The Riskiest Agentic AI Use Case in Finance Is Lending

Host: Foreign, Uh, is I would say one of the riskiest or more complex agentic, ah, AI use cases in finance. While everybody's talking about agentic, uh, commerce, agentic, uh, payments, stablecoins, new protocols. For all that, I must say that the agentic lending topic is the next one and I'm very interested in this because it really touches the lives of real people who want to buy a home, buy a car, grow a business, plan for the future. And here in the eu, I must say we have a very clear regulation, uh, through the EU AI act that classifies that when AI is used to evaluate the creditworthiness of a person, then it's high risk. In the US I think there's more nuance and more questions, uh, around fairness, open discussions to go deep on this very important topic for our society and for financial services. I'm joined by Tamara Lane, who's the founder and CEO of, of um, Empower, and she is building agentic AI infrastructure for lenders. Welcome Tamara.

Tamara Lane: Uh, thank you and what a wonderful introduction. I, uh, couldn't agree more that AI is the next frontier of lending and it is also the early stages of it. So it is a place where there are still so many questions, so many questions.

Host: Let's set the scene around innovation in lending. When I think of it, we had maybe two decades of fintechs that have been, let's say, attacking this problem. We've seen alternative credit scoring, we've seen cash flow underwriting, we've seen the use of machine learning to help offer better credit, more credit, better risk management, serve the underbanked. Tamara, in your opinion, what did these innovators fix or improve or uh, attain? And what is left really that we hope that AgentIQ and the developments in AI can add.

Tamara Lane: So you're going to hear a real mixed reaction from me on this as I was thinking about the question. So let's set the stage right. If we talk about lending and the how lending started you have in the 1950s, really the foundation was set for how our modern system was built. So the pipes of how our modern system was built and if you think about in the 1950s, lending was still to men only, right? Banking was really a male task and it only started to become more inclusive. When we talk about the 70s and that's not that long ago, late 70s are uh, when women were really allowed to get bank accounts and credit cards. I was just watching a something that was put out by PBS and it was interviewing Hillary Clinton and Gloria Steinem, who, Hillary Clinton, who went on to be the Secretary of state had to ask, had to get permission and get signatures from her husband when she was an attorney, making more money to have certain financial services. And so that's mind blowing that we're still, let's just say we're still in the time frame where that is something that people experienced. Now you fast forward and, um, the Internet was created. Now you're talking about the last two decades of really growth. So we've had such incredible technical advances since then. And we can't discount the fact that we went from incredibly manual processes to now technology has come in and created really great ways to make lending faster and to try and find tools that create better answers for underwriters. And so by that, now you have new scores that are being produced with machine learning and they're bringing in a lot of data and they're giving more, more indicators to underwriters. And so in the last two decades, underwriting has become faster, but it's still become more of a black box. Right. You have all these mysterious systems and scores coming at underwriters with hundreds, if not thousands of data points coming in. You don't know where they're coming from. The way that the machine learning is being used, you're like, it's great because you feel like there is some underbanked being pulled into the system and you feel like there's progress. But it seems to be these shadow systems popping up everywhere where it's still not transparent. The borrower still doesn't quite understand the scoring system and it's become incredibly complex. Right.

Host: So the risk models have been developed because we added more factors or richer data sets. They're still not explainable and we're not sure if there's biases. We can say that we've broadened the, uh, serviceable.

Tamara Lane: And then the question really goes to, have we solved the problem? And when I started working on Empower, the stat that really struck me and was like one of the foundational statistics that I couldn't move away from. And why I started to work on this was the gig economy. Right? We're talking the gig economy, the new economy. It's about half the U.S. workforce and 6 out of 10 said they've been denied a loan. So if that's 60% of half the U.S. you know, work population, that's a staggering number that is still left out. So we can't say that we've solved the problem with machine learning and with

Host: the alternative data that goes beyond the payroll.

