# Truly AI Native Products | DC Fintech Week 2025

- Channel: [DC Fintech Week](https://streameth.org/dc-fintech-week)
- Date: 2025-11-21
- Watch: https://streameth.org/watch/yt-815jnk0gKQY
- YouTube: https://www.youtube.com/watch?v=815jnk0gKQY

## Transcript

All right, everyone. We could uh we're going to get going here. Thanks. We're going to get going here with our uh second panel. Welcome back, everyone. Um, so AI isn't just something that you tack on anymore. Building AI native products requires design, rethinking design, data, and user experience from the ground up. Today's panelists are doing exactly that. Uh, I'd like to turn it over to our moderator for this morning's panel, Penny Lee, CEO of the Financial Technology Association and Washington Policy Whisperer. Uh, I'll hand it over to Penny and the panelists to introduce themselves and dive in to what's sure to be a fantastic discussion. >> Well, thank you Jeff and thank you Gibson Dun for sponsoring this segment of the application layer. Um, I'll have all the panelists come up um and let them introduce themselves because they will do a much better job than I will. Um, but as Jeff was saying, my name is Penny Lee, president and CEO of the Financial Technology Association, trade association representing the fintech industry. So with that, we will get the panel started. Amias, let me start with you. >> All right. Um, hi everyone. My name is Amias Scarity. I lead the US early stage practice for QED investors. We are a large global fintech focused fund and I'm also the vibe coderinchief at QED. So that's why I'm here, I guess. >> Exactly. Awesome. My name is Leah Price. I lead Tinman AI platform at better.com. Better.com is a mortgage lender and I'm leading the effort to transition us to become a software vendor. You can think of us like Stripe for loans. >> Uh, hi John Sun, co-founder and CEO at Spring Labs. We are a AI conversation intelligence platform. We basically help financial institutions understand what their customers are telling them using uh AI and uh my background has been in uh FinTech basically my whole career. So before Spring Labs I was the uh co-founder and chief risk officer at Avant and uh really started my career at Unova where I ran credit risk and >> and a special shout out to John. You just flew in on the red eye. So we'll get you some coffee. Uh, thank you, Penny, and thank you, Jeff, and thank you to Zanb and Chris for for having me and for an amazing DC fintech week. I'm just blown away. Um, I was just commenting that it's changed so much since, you know, the early days when it was largely this academic gathering. So, I'm Kabir. I'm with Flourish Ventures. We are also a global fintech fund. Um, close to 900 million in assets under management, 100 portfolio companies. We were early investors in Chime, Alloy, Hummingbird, uh, Spade, a bunch of other companies. So excited to be here. >> Let me start off um, Amir with I mean with Amias with you and that is I would argue that fintech has always been kind of AI enabled or had embedded into it some type of AI services. But we're seeing a shift especially with you know um aentic commerce and a and a lot of different things. So talk to you talk to me a little bit about kind of what you've seen from QED from when you were first there and kind of the early modeling of AI use in fintech to kind of maybe where we're progressing now. So I think I think this is such a good way to frame it because fintech has always been the promise of using technology to make financial services better. And for a long time the lack of success for any given fintech often hit against a wall that X or Y or Z was just not quite automatable enough. Like when I started at QED, you know, I see Joanne and Chuck in the in the second row here. Like we spent a ton of time talking about compliance automation and we talked about I mean Joanne has done so much on machine readable code, right? That we had to like rewrite the entire law code so that it could be machine readable and like that was a great effort and it didn't quite work. Um and now the idea that a machine can read the code without any effort by the lawmakers is just true. I actually think um and when I go talk to large incumbent financial institutions, a lot of what I'm doing is sort of the pied piper of fintech in those contexts is saying to people that the promise of fintech is only truly going to be realized now because it is this advent of generative AI that is going to f for the first time allow us to actually deliver on the promises that some people on this panel have you know making and trying to make real um for for going on 17 years >> and give I kind of talk to you a bit about your portfolio too because you've also see this transformation. >> Yeah, I couldn't I couldn't agree with Amayas more but we we've only been at it for 10 years Amayas. So it's not it's not been that long. >> We were founded in 2008 I don't know about you. >> You be longer. You've been longer. But you know, you know, the thing that's on our minds is the anthropic co-founder uh characterized this in a in a podcast that this is the largest infrastructure buildout what's happening with AI in human history. So you know Apollo was $50 billion. Manhattan project was $300 billion. AI already this year has been $300 billion. 