The Promises and Pitfalls of Agentic AI for Small Business Lending

Small businesses often struggle to access capital. Traditional business loan approval processes are slow, subjective and oftentimes rely on outdated models or incomplete data. Using AI to improve loan decision-making can be a game changer for both lenders and SBBs. Our panelists will discuss how AI is transforming lending through automated data collection, enhanced risk assessment, and faster loan approvals. The panel will also discuss the potential risks to using AI in lending decisions including algorithmic bias, data privacy concerns and a lack of transparency about lending decisions. 

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

Bailey Reutzel (00:08):
AI session for you. Everybody's excited to talk about AI, I assume. AI has been around for decades. You all have seen Blade Runner. Wooz. Wooz would be acceptable here. Blade Runner, Metropolis. Minority Report.

Welles Borie (00:25):
Yep.

Bailey Reutzel (00:25):
What about WALL-E? Any WALL-E fans? Oh my God, that robot will make me cry every time. And then what about Scarlett Johansson's sultry AI voice in Her? No? You guys have homework. You have to watch these movies. We are talking about AI today. Like I said, it's been around for decades, so we're going to try to explain what's the hype now and what's real now. Lots have changed. And so we're going to start there, but let me introduce my panelists. I'm Bailey Reutzel. I am the senior director of live media at American Banker. And then let's go down. We have Welles Borie right next to me. He is the principal product manager at Amount. Then we have Jagdeep Dayal, Chief Credit Officer at Optum Bank. Next to him is Thomas Ritchie, Chief Financial Officer at Biz2X, and Paul A. Pavlou, Dean of Miami Herbert Business School, University of Miami. All right. So we're just going to set some groundwork here. As I said, AI has been around a while. What has changed here? Jagdeep, I think I'm going to start with you. And remember, we're at the Small Business Banking Conference, so maybe we can direct it towards small business banking as well.

Jagdeep Dayal (01:37):
Absolutely. So I think if you look at it—and I'm a banker by experience and background—banks are taking an incredible amount of... they're investing a lot of money, both in people and technology, to bring AI to the forefront in lending. Now, it is first being applied to consumer and small business lending. And it has various aspects to what the banks are investing in. It's how you originate new credits, how you bring productivity and fast decisions to small business applications, and then how do you manage these applicants ongoing so that you can keep track of payments, what the other needs are, and so on and so forth. I'll add a couple of quick interesting pieces. I generally put them in three categories. The first one is using AI machine learning to originate new credits; it's a big area of focus for the banks.

(02:30):
The second area is around productivity. How do you bring efficiency and fast, cheaper products to the marketplace? And then the third place is customer experience. So a lot of AI tools are being used to drive ease for small businesses to understand: what products do you have? How does it fit what you're doing today? What is your need? And really sort of being hyper-focused on providing solutions to small businesses.

Bailey Reutzel (02:56):
Wells, I'm going to pass it to you as well. We're sort of here discussing there's AI decades ago, there's GenAI—which sort of started with the emergence of ChatGPT—and then we also have Agentic AI.

Welles Borie (03:17):
Yeah. So as you mentioned, AI is a 70-ish year old technology. Its earliest form was machine learning, which was largely confined to the space of data science where it's really good at classifying and spotting patterns that humans otherwise wouldn't be able to. What's different now is, on the one hand, you have generative AI, which can read and write. And then on the other hand, you have agentic AI, which can reason and act. And so that just opens up the space of applicable use cases vastly beyond the more esoteric data science realm to lots of different applications in that end-to-end loan journey that Jag had outlined.

Bailey Reutzel (04:00):
Yeah. And then Paul, I'm going to pass it over to you. I know you all have done some research in this area. Tell us about some of the research that you found. What are banks and small businesses using the Agentic AI for?

Paul A. Pavlou (04:12):
Absolutely, yes. So as you said, this is definitely older technology, but the speed of processing, the sophistication, and the ability to run much more complex models has changed. And the most important part for agentic AI is the ability to learn by itself. So you can actually have an agent that you provide some basic training tools and it has, because of generative AI and the large language models, the ability to learn given certain dependent variables. Because we have this tremendous processing power these days—and that's why we're building these amazing data centers left and right—basically we can train those tools to be extremely sophisticated and they can beat any human being, any banker, anybody here. And increasingly, they can do things that human beings used to do. So I'll start obviously from the basic efficiency in terms of processing, whether it's underwriting or whether it's getting data.

