Good afternoon. Thank you, Governor Barr and Tim, for your insights. Our panel turns to one of the most consequential forces reshaping the economic landscape, artificial intelligence. We will explore AI's promise and risks for financial inclusion as this technology spreads through our financial system. As a Fed governor, I care deeply about financial inclusion because it connects directly to our dual mandate. When people have access to affordable credit, saving tools, and sound financial advice, they can participate more fully in the economy. The topic of AI and financial inclusion is central to my work at the Fed. As an FOMC member setting monetary policy, as chair of the Financial Stability Committee, and as a member of the Committee on Consumer and Community Affairs. The speed of AI adoption is remarkable. In our most recent small business credit survey, nearly half of small employer firms reported using AI, and 71% saw increased productivity. But we are still in the earliest stages of understanding how AI will reshape financial services. I am optimistic about AI's potential to demystify sophisticated financial tools. Around here at the Fed, I've been told I am too optimistic. But I'll take that. Historically, personalized financial advice and wealth management were available mainly to those who could afford them. AI could make high-quality financial guidance accessible to everyone, regardless of income or wealth. But I have been asking hard questions about AI for several years before coming to the board, and we need to continue to ask hard questions. What happens when AI tools fail? Who is accountable? How do we ensure AI does not perpetuate existing biases in lending? How do we protect consumers' data? Are the benefits reaching communities who need them the most? Every general-purpose technology, from the steam engine to the internet, has brought opportunities and disruptions. My research on innovation has shown that when access to new technologies is equitable and appropriate guardrails exist, society benefits broadly. Today, we'll cover how consumers use AI for financial advice, how institutions deploy these tools, and which risks we need to mitigate. I'm eager to hear from our panelists about the remarkable possibilities and real challenges. First, allow me to introduce them. Starting to my immediate left, we have Sanjay Subramanian, who leads PwC's technology alliance with Anthropic and the firm's global AI analyst relations practice. With over 25 years of experience, he helps large organizations put emerging technology to work. Sanjay will ground us in practical real-world applications of AI in business and banking today. Laura Blattner is head of impact at The Bike Shop at MIT. Now, you know, when I saw this, I got so excited because I used to live in Cambridge when I was at Harvard. I had a bicycle, and all those bike shops were closing. So, I thought, finally, one is opening. [laughter] Okay. And apparently, this bike shop in Cambridge is elusive because this one is The Bike Shop at MIT and applied AI lab. Still have to keep looking for them. And advises philanthropic funders on evaluation standards for AI products serving low-income Americans. Formerly a professor of finance at Stanford GSB, Laura brings the academic rigor we need to understand AI's deeper implications for financial inclusion. And then we have Sarah Willis Airtour. Sarah is executive vice president at SaverLife, a non-profit financial technology company. SaverLife has developed an AI-powered navigator tool to democratize personalized financial guidance. Sarah will help us understand how AI can serve consumers who need better financial tools the most. Thanks to each of you for being here today. Laura, let me start with you. So, the bike shop research lab here is behavioral science with AI research, which is unusual. Why does behavioral science add when we think about AI and financial inclusion? What does it add?