7 trillion is already baked in. It's the largest in human history. And it's also the fastest moving technological moment we have ever lived in. I was talking to a founder who is a winner of the last decade of fintech. They had a big IPO earlier this year. Uh you know brilliant engineer very well respected and he was saying he is building something new unlike any other moment in time. the technology is moving so quickly and so despite all the investments we've made up to this point I think we come to this moment with tremendous humility and we you know with a real learning mindset because it's moving so quickly and uh so I I just we don't have the answers for what the world will look like but clearly it's going to be nothing like we have seen before. >> Leah can you explain to your product journey as well? >> Sure. Um so uh I'll talk about better.com's product journey. I'm I'm lucky enough uh to have joined after the technology was built. I just get to release it out into the world and that's probably the most fun. Uh but better started their journey as a mortgage lender about 10 years ago. Uh I'll if you're not famili familiar with the mortgage industry, most of the technology that is in place today was built in the 1990s. So something that was >> so very modern. >> Yes. Yes. Uh so for us having something that was built in the last 10 years positioned us very well once the whole Gen AI thing took off. Um it's really difficult in our industry when you have these different pieces of technology that are very much silos. You have data providers, you have a point of sale, you have a loan origination system. And um we had built one integrated platform because we thought that that would be much more streamlined. And it turns out that you can just once you build the machine that's able to do number crunching and making decisions then you can just layer on a large language model on top of that. The LLM doesn't have to make any kinds of uh do any kind of number crunching do any kind of deep data analysis. The large language model is just taking everything that comes out of our loan origination system and turns it into something that can delight a consumer that can help them make a decision about one of the biggest uh most meaningful financial decisions in their life. >> And are you is are you seeing more consumers now being able to take advantage of your software because of your LLMs and kind of discovering new consumers? um what has the impact been? >> Yeah. So, one of the things we think about is how do we meet consumers where they are? We are all so used to just like looking at our phones and being glued to our phones and clicking and typing. But that soon is not going to be the norm. Soon I'll just be able to say to my refrigerator, I want to buy a new house. Just figure it out. I do I do that already, you know. >> Right. [laughter] >> What kind of results do you get from that? >> I'm working on it. [laughter] [gasps] >> Right. Or >> what kind of refrigerator is that? >> Or if I don't want to use voice, then I can chat because I don't want to I I'm in a doctor's office and I just can't handle it right now. Um or maybe I do want to speak to a human. We're obsessed with meeting the consumer where they are. and you want whatever interaction model to have the context of your last conversation. So, one of the things we're doing is we're making sure if you have a conversation with our phone bot, Betsy, better Betsy, that's the name. We'll rebrand her whatever our lender wants. But if you have a conversation on the phone with our voice bot and then you transition to the avatar because it's in the middle of the night and you're just nervous about what's going on with your mortgage and you need someone. If you saw my demo, she's very comforting uh and very funny. If you want to talk to her, she'll remember everything you told her on the phone. And then if you want to text a question uh on the other side of the text, the bot will understand. And then the human has to actually catch up to to what the bot has been said. It's actually harder for the human at some point. Um, yes. So, we're really meeting the consumer where they are and these large language models make that much easier. >> Leah, can I Sorry, Leah. Talk about voice for a second because if people I would have said a year ago that human beings are going to be totally creeped out talking to AI robots, but I will tell you everyone in our portfolio that is not true. And you're living. Tell us about the consumer response to Betsy. >> Yeah. So, >> we're not an investor. >> No, I I've never met him before. Did you see my avatar? >> I did. >> Oh my god. Okay. Well, it she killed this morning. All right. Okay. So, not about the avat not about the avatar, but about voice. So, I had the same question when I was at FHFA um last year. I asked the CEO of better.com. I was like, I actually personally prefer to text more than to speak. Is that cool? Will that work? And he was like, no, the consumers, they love to talk to Betsy. Uh, we had 600,000 consumer interactions with our Betsy voice bot in second quarter. 