(05:26):
Literally, theoretically right now, they can process any amount of information, any amount of data, and they can search theoretically everywhere they have access to. And that's where the control versus the open models can find this power to find any data around us and learn from those data as long as they optimize what they were designed to do so. And of course, depending on how you structure and design those tools, you can change everything accordingly. So they can learn to even optimize certain things. And of course, we're going to hopefully get into explainability, ethics, and how to better control them. We have a very powerful technology that gets more powerful by the day and by the minute if you think about it, plus it learns how to be much more sophisticated than us. So there is tremendous power there. And then at the very least here is how we augment our human intelligence with this very powerful tool.

(06:27):
So that interaction is a big promise and hopefully we can cover it in these and other sessions.

Bailey Reutzel (06:34):
Yeah, of course. I mean, look, you're saying it could be any banker. It could be me as the moderator up here. It's going to take all of us out of a job. So Thomas, I'm going to pass that one to you. How do you see this connection between the humans—"humans in the loop," I'm sure you all have heard—and the agentic AI that you're deploying?

Thomas Ritchie (06:51):
Yeah, sure. I think looking at it from a competitive standpoint, if you're not improving the merchant journey, reducing the amount of time they spend interacting with us, making it more seamless to them, making better decisions around credit, and delivering returns to our investors at more appropriate risk-adjusted returns and making those decisions quick... so merchants are on a straight line to funding. That's how we think about application of AI. At Biz2X, we take in 25,000 applications a month. The amount of data, the amount of things that we need to sift through quickly to make good decisions, it's a lot. So we talk about the interaction of people and how we're thinking about how we deploy our people. We look at it a little bit differently. We don't think that this is going to substitute necessarily for people. What it will actually do is the opposite and allow us to address those 25,000 applications, as I said, more quickly, adjudicate for risk more precisely, and effectively get our merchants funded with much more transparency and much more speed.

(08:07):
And so that's how we're looking at it. It's a force multiplier as opposed to a cost reduction tool. Obviously all of us in this room are somehow related to small business and small business financing. It's a gigantic TAM (Total Addressable Market), right? I don't think we even understand how large it could be. So I think we're thinking about it in terms of top-line growth or growth of financing as opposed to sort of compressing our cost structure.

Jagdeep Dayal (08:31):
If I can add just one thing to what Tom was saying: in a previous panel, there was a question around how a small business owner knows what a banker is going to need. I want everybody to think about it this way: it's going to change. Agentic AI becomes your assistant. So the relationship is still owned by the banker, but the information and the speed at which that assistant—the Agentic AI—can help that banker in providing a solution, getting the right information to be underwritten at speed, and at a speed that we haven't seen yet in banking is what's going to radically change. From the inside looking outwards, there is a ton of investment in banks going on behind the scenes where the relationship manager doesn't change, but how they operate is going to change with agentic AI.

Bailey Reutzel (09:21):
Yeah, 100%. I mean, we just had our most powerful women in banking conference last week in New York, and several of the big banks—BNY Mellon was one of them—in their payments department, they're saying they have 12,000 people building AI agents right now and doing actual transactions for some of their clients as well. I'm wondering, it does feel like a data problem or solution where if you have a bunch of siloed data throughout your institution, it's hard for that agentic AI or AI broadly to make the most out of it. And so I'm wondering how you guys are thinking about unlocking some of those data silos so that the AI can be as beneficial as it needs to be. Wells, I don't know if you want to—

Welles Borie (10:09):
Yeah, I think what's exciting for us at Amount having recently been acquired by FIS is we're focused on the origination journey. So from application to decision to close, but prior to that acquisition, we'd have to build a lot of bespoke core integrations and whatnot. And so we didn't really have that bidirectional data feed. But now what's exciting about being part of that broader ecosystem is we can actually have that full customer lifecycle journey from app to close and then servicing and beyond, which can be used for everything from cross-selling to pre-qualified offers to spotting risk signals early on and mitigating those. So I think that's one area for us as an organization where we're seeing those data silos be broken down. And the other point I would make too, which is interesting, I think also for Biz2Credit, is Amount comes from a legacy of a number of online direct lenders.

