empower both of those groups. How do you take a product person who understands mortgage and create requirements to hand off to your engineering team? In the past, you would meet with your loan officer and ask, "What are the 10 things that you hate?"

Mark Hansen (06:26):
We would design it, go to the UX team, they would build it, and three weeks later you'd have a requirements document to hand to engineering. They'd spend three or four sprints to build it. That is a thing of the past. Using AI, a product person can use tools like Bolt.new, Claude, or Gemini, and you can actually not only build working prototypes, but then have Claude extract the requirements documents from those prototypes. You can do this in a matter of hours. Come sit with me at my laptop, give me your brilliant idea, and in 15 minutes we'll have a working website of whatever you want. It's fascinating. But the real conundrum is that the artifact you are now giving to your software engineers doesn't need to be human-readable because the software engineers are also using AI like Claude.

Mark Hansen (07:26):
The instruction needs to be a program or set of instructions that their interface can then use to build those pieces of technology. What would take six or eight months is now taking a matter of days or weeks to get out of the system. All the tools you're seeing that have AI in them are great, but there's a lot of work behind them—how do you use AI to actually build those pieces? It's kind of like AI building AI. We're not AI fanatics; there's obviously a human in the middle and you need to test things, but you need to use the tools from the complete spectrum. You can't have a sacred cow if you have technology that may be obsolete. You have to be comfortable with that. Most importantly, you have to train it. How do you get a non-technical product person comfortable writing software that's going to be handed off to an engineer? It's a very big change in methodology.

Bailey Reutzel (08:33):
That's really interesting. How many people here are using this on the engineering side to do what he's talking about? Show of hands. Okay, so it's definitely not half, but there are some people using it. One of the things that I wanted to talk about is the "golden cow" situation regarding LOS systems.

Jay Promisco (09:05):
There's a reason why it's called a POS, by the way.

Bailey Reutzel (09:07):
Fair enough. How are you trying to integrate this AI into these ancient LOS systems? What are the hurdles and tips for folks trying to do this?

Jay Promisco (09:26):
I've been blessed to be on proprietary loan origination software. We have an advantage over a lot of other vendors sitting on dinosaur platforms. Integrating AI into our platform is a lot easier because it's my API and my code. If you're a mortgage banker working on one of the big box companies, you're beholden to what they will allow you to do. A lot of these companies over time have developed effectively the "Frankenstein" LOS or point of sale, and it's really hard to unwind all that garbage once it's in there—it looks like spaghetti.

Jay Promisco (10:19):
If you want to make a change, you find out somebody made a code change two years ago, that guy doesn't work there anymore, and nobody knows what the code is for. My advice is to look at the entire manufacturing process instead of just starting in one place. Every single component of the loan origination system can be changed upfront to affect the backend. For me, the most important thing we did was look at the manufacturing process first. If I can get my manufacturing cost down to zero, I can outpace and price out everybody else in the market. A lot of other people are focused on making loan officers happy; that's not what we did. We looked at the entire loan origination system as a holistic environment, making sure to integrate agentic AI throughout the entire process.

Jay Promisco (11:16):
The key for anybody to survive in this market is artificial intelligence creating elasticity in your fulfillment. We've all lived it: things get busy and I have to go try to hire a hundred people; things get slow and I have to lay off a hundred people. The golden goose is being able to use the same amount of people for 200 loans as you do for 2,000 loans. Just from an ROI perspective, I have a vendor that has done a phenomenal job for us. In the last six months, we went from about 82 people offshore doing work for us down to two. There's ROI immediately there.

Bailey Reutzel (12:08):
It's interesting because typically when I talk to people about implementing AI, they say "get specific"—build out one use case and go from there—and you're saying something different. I don't know, Dawn, if you wanted to add to that.

Dawn Svedberg (12:35):
There are so many products and different systems that have been put on top of the Frankenstein base system that is the LOS. There are PPEs, pricing engines, and CRMs bolted onto systems that are 30 to 50 years old. Changing that out is extremely difficult. We have some new entrants like the Vestas and others working hard to solve those core deficiencies and starting to make some headway. What we have found at CI&T is that our clients are basically saying, "No more; we're not bolting on anything else."