Tamara Lane: Exactly. Because our system is still stuck in Some of our old structures that we've come up with new tools to create better insight, but yet we haven't moved away from the old structure and thoughts, thought process of how our economy and how a borrower looks in order to really make a structural change in the financial thriving of our economy. Right. Of our workforce, of our borrowers. So as where I'm trying to not be like gloomy, I'm saying yes, technology has really done a good job and there's been a lot of advances, it hasn't actually solved the problem of getting more loans into good borrowers hands.

Host: Okay, so you still see a gap and that is what you're trying to exploit. If you want and design better systems to serve more people. And tell us a little bit about Empower, are you building a dual ecosystem? One that is serving retail individual consumers that need credit and another that is really a B2B offering for banks and credit unions to better serve their customers.

Tamara Lane: So we are purely B2B. So my goal is to help institutions reach more customers. And uh, one of the things that we created when I started, I very much, and I still talk about the new economy. It is a passion of mine to get more borrowers into the financial system to get them better access, better rates. But what we know is that banks also want to serve their prime customers. And so how do we blend that together without giving banks a new scoring system? Right. Without being like adding another shadow system into this ecosystem for them to try and make decisions when they're already having a hard enough time parsing through the kind of systems that are there that are stacked on top of each other. So at uh, Empower we designed the new, the first agentic ecosystem for lenders to create better lending experience for their borrowers, whether they're prime or subprime or a new acquisition or gig workers or creators. Our agents can parse through the data, do the manual work. They're all policy bound and compliance first. So as an AI native company we're able to come in, be more agile and, and policy bound means that we are always up to speed with the compliance and regulations that are in the region.

Host: How uh, does your intelligence, where do your agents sit with respect to the loan origination systems that every bank will have, Be it through and chino through encompass whatever provider their own system. Are you proposing a uh, replacement and integration? I don't know. It's an additional layer, give us a sense of we can and why.

Tamara Lane: Yeah, we can sit on top of or be an LOS system and we understand That a lot of organizations like their tech stack or they are in a relationship with their providers that they are in for a while. And so organizations that want to incorporate AI, especially policy bound AI, not just a chatbot, we uh, knew that we would have to go in and be able to integrate into their system. Right. And so that's how we built because we didn't want someone to have to wait two years if they were in a contract with an LOS that they liked to be able to use our system. Uh, we wanted to make sure that if a provider wanted to use responsible AI that we could assist them with that.

Host: Yeah. Can you give us a uh, sort of a high level idea of the architectural design without getting technical of what we are talking about here? Is it very technical? Are there different layers or maybe what kind of agents have you designed? I assume there's multiple agents, how they connect and those types of questions.

Tamara Lane: Absolutely. We spent a lot of time on our underlying proprietary technology that was, that is the policy agentic so structure which means that we can pull in a lot of information and we can hook into a lot of systems and build policy bound agents on top of that that are incredibly agile. We've started with the first six agents which are our acquisition retention portfolio, risk mitigation, so pre collections risk mitigation agent and financial fluency agent. And so with those agents they are action oriented. So our acquisition agent actually does the work of collecting all of the information and from the borrower and the third parties that need to bring in the information to compile it. Oh we got underwrite to compile it and synthesize it for the underwriting agent. The underwriting agent then provides all the information that the human underwriter needs in order to make a fast decision. So those two agents, they basically compress onboarding time and increase the, the human underwriter's ability to make a good decision. That agent there to assist with that. And then we've got our other agents that operate off that. Like I said, our risk mitigation agent, which is an agent that can help pre default. Right. Before something goes to collections, there's moments where you can intervene, the organization can intervene and say, hey, we noticed you just missed a payment. How can we help you?

Host: And the organization doesn't have to wait for month end or whenever you get the report of what's been missed and matching all those. Yeah, yeah.

Tamara Lane: And sometimes there's so many reasons that people miss a payment. Sometimes someone's traveling, right. Sometimes someone just forgot to set up automatic payment or someone Thought they set up automatic payment and three months later they're like, wait, this hasn't been. I can say from experience that's happened to me.