200,000 of those were after hours. Um, so in terms of voice, what having a voice bot really enables you to talk whenever you want and that's probably not 9 to5 Monday through Friday. That might be while you're at an open house in the afternoon on a Sunday. >> John, get you into the conversation. You obviously were part of the kind of the OG of marketplace lenders with Avant and now you're at Spring Lab. talk to us and obviously Avant was using LLMs and others from almost its beginning, but talk a little bit about kind of where you are seeing some of the market fit needing to be in the use of AI in that. >> Yeah, I mean I think um I'm I'm going to actually piggyback off of what Amaya said. I I think um LLMs finally allow what was how did you phrase it? Allow fintech to deliver on the promises that we've been making all these years. So Avant was started back in 2011 2012 right and I think the main innovation there was kind of the AI of its time which is machine learning for underwriting being able to use these machine learning models to ingest thousands of attributes for better you know capture of of false positives and false negatives. Um and now all of a sudden what we have is something that is much more accessible to kind of most users within an organization whereas there's a small team small. So we had a team of 80 people at Avant that were like machine learning engineers and data scientists but still a relatively contained portion of the company that was using these traditional machine learning models for underwriting. Now I think with Gen AI models it's really transformed the work of everyone within these organizations. Um, and you know, take what I think of as as kind of the, you know, one one of the worst processes today, which is um, how do you convert business requirements to data, right? Like how do you communicate with your BI team, right? And and that process has been transformed so dramatically by AI that like within Spring Labs, we don't have that translation anymore. business users kind of use AI tooling to directly grab the data that they need for the work that they need and it's much more consistent. It's much more reliable and it answers questions in a way that was kind of never possible before and that's just one of a very very narrow number of use cases that you can kind of apply generative of AI to within financial institutions. >> A question for everybody is is there dangers in the over reliance now of AI? Are we seeing accuracies in the model or you know as they say the machine learns against the machine like how are we ensuring that the information that companies are building their foundation on are accurate and actionable um and not almost relying too much if that's a possibility. >> Um so look I think it's definitely uh a risk. I think that it's actually a little overblown. Um, so I think you know when when John and the Avant crew and the marketplace lenders were first trying to use machine learning, there was a lot of concern and a lot of real evidence that the biases from the real world were getting embedded in these machine learning models. Um, and I think to some extent that's probably still true. And also LLMs are just sort of not really relevant in that same way. Um so they are surprisingly good actually at m traditional machine learning tasks which is one of the like funny and scary things is that these language models are actually good at um statistical analysis. Um, but I think the application is a little bit higher level and it and and so some of those those models that I think especially here in DC people have of like oh well if it's trained on the internet and the internet is a cess pit like won't it be a cess pit? And that's definitely true. Um, but also because they're so responsive to the prompts that you give them. I think the emerging best practices that all of our companies are are discovering or failing and then discovering um are that you actually can control the inputs and outputs relatively um relatively more easily than some of the machine learning models. Even though the underlying LLM is in some ways harder to um harder to understand, but the the the way people are using it and the way you control it is actually a lot easier for business users and so the the people who are sensitive to the need are closer to the controls. >> Oh yeah, sorry. Just to kind of add on to that, you know, that was something that surprised me as well that LM are actually pretty good at statistical analysis, traditional machine learning, but it's not because transformer models are great at it. It's because LM are good at writing code. And if you're good at writing code, you're good at everything. So I think that's kind of the surprising twist that you know the last couple of years brought is if you pair something that can write code can access data through MCPS and access services and tooling through MCPs then you can do almost anything kind of within within an organization and it's it's kind of been quite a wakeup call. I appreciate the question though and um you know I would I think what Amias was saying I hope it ends up sort of not being a material issue but a number of people are concerned about sort of the reliability of of these models and from an investor perspective it's an opportunity you know so we are looking at infrastructure players that can I won't say ensure reliability but can help you work towards more reliable outcomes and and you know this idea that eventually that becomes part of the infrastructure of finance that there's a whole layer that ensures some degree of reliability of your of your models. Remember the challenge here is autonomy. These models are we are building them to be autonomous and I think that's where the reliability risk is different than in the previous generation. >> Well, when you say autonomous, I would say probably every regulator in the