(11:03):
And so with that, we had a lot of human operational folks focus on that end-to-end conveyor belt. Because of that, we have a lot of training manuals and standard operating procedures (SOPs) that we built for ourselves; that's exactly what you need for training Agentic AI. So I think for FinTechs that started as online direct lenders that wrote manuals for themselves, we have our own rich data—proprietary data—in the form of SOPs that we can then use to train these Agentic AIs to serve as copilots. And the last point I'll make is echoing what some of my colleagues on stage have said: it's not about replacing human judgment, it's about amplifying human understanding. It's having a copilot that can actually execute workflows and serve as the supercharged, ever-present, ever-patient intern that can sit there and help you with what your day-to-day is.

Bailey Reutzel (11:58):
Yeah. I want to come back to the interns at some point, but I'm going to pass it to Jagdeep first to talk about the silo busting. I think inside of a bank also, there are reasons why the silo existed. So are we happy to share information across some departments? Take us through that.

Jagdeep Dayal (12:15):
Yeah. I think this is an imperative at all banks that the information is shared across business units. I'll give you two examples. A small business owner that is your customer is also a consumer who has banking needs. And usually at your typical banks, you'll have consumer businesses, small business lending units, and otherwise. The larger the bank, the bigger issue it is in solving this issue. How do you break down silos and bring all this customer data into unified databases? That is starting to happen. What I would tell you is the banks that are moving fast are leveraging state-of-the-art external partners, whether it's FIS, Biz2X, or others, that can provide those cloud-based services and they're leveraging that to bring that data into a secure environment. Remember, in banking, security is really, really important, but as long as you can do that, what it enables you to do is to actually use that information to create more solutions for your customers—whether it's your small business owner or the consumer—and be able to bring it to market faster.

Thomas Ritchie (13:22):
Yeah, I was going to say this silo issue is an enormous issue for us, and partly that's because of what we do and where our business is, but we've got numerous origination channels. Those include people who come to our app or our website because they saw some of our content or they saw us on TV. We have got a significant set of digital marketing channels where the spend and how we think about those could be exceedingly complex. And then we've got a whole host of embedded and referral channels that folks come to us with. Tying the efficiencies in how we treat those channels to our underwriting and to our backend—effectively our backend performance or assets—it gets very difficult. It's something that certainly outstrips the human mind to be able to latch those things together and optimize.

(14:20):
So one of the things that we're quite excited about is: can we be more, frankly, reflexive in our origination channels to optimize those—both for CAC (Customer Acquisition Cost), for what we're asking for from data, and frankly, what we're making for offers to merchants coming at us through those channels based on what we're seeing in data in other parts of the business? That is something that we're really trying to lean into quite a bit.

Bailey Reutzel (14:45):
Yeah. And Paul, I wanted to pass it to you. Again, the research that you've done, what sort of metrics are we tracking when banks deploy agentic AI without silos, et cetera?

Paul A. Pavlou (14:56):
Well, let me talk about the data science part, which is very important as everybody mentioned. What is the early advantage of a small bank compared to the BNY Mellon that you mentioned? Obviously the quality of the data is so paramount. People forget, when talking about AI and agentic, it's all about the data that the AI gets access to, especially in a banking environment that is heavily regulated. So it's not going to be opening it up to the web and finding the most appropriate data; it has to be very controlled and very well specified. You have accountability and explainability regarding why you're making certain decisions. This is extremely important, and this is what we've been teaching all of the data scientists and analytics students. You have to be very careful about what data you feed into any model, whether it's AI, agentic, or any other kind.