Thanks, Lisa. Excited to be part of the panel today. So, I like to say I keep a running list of AI products that you should not build. And at the lab at the bike shop, we work with a lot of organizations helping them test and prototype AI tools. And when these products fail, it has usually very little to do with the technology, but everything to do with how we design the product. Because ultimately what we're trying to do is we want to change behavior, right? We want to help people make better financial decisions, better education decisions, better health decisions. But it turns out changing behavior is really, really hard. So, let me just give you one example of that from an organization we're currently working with. So, this is a financial counseling organization that helps low-income workers. And when you look at the calls these counselors have with their clients, what strikes you is that, you know, maybe 20 minutes out of a 30-minute call is being spent on intake. You know, tell me about your spending. Why are you spending that much on a hairdresser? Like, tell me about your income. And so, the first thought you might have, well, that seems pretty inefficient. Can we just outsource that to an AI bot? Can we just have you talk to an AI and then the financial counselor, you know, gets the perfect summary and already knows what's going on when they pick up the phone and talk to you? But let's flip that. So, imagine you are the person talking to the counselor, and this organization you've never heard of is asking you to talk to this AI bot that's asking you very intimate questions about your financial life. Why should you trust that thing? Why should you spend your valuable, scarce time on interacting with this AI tool? And so, we quickly realized that that's not the way to go. In fact, the organization already has an intake flow that almost nobody uses when they first come into this counseling conversation. So, what we're building instead is a tool that helps with intake once the counselor and the client are live on a call. So, once the counselor had a chance to actually establish that trust, show the client what they can get out of the session, then it becomes very natural for a client to say, 'Okay, yes, I want to give you my bank statements, my credit card statements. I want to tell you about my financial life.' And at that point, we plug in the AI to help us crunch the numbers, digest in real time what the client is providing, and then flagging those places where the client and the coach can really dig in and figure out what's going on. And so, what this and many examples we have seen at the bike shop have in common is that if we want to truly change behavior, we need to figure out what stands between a person and a better decision. And if that's not, you know, like a pure information problem, but something about an underlying fear or worry or some emotional barrier, then actually building a standalone AI tool is probably just going to be an expensive build that nobody uses. And just to tell you why we call the bike shop is because we believe that the way we should be leveraging AI is actually by augmenting what a human can do. So, for example, augmenting what one of those financial counselors can do when they have a conversation with the client.
That was really smart, Laura, because when I saw bike shop, I got this warm, cuddly feeling. So, thank you all for being thoughtful about that. Um, Sarah, what are you learning about how low and moderate-income households are actually using AI for financial decision-making? Are there any surprises in the types of questions they're asking?
There are many surprises, Governor Cook. Um, and by the way, we not only share your optimism in the potential here, but we're also seeing it real time. So, for those of you that don't know SaverLife, which I find are not that many recently, um, we've historically focused on building savings and financial security across a really wide range of low-to-moderate income consumers. We're across all 50 states, we're digital, um, members to date. So, pretty large swath. Um, and when we started to think about where AI can benefit our members, the number one challenge that we saw our members facing in addition to all of the things that you economists and researchers know well, like affordability and cost of living and inequality, was the complexity of financial decision-making. Um, the constant nature of decisions, whether it's consequential, like where to go to school, how much student debt to take out, or it could seem really minor because life threw you a curveball and you don't know how to pay for your father's funeral expenses, right? And so, these decisions don't come with a user manual. And even if one could exist, people aren't going to be able to have the time and the fortune to review that manual. And so, when we deployed this, we call it a navigator, this tool across our membership, we saw a number of really surprising things. So, Laura, you touched on trust. Um, this is interesting and Tim mentioned that consumers are more likely to trust potentially AI based on the provider it's coming from. SaverLife's a nonprofit. We've been around a long time. We have a really deep relationship with our members. So, when we launched the MVP of this tool, not a single member, again, noting they were a highly engaged subset of our membership, opted out of the personalization that we were offering. So, we asked them, 'Would you like us to pull in your transaction data? Would you like us to pull in your goals that you've given us during onboarding? Would you like us to analyze how you've used our app in the past, what type of articles you've read, what type of actions you've taken.' And no one opted out. So, that was really interesting. Secondly, we noticed that people were coming back repeatedly, not just for the one-off kind of question, and that the questions that they were asking the AI navigator were not, you know, basic financial literacy questions that are, you know, something you can Google or you can ask your parents about. Um, they were 'Hey, I'm in this situation. Maybe it's debt. Maybe I'd like to buy my first home. And I want to know what to do next. I don't want 10 different recommendations that I could do. I want the AI to do the heavy lifting for me and give me that one next step.' And so, that's what we've tried to build with this navigator is really streamline the complexity, give them the one next best step. And interestingly enough, that is yielding significant results in terms of likelihood to act. So, for those that received an AI-generated recommendation, they're 10 times more likely to take up that recommendation. In this test case, it was referral to a credit counselor, like a GreenPath or an MMI type entity. And I know, you know, this is the crux of the matter when you're talking about matters of stigma and shame, which debt brings a lot of that. Um, people are usually likely to kind of put their head in the sand as opposed to take action. So, that's been really, really fascinating to see the uptake. And then our core outcomes that we've always been seeing with our kind of one-size-fits-all model, now with AI, are accelerated. So, people are saving more, $200 median increase in savings prior 3 months without using the AI navigator. Um, and they're depositing about half of them are depositing over $500. And so, again, the famous $400 Fed rallying cry of the sector, we're seeing people overcome that barrier through some really personalized, actionable, and empathetic support.