room probably got their backup a little bit on that aspect. But I mean how do you though protect against this machine that might be out there autonomously working on you know engaged and yet ensuring that consumers are protected that the products are saying what they're actually functioning to >> you know it's software and so even when you built the previous generation of models we ran tests to see whether they would actually give you the outcomes you want. Now imagine those tests running continuously as a separate layer on these autonomous models. So it we know how to do it. We just haven't built that infrastructure layer out yet. And I think that that's where a lot of smart engineers are spending time because it's a big problem. >> Yeah, I can talk a little bit about this. So I think I've I've uh said earlier in the panel that uh the way we've set it up is our core platform is doing all of the calculations and all of the sophisticated data analysis and all of the agentic stuff. all of the actions happen in the core and we use large language models just to be the human interface. Um, and I was thinking about like how do you distinguish between an AI native platform and a nonInative platform. So a nonInative platform might just have a chatbot slapped on top of it. And the way you would know it's not fully in integrated is that that chatbot wouldn't be able to take any actions on your behalf. It'll be able to answer some questions, but it won't know very much about you. Our AI native platform, a chatbot that's integrated, it knows everything you've already told it. It knows exactly where you are on the page. It knows what you want. It knows what your preferences are. And it can take actions. it can get all of the right information to create a loan application and it can present you with options for your loan. Um, that's nearly impossible if you just slap some kind of a bot on Expedia or something like that. >> Another risk I think Penny that that I want to sort of highlight is I think it is traditional in financial services, right? Like our common law principles always rest on the idea there is a human who puts their hand over their heart and promises. Like that is the fundamental common law principle in everything we do. And there is no way for any AI to put their hand on their heart and promise to do anything. But I think we all know how weak and how riddled with problems and conflicts and criminality and fraud that handon heart promises. >> Bias and bias >> and bias. Exactly. I There was a group of community bankers. I They all came in and they yelled at me. And there was one particular group that came in and said, "We do relationship banking." and a staffer came to me and said,"I never felt worse about relationship banker banking than when that person um talked about it." So, so exactly right, Kabir. And so the thing I want to say to people when we think about this is the default natural answer in this moment is human in the loop. But I actually think human in the loop might be the worst answer. It's the best answer for today, but it might be the worst answer because human in the loop is an arbitrage. It's an arbitrage between or it has a very real risk of the machine making the decision, the human deferring entirely and then the human putting their hand over their heart and that is that's a real dangerous system. And so, but you think it's all going into machine like there's going to because I mean that's the problem. You know, the promise of Moneyball is that it's going to have all of these algorithms out there and decide what the best picture is, but you actually picture is, but you actually still need to see them throw a pitch. I I think what it is is if you're going to automate something, you have to do the hard work of making sure that the automated process is going, as Leah just talked about, the automated process is going to be fair. The automated and if you if you rely at the end on, oh, and then we'll put a human in the loop. I think that's where we get the risk because we're going to get the biases from the automated process and then we're going to actually end up in a situation like if you have a loan officer doing hum hand on heart I reviewed a 100,000 loans today. >> Yeah. >> No, you didn't. Right. So that the fact that your hand was on your heart is completely irrelevant. >> But the refrigerator is going to be better. >> Well, again, if you do the hard work. So it's not to say that that we could actually say what I'm saying is possible. It's not possible. You're going to have to have humans in the loop today. But it is saying that if we rest too much on the idea of human in the loop, we are missing the fact that that system is weak relative to the obligations we should impose on financial. >> I think the risk is we will not invest in the sort of infrastructure to give us that trust because we you know even your reaction I mean I think it's it makes us anxious to think about it and it's all legitimate. we should be anxious about it. But the truth is we can engineer responses to that and we should invest in trying to do that. And >> it's hard. >> It's hard. >> But that's why we should do it. >> Uh we're we're in the process. We recently um implemented this. Whenever a human loan officer or underwriter is going to deny a loan, they're about to click deny. Before they do that, they are required to submit the loan application through our AI bot, Betsy, and