(15:48):
On the data silos, I think that's one advantage that the small businesses have in general. Compared to huge banks that are very intentionally bureaucratic and very hard to share data, basically here you can be more competitive in that regard. Obviously, you don't have the ability to have 12,000 people designing your AI models, and understand that small businesses have some financial constraints, but you have the ability to have a sea in your data where we can find better ways to create loans and have a competitive advantage. Why would a customer come to Biz2X and another small bank as opposed to going to these behemoth companies that can do it potentially more efficiently? And that's where I think the competitive advantage is: trying to overcome the data silos and finding better ways to create and originate loans that somebody else cannot do, taking advantage of this new technology.

(16:50):
I can get into details here. AI is very contextual and very specific, and we need to figure out what data we need, leverage it very carefully, learn how to optimize certain things, and allow you to be more competitive as a small bank.

Bailey Reutzel (17:06):
I'm just wondering, by a show of hands, who is building AI agents within their business right now or personally even? Okay, yeah, there's a few of us. All right. So these are going to be maybe the nitty-gritty details here. I'm wondering what is the timeline from thinking about deploying an agent for a specific reason to actually deploying that agent? Maybe proof of concept through... is there a range of timeline, Wells?

Welles Borie (17:37):
I mean, we've deployed proofs of concept in days. Now, we're not deploying that across the board to all of our banking clients and whatnot. Of course, we need permission and there are regulatory aspects to consider. But in terms of just working with a developer and doing a mini sprint and working on an AI-based fraud risk signal web scraper, that type of agent can be deployed in a matter of days. So the gating item is not necessarily the technology or the SDLC (Software Development Life Cycle). It's more just around safety, governance, and controls, and making sure that your clients are ready to adopt because you have clients that are at different levels of receptiveness. Some are right at the bleeding edge and say, "Give me that thing you just built. I don't care. We'll figure it out with our compliance team." And then others are like, "No, AI is a four-letter word to me.

(18:32):
I don't want to see it yet. Get back to me later."

Bailey Reutzel (18:35):
Yeah. Any other timeline?

Jagdeep Dayal (18:36):
Yeah, two things: they're happening today and it is an evolution. Paul mentioned that the more the data, the better the agent is going to become. So it evolves over time. I'll give an example—and everybody's trying this, not only the FinTechs, but the banks are trying it. There's a very large bank that is known for creating it on the consumer side: think about a car buying experience. The way they would do the car buying experience is they would say, "What kind of car do you want? An SUV, a sedan? What's your price point, some other characteristics?" It's all click, right? You click, you answer, and then they'll say, "Well, here are some of our dealers—which are small businesses—that we can connect to you which have inventory." That has now evolved to being a dialogue. There are no clicks.

(19:24):
You can have a dialogue with the agent. The agent is going to be able to find you very relevant ones that are available in your area, make a reservation for you at that dealership, and you're all set. Now, it's early days, but they're collecting so much information that they're able to build this AI. Subject to, as Welles said, the whole compliance and other things that the banks have to work through, but it's happening. It's just going to evolve and the agents are going to get better with more data.

Bailey Reutzel (19:56):
Yeah, that is a really good point. At first, it might be hard to see the huge benefits, but over time it just gets more and more. Thomas, did you want to add here?

Thomas Ritchie (20:04):
Yeah, I would just say there's so much potential for AI, but it is a little bit of "walk before you can run." There are easy applications that we have done that take days or weeks, and then there are very—I call it deep—hands-off-the-wheel type of applications that are going to be quarters or years before they're usable, or we feel comfortable doing them, or the folks that we're selling assets to are okay with us taking our hands off the wheel on some of these data analytics. So I think for us, the tricky thing is: where do we take our resources because they're finite? Where do we take them and apply them to walk and then to run? And that gets very tricky because you can imagine across the organization, people have got different opinions on this stuff and they are thinking about: how do I enhance my part of this organization?

(21:02):
But you kind of have to be a little bit more holistic about it.

Bailey Reutzel (21:04):
I think there's a lot of interest within banks to build AI agents right now. I'm not sure if they're tracking all your metrics as you're building these agents, but I am sort of interested, and maybe Paul, I'll pass this to you: what's the timeline for seeing ROI (Return on Investment) from these builds?

