Monday, July 6, 2026
7.6.26 Housing Supply; Sei’s Pranay Shetty on AI Agents; Dog Days of Summer
0:00
0:00
Listen on
YouTube
Today’s episode includes a look at the supply of housing currently on the market. Plus, Robbie interviews Sei’s Pranay Shetty on how purpose-built agents that handle borrower calls, process loan documents, do contact center QA, and monitor compliance are driving mortgage operations forward. And we close with the expectations for the mortgage industry as we enter the dog days of summer.
Thank you to FICO. In a market where every loan counts, FICO® Score 10T lets lenders say yes to more borrowers without added risk. As the industry's most predictive credit score, FICO Score 10T combines proven performance with deeper insight into borrower behavior to help support a stronger and more resilient housing finance system. FICO has set the standard for decades, and we're grateful for their support of conversations that bring mortgage professionals together.
The Chrisman Commentary is your go-to daily mortgage news podcast, where industry insights meet expert analysis. Hosted by Robbie Chrisman, this podcast delivers the latest updates on mortgage rates, capital markets, and the forces shaping the housing finance landscape. Whether you're a seasoned professional or just looking to stay informed, you'll get clear, concise breakdowns of market trends and economic shifts that impact the mortgage world.
Full Transcript
Show full episode transcript
(0:02) Welcome to the Chrisman Commentary Daily Mortgage News Podcast. (0:06) I'm your host, Robby Chrisman. (0:08) Topics on today's episode include the supply of new homes, my interview with Sei's Pranay (0:14) Shetty on how purpose-built agents that handle borrower calls, process loan documents, do (0:19) contact center QA and monitor compliance or driving mortgage operations forward, and (0:25) have we reached the dog days of summer? (0:27) It seems that way in the capital markets.
(0:32) In a market where every loan counts, FICO Score 10-T lets lenders say yes to more borrowers (0:38) without added risk. (0:39) As the industry's most predictive credit score, FICO Score 10-T combines proven performance (0:44) with deeper insight into borrower behavior to help support a stronger and more resilient (0:48) housing finance system. (0:50) FICO has set the standard for decades, and I'm grateful for their support of today's (0:54) podcast, and the conversations that help bring mortgage professionals together.
(1:00) There are many reasons why renters aren't owners, like can they not afford a residence (1:05) or are none available? (1:07) We recently learned that May new home sales slid to 580,000, seasonally adjusted, and are (1:13) down around 7% month over month and year over year. (1:17) Sales of completed homes have declined for three straight months and are down over 30% (1:21) since November. (1:21) The supply of new homes for sale, 496,000, is now over 10 months, one of the worst in (1:28) nearly 20 years.
(1:30) Median new home prices continue declining, and all the while most will say that no one (1:35) seems to have any idea what our current national housing policy is. (1:38) We've latched onto affordability, but that has not been put into a policy. (1:43) We've seen plenty of trial balloons about selling federal land, doing away with tri-merge, (1:47) 40 or 50 year mortgages, and assumable mortgages, but nothing has stuck so far.
(1:53) There are small steps, like MISMO updating its mortgage insurance data standards to (1:57) support VantageScore 4.0 and FICO 10T, helping firms prepare for new credit models. (2:02) The dog days of summer are traditionally associated with the rising of the star Sirius, which (2:08) was believed to bring heat, drought, and thunderstorms. (2:11) During those conditions, people don't do much, and certainly the rising of a star doesn't (2:15) move mortgage rates.
(2:16) But mortgage-backed securities and U.S. Treasuries ended the holiday short and weak last week, (2:21) little changed but lower overall, as weaker-than-expected June payroll growth at $57,000 for the month (2:26) and significant downward revisions to prior months were offset by an unexpected decline (2:31) in the unemployment rate, which fell to 4.2%, driven by lower labor force participation. (2:37) Employers remained cautious due to ongoing inflation concerns and consumer pessimism, (2:41) with the largest job losses occurring in the leisure and hospitality sectors. (2:46) The mixed employment report supported shorter-term Treasuries and the steepening of the yield (2:51) curve modestly.