have her go through and see if she would also deny. And if she doesn't, what does that say for this concept of human oversight? Now, it's like bot oversight. And we have found instances where the bot actually pointed out that this this borrower should have been approved. They should not have been denied. And what happens when there's a conflict between the two, between the human and the bot? Does the bot win? >> Someone is in trouble. No. Then you would escalate it and you'd go to a manager and you would What's nice about Betsy the bot is you can read all of her logic. The reason she made every single decision looking at every single data point that takes like she types it up. It's like 20 seconds. You go to the human. You tell me right now why did you make that decision? You're not going to be able to tell me in 20 seconds. You're going to have to maybe go through your notes and look through the system. So, humans are it's going to be hard being a human. But John, you know, and you're are you also seeing the, you know, when you go before regulators, you're also as, you know, going to have to explain these models and going there's going to have to be some type of in human interaction with them and the explanability in your in in what you're building like how do you balance those two needing for that explainability to ensure there's not but make sure but also recognize the efficiencies that it brings and the openness that it brings to a whole host of you know consumers. >> Yeah, I mean it's a great question. Um but we just saw this play out literally 10 years ago with machine learning models in in underwriting. So before you know the first machine learning we'd like to think at Avant we probably had one of the first regularly approved XGB what's called a gradient boosting model for real-time autonomous underwriting and that process was kind of a long arduous journey to kind of get to a point where there was regulatory acceptance and understanding what was going on and I think part of the journey was really kind of being at the forefront of inventing a new vocabulary that helps regulators understand you know what are some of the the risks uh making explicit some of the more you know implicit uh aspects of of of model risk management. I mean the entire kind of machine learning based underwriting model risk management stack was built within like the last 10 years, right? Like you're required to show like disperate impact analysis. You're required to show false positive rates. You're required to show false negative rates. is required to do you know uh what's called um um uh uh independent kind of variable uh plots against you know the the outcomes to make sure that you know directionally all of the variables are kind of trending in the same way that you would expect them to. So there's kind of this like whole package of things that you do but again it's all invented as a result of machine learning models and the vocabulary to understand the risks of these models was invented as a result of that. We're going to see the same thing here, right? And and we're going to invent a new set of vocabulary for how to communicate with regulators on, you know, what the false positive rates and the false negative rates and the hallucination rates and all of these things for LMS are. But I will say like to everyone's point I think everyone whenever they look at new technologies has this tendency of comparing it against this theoretical perfect of like what if it's different from this theoretical perfect but again like what you're comparing against isn't the theoretical perfect it's against humans right because that's the process it's replacing today I can tell you you're not going to be able to generate a false positive false negative rate on a consistent rate basis is for a human, right? So that's the biggest black box. Humans are the biggest black box. At least I can explicitly understand and and communicate what the risks of a model are in a way that I would never be able to do for a human process. >> Kabir, you have invested a lot in some reg tech companies as well. And um Michael Sue had an interesting paper published yesterday for the DC fintech week and that is maybe we should think about especially when we look at AI is that we're regulating or we're thinking through from the outcome perspective not from the modeling perspective. How do you think that evolution because the technology is happening so fast? How do regulators respond and should they be looking at more from an outcome perspective? >> You know, I don't I haven't read that paper yet, so I I want to think about that. But I think that um so we've invested a lot in fraud, compliance, technology, uh in part because we believe technology can help address those challenges. We can fight fraud with technology. We can fight scams with technology. Uh, one of the companies we've invested in is called Spade. You know, between the time you swipe or tap or whatever and the transaction is approved, it's 200 500 milliseconds. These guys can improve the quality of the data stream within 30 50 milliseconds which means that that data now fed into your anti- fraud models will give you better outcomes and they have so they are the sort of the data behind sardine for instance uh they have been you know bunch of banks are using them etc. So we are we are big believers in that power of technology and addressing those challenges. Hummingbird