Yeah, that's fascinating. You work directly with Sanjay, we're turning to you. You work directly with the teams building today's frontier AI. What can these systems now reliably do with someone's money, and where does the hype still outrun reality when it comes to financial inclusion in particular?
It's a great question. I think, you know, I was thinking of the comment that Tim had made and Governor Barr talked about in terms of people have the need to feel competent. They want to feel they're more knowledgeable at the end. And so, I was, you know, when I think about myself, I spent the weekend, I was helping my dad with his finances, and I was able to build something to teach him not only to understand him as a person, the questions he had to deal with, but also help him understand why, understand, you know, where those sources were coming from so he could independently research and validate it, and then to understand what his options were and what the impacts of those were. Because my goal was to help him feel confident and competent, and not undermine him and suddenly become the father to the son right in that moment. So, how do I do that in a thoughtful way? But then, at the same time, I'm thinking about, and so, one of the greatest things that we've seen today is the ability to teach and to coach, but I'm able to do that because of my life experiences and what I'm able to do. So, how does someone who doesn't have that experience know the questions to ask? Know that what they're getting back is valid. And I think that's where I think there is increasing focus from a lot of the technology companies to make that more transparent, to help educate, and to make the ability to teach and help people understand something that is much more democratized. I'm seeing that with a lot of the frontier companies that I'm working with that are trying to help people understand how to minimize that digital divide that we're seeing, right? There is a lot to do in that space, right? Those that aren't doing a good job don't focus on that. They focus on the marketing. They focus on what looks good. They focus on what works in a demo. But it doesn't work for that edge case. It doesn't work that 10th time or that 20th time. And they don't follow the answers, right? So, I just think that that was so well put. You know, people want to feel competent. They want to feel in control. And then once you have that, we've got these devices on it. We can't, you know, we all have our devices. We have the technological access. We don't have the knowledge to know how to use it in the most thoughtful way. Once we do that, that's going to unlock a lot of inclusion.
So, it's interesting. I hear all of you, all three of you saying some version of information is power. And you can empower people not by telling them what to do, but by giving them the options and letting them use what is uniquely human. And that's the decision-making capacity. That's what I'm hearing. And that can be empowering. That's great to hear. So, I'd like to ask a question to the entire panel. So, I'd like all of your views for this next question, the perspective of the work that you're doing is fascinating. How have you seen AI benefit consumers in the context of financial services? You want to start, uh, Sandeep?
I think we talked about education. I think also in your living this on a day-to-day basis is that the old financial system was not built for the edge cases. It was built... I heard a great analogy the other day. It was sort of this: imagine a world in which you have every person goes through multiple choice questions. As long as you could answer A, B, C, or D, you could fit into that system. But if you didn't, you couldn't fit into that system. Right? And so, now with the alternative data that was talked about and looking at alternative ways of measuring a person, the generative AI is built and LLMs are built to understand the messy, the complicated, and piece together something unique about the context of someone's life and what they need, and then to make recommendations as a result of that. Suddenly, there's a price point that allows that to happen in a thoughtful way. So, that's why you're seeing a lot more products that are there. So, I think the economics that we're seeing is not just about access. It's also allowing markets to open up and people to be touched in a different way that they weren't before. And you're seeing that, right, in the work that you're doing.