(2:52) But things are active on the issuance front. (2:55) June marked the fourth consecutive month of agency mortgage-backed securities issuance (2:59) exceeding $100 billion, with gross issuance reaching $118.5 billion, the strongest June (3:05) since 2022, as overall mortgage production remained healthy despite a modest monthly (3:09) decline. (3:10) While refinancing activity continued to support issuance, its share of total production has (3:15) fallen meaningfully since the first quarter due to relatively stable mortgage rates, leaving (3:20) purchase originations increasingly important to sustaining supply.
(3:26) Conventional and Ginnie Mae issuance remained concentrated in 5.0% and 5.5% coupons. (3:32) Loan production continued to grow year-over-year, and with refinancing expected to remain subdued, (3:37) future issuance above the $100 billion threshold will likely depend on the strength of the (3:42) purchase market. (3:46) For today's interview, I wanted to welcome to the show Sej Pranesh Chetty to talk about (3:50) how purpose-built agents that Handlebar recalls process loan documents to contact center QA (3:56) and monitor compliance are driving mortgage operations forward.
(4:00) He's co-founder and CEO of Sej AI, which is an operations partner for mortgage and (4:05) banking. (4:06) He's worked across investment banking, private equity, and fintech, but now focuses on Sej's (4:12) AI operations platform for mortgage lenders, servicers, and banks, helps Handlebar recalls (4:17) process loan documents, and monitor compliance. (4:20) I think people now have a much better understanding of what AI agents are capable of or are doing, (4:26) but when it comes to mortgage, you've obviously been intimately involved with this.
(4:30) Thoughts on where we are seeing AI agents deployed? (4:34) Maybe people don't even realize they're being deployed, but where are we seeing them deployed? (4:38) And then the pie in the sky capabilities for where we're headed with them. (4:43) Yes. (4:43) What's really interesting is if you talk to everyone, not just in mortgages, broader financial (4:48) services, right? (4:49) Like every leader's saying, we're an AI-first company, we're using AI.
(4:53) You dig a little deeper. (4:55) The vast majority of them are using like a chat GPT enterprise or Microsoft co-pilot, (5:03) right? (5:03) So I think we're still really, really early in like actual AI adoption. (5:10) Outside of co-pilot and note takers for meetings, I think we're seeing some really interesting (5:16) applications that are already live with customers, right? (5:19) So one is voice AI, right? (5:22) Both on the origination and servicing side.
(5:24) So origination, the obvious use cases are AI calling leads, doing qualification, and (5:31) then handing those over to an LO. (5:34) Another use case we're seeing is around re-underwriting, right? (5:39) So looking at documents as they come in, basically, can you give every loan officer an expert (5:46) underwriter so that all the checks can be done right at the start of the process, right? (5:51) With the ultimate goal of how do you reduce clear time. (5:54) What you're seeing is a lot of companies just become, the companies that are adopting AI (5:58) become a lot more efficient.
(5:59) And we're seeing that in our sort of customers as well, right? (6:03) They're able to move much faster. (6:05) Things that would have taken four hours, like reviewing a loan file that would have taken (6:09) four hours is now being done in 45 minutes. (6:13) Maybe now is a good time to talk about what's happening at your company.
(6:18) You focus on voice agents, document intelligence, compliance, and maybe let's start there. (6:24) Can you explain the company to people and then we can dive into some of the specifics. (6:29) Yeah, so both me and my co-founder come from a background in financial services, right? (6:33) So I've worked in banking, private equity, fintech, spent some time in real estate as (6:39) well.