is another great example of that. uh uh Joe who is the CEO of Hummingbird was telling me that you know they have I think millions of financial crime investigations have taken place on the Hummingbird platform that would not have been possible otherwise. So truly believe in that. I think in terms of sort of what does that layer look like for the AI world uh maybe it's an outcomesbased framing. I think one of the things is what I was saying earlier is that we need systems that help us understand whether these models are indeed reliable or not. Are the outcomes indeed what we want and so I think those could be the next sort of opportunities for us to invest in the rect layer. Um but I don't think they will necessarily be fintech or financial sector focused. I think a lot of the things that we are seeing are actually sector agnostic. they are sort of how do you get these applications to work across multiple verticals and to give us those reliable outcomes across multiple verticals >> and MI is one thing that I know we were um at a dinner together the other night and >> I don't have rules Penny >> oh yeah sorry but we we can see we were at a dinner [laughter] um but since we are not together but one of the things you made your point you made was all of this is underlined with trust and how do we I don't know that you can regulate trust. Um, but how and maybe it's incumbent upon you um as an investor to ensure your portfolio companies instill this like how vital is trust and and how do we go about ensuring in either the models the outcomes the algorithms how do we ensure trust >> so I think um one of the things that you know whether it's the Mike Sue paper or some of the the rag tech ideas that we've been working on is the idea that sampling just isn't that good and that in the era that we have now entered and you know some of us thought we entered this era 10 years ago um but I think we're clearly entering the era now the idea that you wouldn't universe test is is just kind of unacceptable and look it's going to take time before we can ask every small bank to universe test every transaction but like 10 years from now 20 there's some point at which the idea that your compliance is based on sampling will just be an unacceptable framework. And then the second thing that that follows very very quickly from that is and again using you know some version of what Leah talked about like the idea that you could take a non-compliant action like actually you shouldn't be able to take a non-compliant action like I was breakfast with Wade Arnold at move he was talking about um a number of his customers are finding that there's often insufficient funds when they bill for subscriptions and that's a big problem. It creates operational headaches. It creates, you know, overdraft fees for those customers. Well, one of the big things that Wade is doing is just if you shift the payment rail, you can get real-time data. If you do real-time data for the cost of a few pennies, you find out whether the money is there or not. And if the money's not there, you just don't draw it. Right? So, that's a good example of a way that we take automation, take data. We're not just saying, oh, you know, did Wade in inadvertently create, you know, excess overdraft fees and if that is that a compliance problem? No, no, no. We're going beyond that and we're saying this is a problem. We should have the data. We should have the universe testing and ultimately we should have agents take actions such that non-compliant actions aren't actually possible. And I think look that's a hard hurdle to clear but I think that's the hurdle that we in finance should be aiming towards and and set our sights on >> but I think the crisis of trust in the financial system we all share responsibility for that like you know I'm sorry even the regulators share responsibility for that I mean you know Aspen put out this report about you know it's a joint report with 30 uh corporates they quantify that $300,000 per minute were lost to scams by everyday Americans last year. $168 billion. Why would you trust the financial system and why would you trust the economy if that's actually going on? So I think we have to address from a consumer lens if you're not solving for those things. We are not going to get to trust whatever the plumbing looks like. So I think we need to reframe that and think about what the consumers really want. And we say also I'll put a plug in um obviously our we have been advocates for an open banking system um to allow for the data portability for your financial data to be able as you permission it as consumers to be able to permission it not only own it but permission it is foundational for so much of what each one of you all do and Lee or or John in particular I don't know if you have an a thought on kind of that vitaleness of ensuring that we do continue to have a data permissioned or the consumers be able to permission their data. >> Yeah. I mean I I think I'm a huge data geek. So I'm always a fan anytime anyone's like do you want more data or less data? It's always more data, you know, and and there is this walled garden risk, right? If you if you let the financial institutions control this data that's essentially the the joint you know effort between the consumer and the bank if you give sole control of that to the bank then there's a risk that that data becomes less available and that means worse financial