Host: Yeah.

Tamara Lane: Uh, I was like, oh, shoot, life just gets away from you sometimes. So there's a lot of reasons for missed payments that have nothing to do with not having enough money. Right. Or so being able to have an agent who can intervene in these situations drastically help the bank's outcomes.

Host: Yeah.

Tamara Lane: Or the lender's outcomes.

Host: Yeah. And in terms of the, what um, I would call the three elephants in the room, which is the governance, the auditability and the explainability of what these agents are doing and why, how. I assume this is you. Do you have an agent that is focused on that? How have you thought of all this at every step?

Tamara Lane: Yeah. So that's a really great question and a very complex one, so I will say so using AI for the actual decisioning in any lending is problematic. And so we don't use AI in the decisioning moment of any underwriting. We use a decisioning matrix that pulls that in for the underwriter and the human makes a decision based on really good data put in front of them. And I'll back up a little bit. I have an Emmy behind me. I was an investigative journalist and a financial news anchor in the beginning of my career. And one of the stories that I did that actually won the Emmy was on ethical AI and stock and activist investors who wanted a code of ethics in AI development, specifically for Northrop Grumman and Amazon. And I was talking to a whole bunch of experts at the time, PhDs, people working on the development of this. And what was drastically clear was that, uh, anything that you create, any software that you create has bias in it. And how AI is trained has bias in it. We know the underlying. The models that are out there right now that are. The big models were trained, uh, in places like Reddit and Twitter and different places. And now we've brought in better information, the agents are better trained. But there's still an underlying way that our technology is developed and trained where you can't 100% say that these.

Host: There is no bias. You can't. There's a narrative in each, whatever language in each, uh, area, whether you're scraping, uh, digital, uh, media or whatever, you're scraping, there's a narrative that's hidden there that going to be.

Tamara Lane: And so there's use for AI in decisioning, but it is not to make the decision. Because in my opinion right now, and from all the work that We've done. And uh, is that still something that you can't 100% guarantee and so therefore you should not use it? And so that's why we've created our system differently where the decisioning and the actual final decision is human led. And I am a firm believer in designing intelligent AI in an ethical way to maximize human excellence. And so what we do is we design our product so that the human in this loop can thrive. And that's the underwriter and the loan agents.

Host: So that's why I realized that you keep repeating AI prepares and human decides. That's a mantra that, that you're using and it's part of the vision of what you're building. Does that mean, is it a soft version of what the other day, Brian Armstrong of Coinbase, they laid off 14% of their workforce. And AI is one of the, is the narrative excuse.

Tamara Lane: And there's some other factors that we're going into that I've been reading that aren't necess necessarily all AI based.

Host: Exactly. Uh, but what's interesting is he described the vision of using AI to end up with one person teams. So for example, you have a product, whatever the product is, in Coinbase's case, the consumer facing app, because they have different products obviously. So his vision is to use AI to get one person team for each product. So a very, what I would call a, ah, very lean organization versus an organization that does leverage AI, but does so to serve more, uh, and in different ways, new services and so on. Where do you stand? What do you see for lending? Because you're focused on lending.

Tamara Lane: So I've got two thoughts on this. First, many of the lenders that we speak to right now are short on human capital. And so one of the reasons that they want to employ our technology is not to replace anybody. It is that they do not have enough people for the business that they have right now.

Host: They don't have the capacity, the human capacity. Yes.

Tamara Lane: And so that's one of the things that we see in front of us all the time. Second, I go really back and forth on this. I do believe that jobs are going to be changing, impacted. But I don't believe in the lump of labor theory. Right. Which, the lump of labor is that there's only so much work to be done and that if AI is doing that work, then the human doesn't get to do work. I believe that in thriving capitalistic societies that there's, that work builds with capacity. So as you interject AI into the system, it's Going to just create a different type of job to service that AI. And so you're going to have work that builds. And if you ask anyone right now who's working on AI, our workload has tripled or quadrupled. Right. And so there's a changing workforce. But right now I don't see work going down. It's just there's going to be a really harsh, weird transition, um, into the new. And that's what I believe right now. And things change so fast. You can talk to me in three months and see where I stand.