Paul A. Pavlou (21:23):
Well, yeah. I mean, there was this informal study by MIT saying that 95% of companies don't see a return on investment from AI initiatives. That kind of examples the early stages of the internet 27 years ago: you cannot always... future investments, but no return. Obviously for the companies that actually provide the solutions, they already made money, so that's a good thing. But I guess in our case here for the clients of this AI, when do we expect to see the return on investment? Obviously the idea is to use the forecast initially on efficiency aspects—how we can automate certain aspects so we can see more immediate results. If you trim five minutes from every loan decision and you have 25,000, you do the math: how much you save on human capital, you can calculate the ROI.

(22:20):
And regarding the decision part, that would be a little bit more challenging, but if you say, "I improved the decisions through AI, we have fewer defaults," then you can calculate the ROI depending on your volume. And then the next thing is more like the quality of the overall experience. For example, one thing that is very important for small business is the notion of personalization. That's how you compete with the big players. So how can you use data to personalize more for my clients? From the 25,000 people, you identify those who can make a better decision so you can serve them better. For those who will be your client, how can you personalize it so they feel, "I'm working with a small business because I get that personalized attention," as opposed to going to any huge company out there?

(23:10):
So I think those are the three aspects: efficiency, quality of decisions, and long-term loyalty in terms of where you can see the ROI. And everything is very contextual, as I mentioned. All of us need to make our calculations how this will pay out in terms of ROI.

Thomas Ritchie (23:36):
I was going to say—side by side with this ROI thought is the "build versus buy" decision. We've got a fairly large engineering staff, and we have historically been a builder rather than a buyer just because we've grown up technology around small business lending in the US. Historically there hasn't been a ton of great tech to do that. This will be sort of the inverse of that. There's some amazing AI that has already been developed. The way we approach it is we're looking to see what is offered to us and then looking to see what we need to add on top of that to effectively get an ROI a little bit quicker and to be realistic about the tech burdens of building in this environment.

Bailey Reutzel (24:27):
Yeah, sure. Jagdeep.

Jagdeep Dayal (24:29):
Yeah. Actually, Paul hit the nail on the head. There is some very interesting growth behind the scenes inside the banks. The first one is, as Paul said, everything is focused on efficiency because you can show an ROI. But in the last year, the whole discussions have changed. Yes, that's important, but what are we going to invest in where we're actually bringing in customers that don't come to the bank today? How do we make that experience better? Because the ROI is in the future, it's not today. And more and more of my colleagues across the banking system that I talk to, that is happening very fast. And then second thing: to break down the silos, banks are creating these horizontal organizations. For example, they will be on top of deposits, small business lending, and consumer lending, and bringing everything together. It is moving in a very different direction, and those discussions are changing rapidly.

Bailey Reutzel (25:23):
Yeah. Wells, did you want to add?

Welles Borie (25:24):
I was going to say, it's also important to think about whose ROI. Is it the small business? Is it the client? Is it the company? Because I can drive a really powerful ROI for our clients and our small businesses, but if that solution scales rapidly and the inference costs start to rack up, then you're underwater immediately and your unit economics are completely blown out. So it's really important to just think about whose ROI you are really serving here. When you're a hammer, everything looks like a nail and you don't want to be a solution looking for a problem. It's really about focusing on the friction, not the feature, and trying to map that onto inference costs of today, but also projecting out to what that could look like in the future.

Bailey Reutzel (26:11):
Yeah, for sure. So I'm going to take it dystopic now. We've sort of talked about all the fun things you can do with Agentic AI, but if you have seen any of those movies—go watch those movies—they get pretty doomer pretty fast. So let's talk about it. AI in your banking institution, certainly there are a lot of data security risks. Certainly there are explainability risks. Can you explain why this AI made this decision, especially in business lending? If you start to give AI the ability to make those decisions, you will have to explain that. So I guess a broad question: how are we thinking about the risk of AI while we are deploying this? Jagdeep, I'm going to start with you.

Jagdeep Dayal (26:57):
Yeah. There are two or three things you have to take into consideration when we are doing AI work. The first one is security: data security and data integrity. A lot of people forget about integrity of data, and so you have to be very focused on that. The second area is explainability. If I am going to use an AI with a customer—a small business owner—can my models explain the outcome? It is a very, very important topic that has to be solved and is being solved. For example, when machine learning models are built to make an underwriting decision, one of the most important things is it shouldn't be a black box.