(6:40) My co-founder has been a software developer at large companies like Amazon, PayPal. (6:45) So really built core payment systems for these companies. (6:49) And playing around with chat GPT like two plus years ago, it was a big moment for us (6:56) because we could see that so much of the work that we had done, like I still remember in (7:01) my first year as an investment banking analyst, sitting up at 3am updating spreadsheets, financial (7:08) models, updating PowerPoints, and I could see that that work no longer needed to be (7:12) done.
(7:13) And so our thesis when we started the company was that there is so much work in the back, (7:18) middle, and front office in financial services that is just very manual, very repetitive that (7:25) is either done by hiring a ton of people in-house or you outsource it. (7:29) And so what we're seeing across the few of the use cases that we're working with customers (7:35) on is they're now, you know, and I go into the use cases, right? (7:39) So one is around sort of voice AI. (7:41) Like I mentioned, most obvious use case is calling leads, right? (7:46) The idea here is that, you know, I think this is CFPB data from a few years ago, but about (7:52) 60% of borrowers go with the first mortgage lender they speak to.
(7:58) And so if you're not calling that lead immediately, you're losing their business, right? (8:04) So we did an experiment a few months ago where we had our AI call, I think it was about 300 (8:10) lenders across the country. (8:12) And this was both during and after office hours, right? (8:15) And after office hours, about 70% of them, the calls just went unanswered. (8:20) There was no callback, even when we left voicemails.
(8:23) And so it's shocking how a lot of people are basically just losing business because they're (8:28) not there where the customers or the borrowers really need them. (8:32) And so the way we're seeing AI is, again, AI is not going to like replace a loan officer. (8:37) What it is going to do is allow the LO to spend time on the more qualified leads.
(8:45) It's going to do that drudge work of calling the leads, doing the qualifications so that (8:49) they're not spending time just calling 100, 200 people. (8:53) Something similar on the pre-underwriting AI agent that we've got live as well. (8:58) So a processor may be looking at 1,000, 1,500 pages, a processor and an underwriter may (9:06) be looking at 1,500 pages for every single loan file, comparing, extracting information, (9:12) looking at agency guidelines, investor overlays, all of that takes a lot of time.
(9:18) It's not a great experience for borrowers because there's a lot of packets. (9:22) I'll give you one example from a customer of ours, their team, when the loan file would (9:27) hit their desk, would spend about four hours reviewing everything, right? (9:32) So 1,000, 1,500 pages, 500 rules and conditions. (9:37) And so it's a meaningful improvement in efficiency for them.
(9:41) Right. (9:41) So, Pranay, you cut out, AI is able to automate how much? (9:45) About 80%, right? (9:47) So if they're looking at 500 conditions for every single loan, the AI today is able to (9:52) do 400 of those. (9:53) So when they would have spent four hours pre-AI doing their work, it's now down to just one (10:00) hour.
(10:01) So think about what that means for like how quickly you can move loans through your pipeline (10:07) and where your people can actually spend meaningful amount of time and effort on. (10:11) Can you talk a little bit about unifying front office and back office operations? (10:19) I think for a long time they were treated as separate, but now we're seeing them become (10:23) more blended. (10:24) Yes.
(10:25) I'll start off with just like an example, right? (10:27) And I can go into more details. (10:28) But one of the things that we're, because we have both sort of like voice chat agents (10:33) and the pre-underwriting agent, we've been able to combine them, right? (10:37) So think about a simple scenario where, you know, you're a borrower, you send in your (10:41) bank statement. (10:43) Today, a processor may look at it, realize that, you know, it's missing one month.
(10:48) There is a large deposit and there's no LOE. (10:51) They may look at it that day. (10:53) They'll probably look at it a couple of days later because they've got other work to do.
(10:57) And you just, when they do look at it, they'll call you. (11:01) You may be traveling. (11:02) You take another couple of days to get back.
(11:04) Now, four or five days have gone by and we haven't really moved further along in the (11:08) process, right? (11:09) But imagine a situation in which you send in the documents, oh, you know, in this case, (11:15) the bank statement, the AI analyzes it. (11:18) It's immediately within two minutes able to figure out what the missing information is. (11:25) It can call you or send you a text.