Host: I've always thought, and I'm not a deep expert in lending, my banking background is more investment banking rather than consumer banking. I've always thought of this idea that if money can be streamed then credit, we should get to the point where you could stream credit, whether it is micro credit or major credit. There's very simple examples. I give you a personal example. Obviously, like many of us, I have many credit cards. Some are personal, some are, uh, for my small business. Right. What happens is my needs for credit through credit cards are not uniform across time. Most of the time I'm fine, I'm um, largely under utilizing the credit limits that I have. But then there might be a month where I need triple the credit that I have. And I don't want to be trying to add a little bit here, split it there, split it over there. Let alone I don't want to mix my personal with my business. So there isn't an easy way to do that for that one time need. Right.

Tamara Lane: Yeah, you're right.

Host: And this is a very simple thing that I'm sure a lot of people, uh, face. But there's all sorts of other cases where you need to have credit. What I would call all these edge cases. So there's so many problems if you want that we could solve beyond the. Also people that don't have access at all or what makes sense, how to optimize, does it make sense to borrow? Does it not? All those questions, who's going to build the agent for me and you to give us advice.

Tamara Lane: Right, exactly. That's a really interesting thing. And especially one of the conversations that I get in a lot is the personal shopping agents.

Host: Yeah.

Tamara Lane: Because that all that kind of ties in with your personal shopping agent, like anticipating your needs and acting on your behalf online or financially. Right. And so that goes that you're going to have to have agents that are certified personal wealth managers that are taking the tests and updating themselves. And I know someone who's Working on a bank for agents to transact money because our banking system is not set up for agent transactions.

Host: Yes.

Tamara Lane: How wild is that? Thinking about that makes me go, whoa, where are, uh, we going?

Host: But a hundred percent, because if you think of it, banks are there. I don't know how they are, by the way, going to get licensed, how, ah, they're going to discuss with the regulators to get the license, but it's definitely a need. And right now what's happening is there's other entities that are trying to position themselves as those that verify and authenticate agents, give them an identity, give them a, uh, reputation. Visa and MasterCard are in that business. So it's a world that is changing really fast. Tamara, I want to touch a bit on the, um, U.S. regulation around lending, and I'm, uh, not very familiar with the outstanding questions that are out there around algorithmic. If you want credit decisioning, all these issues of fairness, AI, you want to share with us a little bit where the landscape is and how that, uh, affects what you're building if it does.

Tamara Lane: Yeah. So right now, in the us, unlike the eu, which you're very lucky, there's one kind of overarching AI regulation in the eu, which is very nice, in the us you have a lot of popcorning regulations coming up. And my personal opinion is either there needs to be no regulation or there needs to be one regulation. But insecure regulations that are popping up in different jurisdictions are really instable for innovation. Right? Just say when that's hap. When that happens, investment goes down in those areas because investors are waiting to see what will happen with regulations.

Host: Right.

Tamara Lane: And so then you've got a lot of good innovators with technology that it's being stifled. So I am a huge fan for clear regulations, and I'm lucky. Being in a field of fintech, we're already a highly regulated space, so we have a lot of guardrails to work off of, which actually make my life easier because you've got those, those clear things that you can say, yes, we are absolutely compliant with this. We are absolutely compliant with this. Right, so one of the ones you're talking about, I believe, is SR11 7, which is one of our leading regulations. And so just to give some plain language around this for anyone who is wondering, it says that if you use models to make important decisions, you must know what they are, how they work, how they can go wrong, and how you are monitoring and challenging them over time. So it expects strong governance, ownership approval, independent validation, all of these things. Which are incredibly important when it comes to decisioning. So that's, that's what you were talking about, correct?