products for for everyone involved it means worse underwriting which means less access to credit in a lot of cases. So I'm always kind of obviously a big fan of of data you know portability and and and openness. Yeah, I was as you were talking I was having this like surreal moment of like convergence between the blockchain crypto digital assets world and the AI world because I was in the blockchain space before I went to FHFA. Um but I'm totally 100% behind um permissioned financial data for consumers that would help all of us. going to go back to the beginning which was um are we do you slap on AI or um is it now AI native in you know from your port I I'll talk with Amias and Gabir first from a portfolio perspective is is it a journey whether or not you're in a growth phase and you and you are acquiring or you're seeing the ability of startups now that are AI native to like actually supersede the incumbents right now because of the incredible amount of technology that's available and um the LLMs and others and they can just iterate so much faster. I think the the truth is that at this stage of the game like AI native is mostly a thing that you say because if you you believe that if you say it in just the right way with just the right intonation that um >> your valuation goes 10x. >> Yeah. Correct. And look, if you look at the data, the um funding in venture for quote unquote AI companies looks like 2021 and the funding for quote unquote nonAI companies looks like 2022. So there's a lot of pressure out there to be quote unquote AI native. And I look at our portfolio and we've got a mix of quote unquote AI native and nonquote AI native. And I would say I'm not sure that that's a distinction that actually is relevant for those companies that are really pushing. Now, it's certainly a distinction that's relevant. There are a lot of people who are not embracing it. Um, but I think as Leah just talked about, right, like the system of record isn't an AI system. The system of record is like a hardcore logically consistent, mathematically correct, well- tested in QA system. and then they've created an AI native product on top of that. So I think we don't yet know what the emer what the emerging best practices and the emerging definitions are and they're starting to come up. Um I'll give you just like a very simple example but really illustrates the the challenges of what this means. So Sam Alman, good job. I mean wow. Uh but Sam Alman is out there saying there are going to be billion-dollar companies, trillion dollar companies with two employees cuz again he doesn't understand how many zeros uh separate a billion and a trillion. Um but nonetheless, so that that's like a big idea, right? Well, I got pitched by a company the other day who said, "Well, wait a minute. Economics says that if something is cheap, you should buy more of it." So this company with 200 engineers in Pakistan has shipped more product than Toast in 18 months. So Toast has been been around for 15 years. They have an extremely thick product suite. Great company. And these guys said, "Well, wait a minute. AI plus lowcost offshore engineering. We should buy more of engineering. We shouldn't try and build toast with three engineers. we should buy build toast in a tenth the time with AI at the center of it. And so I just as an example, those are very different ideas. One is like me vibe coding and the other is like real industrial, you know, strength. We're just going to ship way faster. And I think that's a good example like which of those companies is AI native? I don't know. Both. >> Both. >> Yeah. John, are you going or sorry, Kabir, >> you know, I I wanted to go back to the open banking piece if I may because I do think that's a huge structural risk that exists because especially for consumerf facing solutions that this whole notion of self-driving money, you know, things that are really going to solve meaningful problems, we cannot have that if data is not flowing freely. And I was talking to someone who's working in the emerging markets. It's the same challenge with the telecom operators in the emerging markets. You know, mobile money has been successful. They are effectively the banks here. You know, I don't know why we have an oligopolistic banking structure in the US, but we do and they are the data holders. So, I think that that is a big deal and I'm very glad that a bunch of folks are still fighting for that because I think we do need a better outcome there. Um I think in terms of what you know digitally native you know AI native or not AI native I agree with Amayas you know I I don't know uh the way we think about it is that the if if you remove the AI functionality does the product sort of you know catastrophically degrade and so then of course if it does then it is you know uh AI native if it doesn't then it's not AI native and so and and in our portfolio everyone like in our lives is everyone we have in people have incorporated the LMS they are trying to engineer all sorts of efficiencies through these product like all of us we are all using it all the time our portfolio everyone is using it all the time and in some cases the advantages are significant in some cases they're not uh but we are seeing things that are uh you know maybe closer to this our kind of mental model of an AI native piece and these are folks that have really specific domain knowledge and they're building workflows goes right