Yeah, absolutely. I think to take it even further, the onus now is back on the provider, so to speak, where it always should have been, frankly. Um, less so, okay, we're going to help educate consumers to navigate this entirely somewhat unnecessarily complex system that we've built up. As Governor Barr said, designed around usually more affluent customers. And so now, I think in order to remain competitive and also think about lifetime value of a customer and increasingly what consumers are going to expect from financial service providers, will be to design around lower income use cases and in a way that the kind of augmentation of AI to assist with their confidence, guide them to the right direction, combined with products and services that truly meet their needs is going to be more of the winning fit. And in institutions themselves, SaverLife is quite small, but we can see when people have negative balance events, right? And unfortunately, it's on the rise. Take a large bank. Any number of the top three, they can see this data on the back end. Obviously, we've had the machine learning capabilities all along, but what would it look like if we proactively talked to our customers, right? Say like, 'Okay, it looks like you've recently had loss of income, right? How can we help?' I know some of the, they're not fintechs, neobanks, not banks, but tech companies, I won't name names, are doing this, right? And I think that is the big opportunity here is that providers have access to this data if only the business case and incentives can be organized around what is ultimately best for the consumer, but unfortunately I think that's actually the hill that we still face.
How about you, Tim? Laura.
So I think where I am most excited about the impact of AI is bringing the expertise that currently lives in the heads of very few people to a lot more people. So Sarah mentioned MMI, and let me give you another example of something we're working on. So MMI provides these debt management plans, right? Help people pay back some of their unsecured debt by lowering payments. And one of the challenges they see is that a lot of folks who should be eligible and who we think would benefit from this don't actually make it through and sign up. And so what we noticed working with them is that there are a few debt counselors who are somehow amazing at addressing these fears, working with a client, you know, talking about why the influencer on TikTok didn't get it right. And they really manage to convert people and have them sign up to these plans who then successfully complete. And so we wanted to know, can we learn what these outstanding counselors are doing, and can we help every other counselor adopt these very same techniques? So we were able to find in the large corpus of calls that MMI has already recorded, find instances of specific behavior that really seems to work. A specific phrasing of how to address a client's fear, a specific way of connecting to them about their past history. And then right now we're building a tool that takes those insights and uses AI to find opportunities in conversations to apply these kinds of tricks and tips, and surfaces that through an application to every other counselor in the organization. So, I think that's a great example of taking something that is in the heads of a few people and making that knowledge, making that skill available to a lot more people to successfully apply.
So, it's empirically based. Really is it empirically based. Um, so that is connected to my next question for you, Laura. It's easy to build an AI model that gives financial advice. It's hard to know whether the advice is good. How do you close that gap? Is this the way you close that gap by depending on empirical evidence?