Dawn Svedberg (13:37):
The time is now. We're going to deal with what we have. We help lenders, including Mark here, make that work together so it's a seamless system that provides the functionality and efficiency. You get higher quality, faster speed, and greater efficiency for loan officers, processors, underwriters, and closers.

Bailey Reutzel (14:28):
Mark, are you on a proprietary system, or are you with one of the majors?

Mark Hansen (14:32):
It's semi-proprietary. We use Byte LOS. I get asked the question every day: why are we on Byte? Number one: could we build our own LOS? Sure. But do we want to spend two years to be where we are today and lose that opportunity? The value we get with Byte, unlike Encompass, is that we're the 800-pound gorilla. We're their biggest customer and we have the ear of the CEO over there. We get to dictate what happens in that LOS.

Mark Hansen (15:32):
We kind of get another company to build our LOS for us because half the stuff they're doing benefits us. That allows us to build on top. It is very malleable and has an API-first approach. Our long-term strategy is to slowly make the LOS less relevant. We've already done it with our loan officers. If you're a CMG loan officer, you typically don't go into Byte; you go into our proprietary Clear system. Now we have processing, underwriting, post-closing, QC, and secondary markets. We are using AI and other techniques to build out those screens and build them better.

Mark Hansen (16:31):
I don't need to bolt on to an LOS because an LOS might have six steps to go do something. We can just turn that into a custom screen for our people and have the LOS actually pop those windows instead. Before you know it, your LOS really becomes your system of record. You use it for compliance and for all of your integrations with third parties like doc vendors and credit. That's table stakes. We want to be focusing on the experience—making things better, more efficient, reducing costs, and most importantly, increasing quality. Several other panels mentioned quality is a big deal, so we want to make sure what we're building not only saves time, but is better.

Bailey Reutzel (17:19):
Having built a proprietary system, would you suggest that to other mortgage providers?

Jay Promisco (17:27):
No. Yes.

Bailey Reutzel (17:29):
All right.

Jay Promisco (17:29):
Let me rephrase this because here's my prediction—it's pretty bold, and I think I'm going to be right on this. In two years, you need a database and a native language model. Doing loans is that simple. You need the fastest, cleanest database you can find and a native language model surrounded with informational knowledge about your own company. That's it. The last time I said that in front of a room, the Encompass guys were there and I think they were trying to beat me up when I was done, but I'm dead serious. This stuff is moving so fast that starting to build your own LOS today will take seven years and 70 million bucks, and you'll still be unhappy. It's better to partner with an AI firm to build it the correct way. We're sitting on antiquated software that has been around forever. If you started a mortgage company today, you would build it for the future, not try to retrofit something that's been doing business for 20 years.

Bailey Reutzel (18:50):
Go ahead.

Dawn Svedberg (18:52):
One of our founders said to me the other day that in other verticals, clients don't even care about the tech stack anymore. They know they're going to swap it out every four years. It's so interesting that here we are so tied to this old system where only one person can be in the loan at a time. This stuff is insanity.

Dawn Svedberg (19:22):
Actually, the headwinds of regulation we've been facing for years is a good thing because it means the data has been regulated and it's clean and ready for AI. Financial services and banking are in a unique position because the data is perfect for AI. Focus on that—getting that ready—and then don't try to build the AI yourself. Just stand back, get the data ready, and then deal with the cultural inertia of your companies. Start to think about how you would do the whole process if you were starting from scratch. Then you'll be in a great position to take advantage of this tidal wave that will reshape everything.

Jay Promisco (20:29):
You want ROI? Using artificial intelligence, the hardest thing for your people to do is going to be laying off people once it's working. When you used to have 15 people doing a job and now you only need one, the hardest thing for everybody in this room is saying they only need that one person. That's back to the elasticity thing. Two years from now, you could probably run a billion-dollar mortgage company using 40 people.

Bailey Reutzel (21:07):
It's really interesting. I would love to talk about the future of work. This is the scary part for some people: figuring out what work looks like in the mortgage industry. The Nvidia CEO said the other day that AI is going to bring about a four-day work week. I'm very happy about that. Maybe we don't fire people; we just say you can only work three days. Mark, you guys have implemented quite a bit of AI already. How are you dealing with that culturally?