Host: Yes, yes. And, and now uh, that you mentioned that My question is your agents that can be plugged in to ah, an LOS or create the LOS stack from scratch, are they model agnostic? Can the uh, each bank say, hey, I want to use this large LLM behind this? Or how are they set up?

Tamara Lane: Technically we can, but we try and make that decision because we have strong opinion. I won't say my strong opinion on it at the moment, but we have strong opinions on this and we are building out our own SLM just because we, we believe that will be a better solution in the long run. But as a startup, right now we are leveraging one LLM that we've picked.

Host: Okay. But eventually you do want to train an SLM that is specific for this. Yeah. Interesting. And have you done pilots with banks?

Tamara Lane: We're in a pilot right now.

Host: Okay. Okay. So I assume you don't have yet some uh, learnings to share with us from this pilot phase.

Tamara Lane: Check back with me in three months. I would happily come on and tell you my learnings.

Host: Okay.

Tamara Lane: Which I know will be a lot because I believe, I really believe in strong feedback and that's the importance of building longevity in products. Is that really strong feedback and analysis? And the beginning phase?

Host: Yeah, yeah, yeah. What would you, what are you expecting from this? What are the KPIs or the metrics that you're looking for to say, okay, this is working, how it should work, how we want it to work. What are they?

Tamara Lane: So our main KPIs are acquisition rate. So are we improving acquisition rate by 10 to 25%? That's like the lowest bar that we are going after for our pilot risk. Are we decreasing risk by 15 to 35%? And then orchestration. Orchestration is huge. It's, it's the time. So are we improving the, the overall. Yeah, the overall timing of the loan. Right.

Host: So are we getting processing from, from the minute there is a touch point with the customer until that all these agents deliver what is needed to the human that makes the decision?

Tamara Lane: Yes. So are we just really, right now, some of the banks that we're talking with their rate to getting a loan to a customer is probably triple or quadruple what we believe we can do. And mortgages in the US are a prime example. A mortgage should technically take about eight days, but on average it takes 30

Host: to 45 days despite the digitalization that has happened.

Tamara Lane: Yes. So that's one of the main indicators that we're looking at. Those are our three big ones.

Host: Interesting. I like to be clear on what are the KPIs and not just because, uh, we can get obsessed about the cool technology and the agents. And I have an extra agent for this too. We can really get excited in this phase about all this and really forget what is the business outcome that we really want to achieve. And I imagine eventually there'll be enough data to see whether the pie has been enlarged by having more deployments in the market overall and hopefully better risk management. Before we close. Tamara, do you have a vision, an idea of what uh, may happen in the next, let's say five years in terms of agentic lending, will it be, I don't know, will vendors, LOS providers all become agentic? Will core banking providers also become agentic? Is this something where do you think it will become a standard and mainstream and maybe even commoditized or not?

Tamara Lane: Oh, I definitely believe agents will be incorporated into all systems of lending. I think that you're going to see right now, uh, I was looking at the rates of adoption of AI just in general. And we've had AI for several years longer, but just like in the ethos of every day several years, right. And it's only 53% adopted in the US by consumers. So if you think about that, only 53% of people use AI in most generally as a chatbot. Right. And it's only 12 to 18% of businesses. So with that as the indicator of how things are growing, I believe in the next three to five years we're going to see maybe 60% saturation in businesses. Right. And then, and with consumers you're going to see more use of agents. But we still have a long way to go. So as fast as we think being embedded in AI, oh my gosh, it's moving so fast every day. It still is not the way that most people operate.

Host: And when we look at the big AI players like OpenAI and Anthropic, they do have a distribution problem, especially at the enterprise level. It's not happening as fast as they would like, as fast as the technology. We have more powerful M models every two or three months, but the penetration in the businesses is not there. It's not happening at the same pace. And uh, I was reading a report by Cinch today saying that over 80% of AI agents that were in production this is and not in financial services industry wide, have been pulled back because the business realizes that they're not really Ready?

Tamara Lane: Yeah.