(27:42):
You have to be able to explain to the customer why you were denied for credit so they can understand it. Your compliance teams are going to ask for it. Your regulators will be all over it. I always tell people: regulators are a few years behind us in grasping this. They're well-intentioned people; they want to make sure customers are not harmed and that the banks are not harmed, but they're a little behind. They want to make sure that what you do is not a black box and that it really puts out information that is understandable. So there's a lot of focus on integrity of data, consistency of data, and explainability behind everything you do in AI.

Bailey Reutzel (28:22):
Wells, I know you have something to say about black boxes.

Welles Borie (28:24):
No, I saw a funny comment the other day, and I'll just read it: "Paleolithic emotions, medieval institutions, god-like technology... what could go wrong?" I think some of the concern around non-determinism, for instance, I find a bit ironic because we've been using non-deterministic agents since the dawn of labor itself: humans. Unlike humans, AI agents, I can guarantee, will actually read the training manual literally word for word. And not only will they read it word for word, but every time they take an action, they have the whole corpus of that manual in their working memory at all times. That's not to understate some of the risks. Obviously, security, governance, and explainability are all super important. But on the lending side, I also find it interesting: the most important of the "C's" of credit is character. Well, how do you quantify someone's character?

(29:22):
That's a pretty wishy-washy, non-deterministic kind of thing to assess. We face a lot of hurdles with the regulators, of course, getting there, but I think trying to frame it and tie it back to what we're doing today with humans—non-deterministic agents already in the loop, already making decisions, sometimes not more explainable than "Jim's a great guy"—I think we'll get there. But again, you have to have security, governance, and explainability top of mind.

Bailey Reutzel (29:56):
I'm just wondering: in an AI world where you have a machine making most of your decisions, does it become so bureaucratic that it's almost impossible for exceptions to be handled? If I actually have a legitimate exception, I can sort of play to my human agent's emotions or their common sense, whereas in the agentic AI world, I cannot, or potentially could not.

Welles Borie (30:23):
You're still going to have, like you do today, overrides. I think it comes down to tracking and accountability. The other thing with AI agents is you can have a twenty-four seven auditor or compliance person. Instead of an auditor pulling a sample of half a dozen out of a thousand loans, you can have an auditor agent literally look at every single one of those thousand loans to spot issues. So on the one hand, AI can scale inherent bias in its training data, but on the other hand, it can mitigate against bias more than any other tools we've had to date.

Bailey Reutzel (30:58):
Yeah, that's a really interesting point. When you brought it up on the call, I was like, "Oh, I hadn't thought about that yet." So that's very interesting. Thomas, I'm going to pass the question to you.

Thomas Ritchie (31:07):
I'll go back actually to what you asked me earlier, which is: one of my biggest concerns is about the people in our business. They need to buy in with the direction that the company is headed regarding the tech strategy. If you've got new technology that has the appearance of replacing people, you're going to get non-participation by those people and you're going to get reduced energy turnaround from the folks who work for you. Eventually they may go look for another job and quit, leaving you without skilled employees. I think that's a tricky messaging point that we are focused on: we need for... I'll use underwriting as an example. We have a lot of very skilled, very tenured underwriters who are going to need to buy in to the force multiplier that AI can deliver to the business.

(32:03):
They need to be the overseers of these models and these decisions. There will be some sort of conversion time that, again, could be years, but those people need to be engaged in that process. So that's one of my biggest concerns: losing them and losing their energy.

Bailey Reutzel (32:25):
And you sort of... go ahead.

Welles Borie (32:27):
I was just going to say one other broad risk here that we haven't touched on is just the AI bubble and what happens if the music stops. I was listening to an analyst the other day talk about how the AI bubble is 17 times larger than the Dot-com bubble. It's four times larger than the '08 subprime bubble. From a tactical perspective, how do you cope with that? It's making sure that you are building for interoperability and swap-ability with different models that can come in and plug and play. I think it's also important to make sure that you are building in these "break glass in case of emergency" levers so that you can turn the AI off or pull the AI out of the loop so you can revert back to your humans.