(11:28) All of this happens within two to three minutes. (11:30) You're probably still at your computer when you get that call and you can immediately (11:34) send the corrected documents. (11:36) And so within five minutes, we're probably further along in the process than would have (11:42) been, you know, pre-AI where it would have taken a day to five days.
(11:47) And this is something that, you know, is not in the future. (11:50) It's not something we're expecting to happen. (11:51) It's already happening today.
(11:53) And I think that's a really good example of, you know, combining front and back office. (11:58) What I think a lot of people, you know, more broadly with AI, a lot of people think when (12:03) they think about AI, they think that the technology is the challenge. (12:07) I think what we're seeing from like all of our deployments with like large enterprises (12:11) is that usually the change management is where things get delayed.
(12:16) Things can potentially fail. (12:18) And so with AI, you have a great opportunity to actually just rethink the workflows that (12:25) they currently are. (12:25) Just because things were done a certain way, because you had to have people doing it, doesn't (12:30) mean that with AI augmenting those people, they need to be done the same way.
(12:35) There is a lot of competition out there in the AI space in mortgage. (12:41) I think lenders and originators now are going, there's all these AI companies out there. (12:45) How do I decide between one and the other? (12:47) Thoughts on best practices for assessing counterparties and making decisions as a lender or originator? (12:54) Yes.
(12:55) So I think, yes, there is a lot of competition. (12:58) I also think there's a lot of noise. (13:00) I think at this point, everyone is saying they have AI, whether that's true or not.
(13:06) And a couple of ways I would like recommend everyone sort of as they're diving deeper (13:11) into this. (13:12) One is actually see, you know, is their product really live with customers? (13:17) It's never been easier to make an AI demo. (13:20) Rob, you can tell me, pick any product you want in any industry, and I can guarantee you, (13:26) I can get you a working demo in like 24 hours.
(13:28) It'll look very impressive. (13:31) It'll look super impressive, right? (13:32) But it will obviously not work. (13:34) And so the weird thing about AI is it's so easy to make a demo, and yet actually making (13:40) stuff work in production at scale is really, really hard.
(13:44) I love giving this example, because it's just so easy to verify. (13:47) But go to Amazon.com. (13:48) While you're listening to this, they have a chatbot called Rufus. (13:52) I was going to buy a guitar a couple of months ago, and I was just asking it for recommendations.
(13:56) And it started obviously giving links to Amazon. (13:59) Amazon probably doesn't care because we're stuck with them. (14:02) But for anyone else, it's a disaster, right? (14:04) Can you imagine, you know, if your AI lead calling agent calls someone, starts talking (14:11) to them about your company, and then the borrower just says, hey, where else can I get a better (14:14) rate? (14:15) And if the AI starts mentioning competitors, it's a disaster.
(14:18) And so I think the first thing I would recommend is, you know, make sure whoever you're speaking (14:23) to, for whatever AI agent they're talking about, right? (14:26) Is it actually live in production with lenders? (14:31) What is the feedback been? (14:33) Number one. (14:33) Number two, I think like industry experience is really important. (14:36) Again, going back to my point of like, never been easier to make a demo.
(14:39) You'll see a lot of people just say, hey, we can make AI industry agnostic. (14:43) It doesn't really matter. (14:45) Which again, it sounds really impressive, it'll work in like 30-40% of the cases.
(14:50) But like when you start looking at like edge cases come up, that's when it fails. (14:53) Because if you don't have that industry experience, if you're not focused on an industry, you're (14:57) not going to be able to spend time solving for all of those. (15:01) So those are the two things I would like highly recommend anyone look at, like make sure the (15:05) vendors you're evaluating actually have products that are live in production.