from the ground up in that specific domain knowledge and uh that's pretty meaningful I think that you wouldn't have seen in the previous era. So one of the companies I'm really excited about it's in a you know not an exciting sexy space they brick you guys might have heard of them they are facilitating licensing financial licensing for anyone state federal US global simple you know idea but really powered by being truly AI native so they can automate the whole damn thing uh at a level that wasn't possible before so that stuff seems very exciting and truly AI native to me >> and John when you are when you as you um launch new products um are you looking from an acquisition or are you still looking to like trying to build it natively within Spring Labs? >> Uh yeah, I mean we we kind of build we're we're very much kind of a build shop, you know. I I think um it it's kind of a carryover from the early avant days when we had to build everything because nothing existed. Um, you know, that being said, I think, um, you know, back to this AI native question for for a second here, like it's a moving goalpost, right? Like the AI what what constitutes AI native is is always kind of a moving goalpost necessarily because of how fast the space is advancing. I don't think people can kind of like take a step back and realize how quickly this space has been advancing. If we were sitting here last year, MCP servers did not exist as we were sitting here last year. And that's basically the API layer for LMS for AI models. Like this this really critical part of infrastructure did not exist if we were sitting here last year. Reasoning models kind of existed. It was a bit of a pipe dream. There were the first versions were out. It didn't work that well. And now they work really well. Right? So if you were to say in 2025, what does AI native mean? Are you AI native? you know, organization. Anyone started within longer than 12 months ago shouldn't be able to claim that, right? They they weren't like the architecture to be a AI native company in 2025 didn't exist. But I think that's why, you know, it it is somewhat of a moving goalpost, but at the same time, it's somewhat of of kind of a mentality and a philosophy as well to be able to kind of continually see advance with the technology as it kind of grows. >> I I think this is a super this is the point I was trying to make earlier. This is a super critical point. And if there was one thing you take away is that this is unlike any other moment in history truly. It's not just a quick nice talking point. It's a fact. And so, you know, it's going to look very different 6 months from now. It's going to look very different a year from now. And so, our definitions are also going to evolve. And that's why we we have to stay very disciplined on what problems we are solving. You know, does it really move the needle on that problem? We have to be very disciplined about that. >> And Leah, are you seeing your product change? Um and has it the um the great startup world pivoted at all just with these new technologies that are enabling you to do perhaps different things or more expansive things or how are you evolving as well? >> Yeah, I mean we started working with Betsy Voice uh last November. So I guess it's been just about a year and there >> Yeah. Um and >> how many how many names did you go through like before you got to Betsy? Well, it's >> Suzie. No, >> our company is better. So, Betsy goes with that. You know, we we'll brand it. We'll let anyone brand it however they want. Um, but it's uh interesting if you look at the uh efficiency of how we've been able to process loans since Betsy Voiced launched. It's a beautiful chart. It sells so well. Um, but you can just see like number of files able to be processed per day because Betsy is able to ask basic questions over the phone that normally it's just hard to track a human being down to to get information. Um, so we are at the bleeding edge like we'll um we've experimented with voicetovoice models that the voicetovoice models aren't fast enough for us yet. So, um, but we're eager. We're pushing the big AI companies to come up with faster and faster models so that we can, um, roll out what our customers should expect from their experience. >> Uh, quick question rapid fire for everybody. Do, um, going forward, does every company need to have AI in its name or is it going to be just expected that they're an AI company? >> For valuations, yes. for reality. The second is true. >> So my title has AI in it. I didn't want it to. I was like, it's just software. That's silly. But for the reason he mentioned, we need it there. >> Yeah. I mean, I I think it's like being an internet company, you know? It's like there's not a lot of companies that aren't internet companies anymore. So I >> I did have a founder call me yesterday say that they bought a domain name withAI and he just wanted to let me know that next year they're going to be.AI. I'm like, "All right, that sounds [laughter] good. Um well with that thank you all for coming. Hopefully you've had a couple of takeaways as we are moving into the ubiquitous of AI I think was the takeaway. So thank you all for coming. >> Thank you Penny. >> Thank you Penny. [applause] >> We we've got a 35minut break here. You can hang out here second and fifth floor. there's uh refreshments and we'll see you back at 11:55 sharp for the next panel on AI and stable coins. Thanks.