Host: There's many things that need to come together to extract that value because also the reality is that you need to stop.

Tamara Lane: Uh, no. There was a fox outside my window. Oh really? Yeah. Am I able to just look really?

Host: Yeah, yeah, yeah,

Tamara Lane: Yeah. It was just so wild. A cute red fox just went right

Host: up to my wig.

Tamara Lane: That has to be good for us because. Yeah. And so sorry I got. Because I was like oh my gosh, it's right at my window.

Host: So I, I assume you haven't seen one before otherwise you wouldn't be so surprised.

Tamara Lane: Yeah, not next to the house. So yeah, so sorry, go on.

Host: But let's go back to where we left the conversation.

Tamara Lane: Adoption.

Host: So enterprise adoption is lagging. That is clear and I think again that lending is a very sensitive with a lot of social impact area, uh, of the business core of course area of banking. And we have to get it right there, there is no room for.

Tamara Lane: And every bank that I talk to knows that they need to implement agents and knows that they need to implement AI. They just don't know how.

Host: Yes.

Tamara Lane: And so that's the clear trend. And so it's not whether or not they're going to adopt it, it's just when and how.

Host: When and how. Yes. And how to start. Yeah. And your approach uh, of AI prepares human decides also gets rid of the issue of the. How regulators will handle this. Of course there's always the question even for the preparedness if an AI agent goes rogue, you're hoping that the human will catch that 100%.

Tamara Lane: That's why that's the human intervention is so important and our agents are policy bound so there's not going rogue. They can't go rogue. Right. That's the difference between, I keep getting asked the difference like why wouldn't a, uh, bank just use Claude and build their own agents? And it's because, because that's wildly insecure and you have to monitor it and you have to have a team to monitor it and you have to. Building your own process is hard.

Host: It's laborious, very hard because there's so much contextual information that needs to be checked every time that it really doesn't make sense.

Tamara Lane: You're responsible for the security, you're responsible for the upkeep, you're responsible for all of the checks and balances. Right. And so for I would say 90% of the financial institutions they don't have huge IT teams. And so the argument that software is dead I think is incredibly wrong. Enterprise organizations need other people to Manage the heavy lifting of creating technology and software. They can't do it all in house.

Host: Yes, there's very few. And actually the other day I was listening on some discussion that Mark Anderson had and was saying something very similar or related to this, that there will be jobs. We need a lot of people to help medium and small businesses to integrate AI. It's there, but people don't have the time or the capacity or how to integrate. I'm seeing a lot of people that are taking on that role. Let me go and help this type of business, be it size, be it in certain industries, to integrate the tools, not building the tools. And what, how do you decide what you need and what you don't need? Right. And then cost optimization. There's so much that needs to be decided and it's not evident. Just saying, okay, the capability is there. It's not so easy. And we can see it in our personal life, in how we adopt the tools. I see it in my business too, as a solopreneur. The way that I progress and decide what tools to use is not linear. It's not easy.

Tamara Lane: Right, Yeah, I feel you on that as well.

Host: Yeah. Uh, and you can imagine in a company, in a business how that is also, uh, reflected. So in closing, empower, who are you trying to empower?

Tamara Lane: Oh, uh, it's a double sided question. I am empowering lending institutions to serve the underbanked and to serve their customers in the best way so that their customers can thrive. So we serve the banks and in my heart we're serving the new economy that's coming up that should be given great financial access.

Host: And when you say the new economy, are you more thinking about gig workers? Are you thinking about lean companies? What is your vision?

Tamara Lane: I see the new economy as gig workers, creators, Gen Z, thin filed borrowers, nomadic workers. There's a new economy that again, it's about half the U.S. economy. And in Europe the numbers are similar. In Africa the numbers are even more booming. That work and live and bank differently and we want to bring them into the financial system.

Host: Tamara, thank you so much. This is a really purposeful mission and worth putting your energy in building this. Thank you.

Tamara Lane: Great conversation.