(15:09) And do they actually have industry experience and use cases that are live in production (15:14) are with customers that are similar to you. (15:16) I want to talk about security. (15:19) I know it's very important to you.
(15:21) I believe your platform just got PCI DSSL 1 certified. (15:26) I don't think that confidentiality, integrity, availability of information is being discussed (15:32) enough in this AI age. (15:36) Data created, maintained, posted, all this stuff is really important.
(15:41) I'm hoping you can speak to just why it's so important, obviously it's important to (15:45) you, why it's so important, and then why you feel like other companies are failing to think (15:48) about this. (15:49) There's a tendency for like a lot of tech founders to come into an industry and say, (15:54) hey, we just know better than someone who's been in the industry for like 20 years. (15:58) Because why wouldn't we? (16:00) And just say that, hey, we can move fast and break things.
(16:03) To be fair, I think, you know, a lot of founders now and a lot of executives at tech companies (16:09) have gotten really smart about this. (16:11) I've heard horror stories from like, this was a lender I spoke to about four or five (16:17) months back. (16:19) They told me that they realized that their LO, so they didn't have any AI tools, like (16:24) no enterprise chat, GPT or Microsoft Copilot.
(16:27) And they've realized that their loan officers were just uploading customer bank statement (16:33) and customer documents into like the free version of ChatGPT and asking questions and (16:38) trying to like use that to help them in their day to day. (16:41) And so this was true in like 2025, right? (16:43) I think a lot of companies hadn't thought through like a risk policy. (16:47) That's definitely changed this year.
(16:49) A lot of companies have got, a lot of lenders have gotten smart about this. (16:53) The way we think about this, I think like a couple of like non-negotiables, right? (16:57) One is none of the data that is used, that is, you know, sent to the model should be (17:04) used for training. (17:05) And what I mean is it shouldn't be obviously used for training the underlying model, but (17:09) it shouldn't even be used by the company to train for like competitors of yours.
(17:14) And so you need to have those like sandbox environments for every single company. (17:19) And so that is one non-negotiable that, you know, everyone as they're looking at vendors, (17:24) not just for like, not just in mortgage, not just for lending, sort of underwriting (17:28) or voice for any AI use case you're thinking of, even if it's like, you know, a document, (17:33) internal document review use case. (17:36) This is, you know, you need to make sure that the models that are being used, they have (17:41) explicit guarantees that they're not being used to train on your data, competitors models (17:47) are not being used, your data is not being used to train competitors models.
(17:50) And then I think besides this, there's a lot of the standard security, which, you know, (17:55) even outside of AI, a lot of companies would have experienced with, you know, like data (17:59) residency, localization, those sort of things. (18:03) The whole agentic AI movement seems so futuristic, but we are here. (18:08) And so I'm wondering how you see AI agents improving from where we are in 2026.
(18:13) How much better can they get? (18:14) Where are we going? (18:15) And you can keep it as it pertains to mortgage banking. (18:18) Yeah. (18:19) So I think I'd say one just like high level across industries, right? (18:24) I think for a lot of people who spend, you know, days and nights on Twitter, like me, (18:31) it feels like AI is just like seeped into every single industry and like everyone is (18:36) using AI all the time.
(18:37) Do you actually go out and like speak to people in the real world, it's just not the truth (18:42) at all. (18:42) Right. (18:43) So I think there is, we are playing this catch up where I think like the technology is much (18:48) further along than like actual adoption and the technology will keep getting better.
(18:53) Like, you know, the easiest example that I think a lot of people would have seen is just (18:59) like the quality of like voice AI agents. (19:02) Right. (19:02) If you had heard this a year ago versus today, it's just completely different.
(19:08) And go further back and it sounds like a shitty IVR. (19:12) And so just in the space of like a year and a half, like the quality improvements in the (19:18) models have been tremendous. (19:20) We're seeing same things happen on like documents and underwriting.