Absolutely. Um, so I live in San Francisco and all our highway billboards have been taken over by various AI companies. But there's one near the Bay Bridge I actually like and I'll phrase it for the purpose of the answer, which is: without evals, your AI financial advisor is just expensive advice trained on Reddit. So, the work at the lab we're doing right now is helping organizations build these applied AI evaluations, or in the billboard lingo, we call them evals. And evals are very different from two other things I see a lot of people spending effort on. One is these benchmarks, right? You may have seen these leaderboards. Think of these as standardized tests about, you know, can our LLM answer basic financial literacy questions correctly? And the other one is sort of the kind of impact evaluation where one group has the AI, one group doesn't. Does the person with the AI tool do better 12 months later? Both of these are very useful, but neither of them really helps you close that gap, you know, build that better AI product. And so, let me give just an example of what this applied eval thing would look like, if let's say we were building that AI financial advisor. So, the first thing it needs to do is actually talk to you and learn about you, learn about your financial life. So, the first metric or the first eval I might want to build is say, how good is it at doing that? So, I might create a couple hundred use cases or user cases from the actual users I'm trying to serve. And say, okay, how well does the AI do when talking to that person and actually getting all of the relevant information? And how many questions does it take to get there, right? If my AI is going to ask you 150 questions, I've already lost you. So, those might be my first couple evals. How well does the tool do at actually asking you questions efficiently? And then once the tool has all that information, I'm going to ask how well does it do, to Sarah's point about reasoning about that information, right? Like, what's the actual thing you should be doing? Not just 10 recommendations, but a single recommendation. So, I might be building again a separate set of test cases that are deliberately made to be quite difficult. So, maybe I have one case where increasing your 401k contribution limit is absolutely the right idea. And then a similar looking case where the person looks like they should be increasing that 401k contribution, but actually they just mentioned that their car is likely going to need major repairs, and you can see that they don't have the cash flow to pay for that. So, likely they would be running up credit card debt if they were to now increase their 401k contribution. So, that would be a case where you would not want the AI to recommend increasing that contribution. So, again, you would construct very real cases like that, and you would ask, can the AI successfully reason relative to what an experienced human financial advisor would do, can the AI match that recommendation? And so you would build a whole stack of metrics that we would call evals that allow you to actually query your product repeatedly and figure out how well is it doing on all of these various capabilities we want it to have, and that allows you to both test and improve your product iteratively.
So that's really interesting. So what about reliability? So you have these AI models that are interactive and they're learning and they're giving advice, but one of the issues is doing that consistently. So how do you make sure that these tools are reliable? Using the framework that you just articulated.
Again, I would say it's an iterative process, right? So think of it as you become more trusting of your product, you see it interact with more and more users, you can start slowly pulling back the guardrails. Think of it like we test new drugs, right? When we first develop a new drug, we have no idea how this thing is going to do when given to real people. So we give it to a very small number of people and we very closely monitor what happens. And as we start becoming more confident that this thing isn't having any bad side effects, we start to slowly increase and increase the number of people that we're willing to test it with. And so I think something similar is what we would recommend, you know, start with a very small controlled experimental setting where you can really check almost every case. Then as you start becoming more confident that it's relatively reliable, you can start automating some of these monitors, right? I just talked about these applied evals, and you can have humans maybe look at some targeted cases that you're particularly worried about. And if you're still seeing good reliable performance, you can increase the user pool and even pull back further to monitors you have to rely on. So again, think of it as we learn by doing because fundamentally we have no clue about the reliability properties of these models. So the only hope we have is to try and build as many empirical evaluation tests that we can run live on the product. And then be very disciplined about growing our user base and how we're monitoring and testing the ongoing performance of the tool.
If I can, I just want to ask one quick question. Um, when you say that you're using behavioral science, are you mainly using behavioral economics or are you reaching out into other fields?
I think we want to bring the insights from whichever field that helps us build better products. Sure. And so I think we have learned a lot from behavioral economics in the last couple decades about how to design successful interventions that we believe we can bring to building AI interventions and AI tools. But if there's a smart insight from another field, I think we should learn and we should be open to any insight that we think is going to move the quality bar for everybody.
Okay, so you hit the daily double. Economists can't be trusted with everything. We're, you know, we can't be trusted all the time. We do need to reach deeply into other fields. You win. It was a rhetorical question. [laughter] I was wondering that.
I would say that there is, you know, just feeding off of what Laura painted a very clear and consistent picture in terms of how to think about this. One of the things we're seeing is there's an ability right now, like so for example, in that farmer model in order to reduce time to value of creating...
Synthetic data and basically looking through and sort of saying where do I start? Like where are the places where I think I can reduce costs and really focus and get something out there quickly. So similarly, you can create these use cases. You can and we're seeing this, the ability to create tens of thousands of edge cases and seeing where they would fail, creating automatic evaluators and trying to have different perspectives. You could have like what is a good plan? You could have a tax perspective, someone looking at purely from a tax perspective and saying this is good or bad, purely from a high growth perspective. Someone saying well no, you want to be a bit more conservative because you want to make it through those tough moments where you need to replace your car tire. Right? And so you can run all of those different scenarios and then you can start to focus on where you want to get things done. So there's the human in the loop which is key. There is backtesting against existing use cases where you're comfortable with the answers which is key, but there's also situations change, data change, models change over time and there is drift. So it's very important to continually have automatic valuation and transparency as you go through this.