Thomas Ritchie (33:14):
Yeah. I think my second thing is exactly that: we are lenders, right? You have to have some discipline not to turn yourself into a tech company because that is a different business. As lenders and bankers, our institutions should have a very, very low rate of failure. Tech companies go bust all the time. I want to be one of these and not one of those.

Bailey Reutzel (33:41):
It does seem that banks are headed in that direction or they're failing more often than I remember previously. So Paul, I want to hand it over to you.

Paul A. Pavlou (33:51):
Yes. Three things I'd like to reflect on. Obviously, I cannot stress enough the importance of transparency. Explainability is extremely important, especially because of regulation and your customers. But ultimately, I think everybody needs to understand—from the individual to the institution—that you are ultimately responsible and accountable for your decisions, not the AI or the agent. We tell our students in the first class—I will tell every employee, every faculty, everybody—that the buck stops with the institution and the human being. Even if you fully automate, you should be accountable for what the AI does. The second thing I'd like to reflect on is the skill aspect.

(34:45):
In banking, there is a shortage of high-quality people. But it goes without saying: if an underwriter can do things 20 times faster, probably you need 19 fewer underwriters. People realize that; they're not dumb. In every single institution, if you can automate what they're doing, or if I can do my job 10 times or 20% faster, then obviously people will lose their jobs. Managing that part is going to be the next five to 10 years in every industry; I think banking will be the ultimate aspect. And then in the foreseeable future, there is this idea of autonomy—people feeling they still have power. We just allow the person making the decision, and on the customer side as well, to feel like, "Okay, I'm dealing with a human being."

(35:42):
Even if the radiologist will actually read my scan and tell me if I have a tumor or not, ultimately I want a human doctor to convey the good or bad news. In that sense, the same thing applies here. How could we empower people in the meantime? That goes to the personalization I mentioned: why the customer will feel, "I want to work with this bank and not some other one," because obviously I'm getting that human touch. I think the foreseeable future is something that will still be there until we automate everything around us.

Bailey Reutzel (36:14):
Fair enough. Well, we only have a few seconds really, but I am interested in keeping this train of thought going: for the folks here, if they want to keep their jobs, what new skills should we be learning today? Plumbing? HVAC? Yeah, fair enough. HVAC is probably good. Yes. So a lot of career shifts in your future, I think. But Paul, I think you—

Paul A. Pavlou (36:44):
I mean, I can tell you more intelligently because we have this issue here of how we upskill the workforce and how we prepare our students. I would say the first thing is AI fluency is very important. But in the foreseeable future, as I was mentioning, people still need to interact with humans. Pretty much everything we said here, machines can do—including the moderator. All of us can do a better job if we have our own agent because we can get all of our information and all our experiences much faster. I didn't sleep well last night, so I cannot really act very quickly right now, but ultimately things will change. But the human aspect is very important. People want to interact with humans. We want to get a decision by a human being. I think this public speaking, what we're doing right here, the ability to communicate, the teamwork...

(37:36):
Some of these—less of the problem-solving, but more like identifying problems, critical thinking, and judgment—these "soft skills" will become much more important, and the hard skills will be done increasingly by machines so those become less relevant. In that sense, that human component in the foreseeable future—it might be three months, it may be three years, roughly 10 years—will become critically important moving forward.

Bailey Reutzel (38:04):
Jagdeep, do you want to—

Jagdeep Dayal (38:04):
I agree. I totally agree. The relationship aspect will never go away and humans are going to trust humans to convey information and outcomes. So that's not changing. Soft skills become more and more important: being able to work in gray areas and difficult situations and how you cope with that is going to be extremely important. The skills that are changing are moving more toward soft skills and a little more creativity, and less about coding. As Welles was saying earlier, looking at a policy and saying, "Do you qualify for credit?"—those skills will no longer be necessary, obviously.

Bailey Reutzel (38:43):
All right. So you all work on your soft skills, work out your smiles and your small talk here. All right. Join me in thanking this panel.