Mark Hansen (21:46):
The "out of the box" answer is that AI frees you up to do more important tasks. The challenge is that not all of your employees are going to be able to use AI. If you force people to use it and they don't understand it, you're going to get "AI slop." I'll use myself as an example. We did a thought experiment a couple of weeks ago. We are building a partner portal so your loan officer, borrower, and real estate agent can all go to one place to communicate. We have a 400-page business requirements document that we spent six months working on.

Mark Hansen (22:46):
We did that the old-fashioned way, and then we needed to create the prototype design. I run the product team and I'm the one telling people to use AI, so I did it myself. I took all the documentation and spent three days building a website. I'm not a designer, but it had everything it needed. Then I had my XD designer do the exact same thing. We presented both, and even though what I had was cool and better than the manual way, my designer created a beautiful, professional website that made what I had look like garbage. Why? Because AI was able to amplify his capability.

Mark Hansen (23:35):
It enhances your skillset. If you're not a good designer, AI is not going to make you a better one; it's just going to make you more sloppy, faster. The learning curve is about how you prompt it. You don't just say, "Build me this website." It's like saying, "Build me a planet," and you end up with Saturn instead of a planet close to the sun with oceans and breathable air. You have to know what to ask and do it in components, like building with Legos. Otherwise, if you ask in one sentence, you're going to get a monolithic pile of sand. The biggest struggle will be teaching qualified people how to use it. In our experience, it requires a "two-week rabbit hole"—you just have to play with it until you figure out the best practices and eventually plateau as an AI genius. How do you replicate that across your whole staff? That will be the challenge.

Bailey Reutzel (25:26):
Dawn, I'd love to hear from you as well about that cultural shift. I have drunk the tech Kool-Aid on AI in terms of it making me more productive. It's not going to take my job, but it does require me to keep striving to be that person. How are you feeding that narrative?

Dawn Svedberg (25:53):
This is new for everyone. No one has been doing this for 30 years. I loved it when Kim said she has "AI FOMO." Everyone does. You have to just jump in. One of the things we do with our clients is provide our "CI&T Flow," which is every AI engine at the latest version. ¹ We put it within the confines of their environment with their data and security and train every single person on it. We give them user IDs and train them how to use the latest models, because the models are all competing and changing every day.

Dawn Svedberg (26:53):
You can't just use Copilot at home; you have to be in touch with everything that's out there. You have to force it because there's going to be pushback, just like there was with the cloud or the internet in the beginning. But once you get that initial momentum, it will start. Another example: if we get an RFP, our founder said the other day it's much faster for us to just build four prototypes and ask which one they like than to respond to the RFP.

Bailey Reutzel (27:49):
That's really interesting. It's totally different. We've got a few more minutes. For the folks out here who haven't implemented AI yet, is there one thing they could play with right now that would help?

Jay Promisco (28:07):
There are two things.

Bailey Reutzel (28:08):
Great.

Jay Promisco (28:13):
If you're running a mortgage company, the goal is to have more loans coming in the door. I don't need technology, legal, or capital markets until I have a loan coming in the door. So, the first thing I would do is look at technology that would make your loan officers more effective. For example, we use Total Expert. They have exceptional AI tools that get loan officers more opportunities with customers to build your funnel. The second thing I would do before implementing AI is actually look at your process. I think the mistake people are making is deploying AI on a mortgage process that has been around for 30 years and is broken. You should reevaluate everything you do and ask "Why am I doing that?" before you deploy an AI solution that's just going to exacerbate the problem you had before. Focus on the top of the funnel and then whiteboard out your process.

Bailey Reutzel (29:36):
I think you all should do that every year with your company and your personal life. 43 seconds left. Dawn and Mark, what should they do?

Mark Hansen (29:55):
I'm not affiliated with these two tools, but I will mention them because they changed my life as a 25-year software veteran. Bolt.new—you go there and prompt it with what you want, and behind the scenes, it's Claude code and it will build you a website and all the code behind it. Alternatively, v0.dev by Vercel—same thing, but it's like Coke and Pepsi. Play with those tools, give them to your team, and they will do amazing things to get you on the track for AI.

Bailey Reutzel (30:37):
Amazing. Thank you so much. Thank you audience for being so attentive and engaged. Thank you very much.