You were going to say something, sir?
Yeah, I was going to add well, first of all, it sounds like Laura, as you described the best approach here, I was feeling like, 'Okay, our engineering team and our product team have done right by this.' These are all words I hear all the time. I don't necessarily always understand what evals mean but thanks for clarifying. We also came at this from an overly cautious perspective. Even though we're not making lending decisions, which I know tends to be the trickiest area to apply AI, we took it one step beyond do no harm and said, 'We're only allowing the model to pull from really trusted sources at first.' It still is only pulling from trusted sources, but we're slowly ungating it. Because again, it's this tension of you limit it too much, is it going to be valuable? If you let it go wild, nothing wild happening in this building, then again, negative things can happen. So one of the things I was going to say is while we're thinking about the use cases around the financial health outcomes that we're all striving to move forward, the use cases are just expansive beyond that, right? Whether it's public benefits, thinking about where to choose your college degree, which student debt to take out, how to prepare yourself mentally before walking into the financing office when you're buying a car. For anyone who's done that, that is a scary moment. And so we often think about how can AI be a bridge to humans as well? Enhancing judgment, not replacing it. We don't do ourselves any favor by replacing human judgment and agency. So it depends on the human.
[laughter] Exactly. Yeah. Yes.
In general. In general. In general.
So, that's really interesting because we can certainly think of different tools that could be helpful. I mean, the example that you bring up, Car Buying, has a large literature in economics. And basically, I've decided to outsource Car Buying to somebody else like Costco.
So, that's a suggestion. That's a recommendation because I'm going to... There's a dollar value put on how much I would lose if I were to do it myself. And I know how much trouble I would get in if I were bargaining for a car. So, outsourcing it is absolutely the best thing. So, that's right. It does depend on the human, absolutely. But, the Car Buying example, I think, it's really instructive. So, Sarah, let's turn to you. You emphasize designing AI to support consumers rather than replace their judgment for high-stakes financial decisions. How have you approached responsible AI and guardrails while building the financial navigator?
Yeah, I was touching on this a little bit. But for every opportunity to make economic mobility gains, there's probably 10x opportunities to also see backsliding and take risks you're not even aware of. Right now, it's been in the conversations, like sports betting, prediction markets, other things that sound like gambling are extremely prevalent and being built in a way that is riding on the existing rails of the financial sector. So when we've thought about our AI navigator, we don't want to discount someone's idea. If they're asking the robot about crypto or sports betting, we try to redirect. We go back to, 'Okay, you told us your goal was to buy a home.' and try to get them back to a low-risk personalized approach. But this is tricky because you don't want to be too paternalistic. We should balance consumer protection with opportunities in AI. The other thing we're thinking about is the business model creeping in, like NerdWallet, which used to be a trusted source of comparisons now feels like you're being pushed into products. So at SaverLife, we're careful not to take referral partners that could compromise our recommendations. We constantly vet and re-vet our evaluation frameworks for what is helpful for a low-to-moderate income audience. It extends beyond the AI navigator to who we partner with and how we pass them on to other organizations.
So, one thing I hear you three saying is that the iterative process is absolutely fundamental. The testing, I was saying to you all when we had an earlier conversation, the thing I absolutely hate is the lack of testing. And this isn't even for low and moderate-income users specifically, but the care taken is appreciated. I'd like to say I mentioned the name of a car buying service, that wasn't meant to be product placement. [laughter] We are here at the Fed, so a generic car buying service. Let me clean that up. So, Sarah, beyond engagement, what early evidence are you seeing that personalized guidance may be influencing financial behaviors or outcomes?