(19:25) So across different use cases, you're seeing that the models get really better. (19:29) But as far as I'm concerned and where I think what's most exciting to me isn't really that (19:35) hey, the models are getting better. (19:36) I think it's just that even where the models were a year ago, like where we've not really (19:42) applied that quality of model to the real world yet.
(19:44) Right. (19:45) It's not helped us reduce the cost of origination. (19:48) It's not helped us reduce the amount of stare and compare work that like LOs and underwriters (19:53) and processors are doing.
(19:55) And so we're in this weird state where, yes, like AI is very futuristic and there is like (20:01) the tech is really exciting, but it's still shocking how little we have applied this to (20:09) actual use cases. (20:11) And I think that's changing very fast. (20:12) Like if we were having this conversation six months ago, I would have said like use case (20:17) are very minimal.
(20:17) Like we're now seeing like we're working with banks, we're working with like top 25 INBs. (20:21) And so there is definitely now a sense from the more forward thinking senior leaders of (20:27) this company is that, hey, we need to move fast because eventually everyone will catch (20:33) up. (20:33) But if we get like a 12, 18, 24 month advantage, we need to do whatever we can to sort of grab (20:40) that.
(20:41) Can we talk a little bit about how you're seeing your product deployed and just where (20:49) people are seeing ROI and maybe, you know, kind of like verbal case studies or success (20:53) stories here? (20:55) Yeah. (20:55) So there are a few use cases that we're seeing ROI immediately, right? (21:00) So one is around voice, both on the origination and servicing side. (21:04) I'll talk about origination to start, right? (21:08) We're seeing voice AI being used both for managing inbound calls and for like outbound calls (21:14) to call leads.
(21:15) In the case of one large originator, we were able to 4x the number of appointments that (21:27) were being set with their loan officers to the extent where like their loan officers (21:31) would be joking and saying, hey, we just have to block hours in the calendar because now (21:34) we're getting so many leads that we weren't getting before. (21:37) Right. (21:37) So that is a very easy to see example of like AI just having a direct real world impact (21:44) where, you know, I'm getting more highly qualified leads that will then move further down the (21:49) funnel.
(21:49) So that is one use case we're seeing. (21:52) The other we're seeing is on the underwriting side, right? (21:56) And so before using us, this is a listed NASDAQ lender, they were, their team was spending, (22:03) you know, when the entire loan file hit their desk, they would spend about four hours reviewing (22:08) all of it, doing income calculations, comparing it against agency rules. (22:12) The AI is now able to automate 80% of that work.
(22:17) And so their team is able to spend time on the 20% that actually requires human judgment, (22:22) is complicated, cannot be automated away. (22:24) And instead of four hours, they're down to under an hour doing. (22:29) And finally, for people interested in more information, best next steps, send them in (22:33) the right direction, please.
(22:35) Yes. (22:35) So our website is Seiright, S-E-I-R-I-G-H-T dot com. (22:40) I am on LinkedIn.
(22:43) Yeah, those two best places to reach out to us. (22:45) Cool. (22:46) Yeah, I was going around the website and I thought it looked really good.
(22:49) A lot of the case studies look good. (22:51) It seems like a no brainer for a lot of companies out there. (22:53) So Pranay, I really appreciate the time, man.
(22:54) I wish you the best of luck and we'll talk again soon. (22:57) Thank you. (22:58) Thank you.
(23:46) She told me to take off her blouse. (23:49) So I did. (23:50) She told me to remove her skirt.
(23:51) So I did. (23:52) Next, she instructed me to take off her shoes, hose, bra and underwear. (23:56) And I did.
(23:58) Finally, she looked deeply into my eyes and said, don't let me catch you wearing my things (24:01) ever again. (24:07) Thanks again for FICO for sponsoring this week's podcast. (24:10) As the industry's most predictive credit score, FICO score 10T combines proven performance (24:15) with deeper insight into borrower behavior to help support a stronger and more resilient (24:19) housing finance system.