The Banking & Payments Show: Will generative AI save banks?

Audio by Rob Rubin, Eleni Digalaki, and Dan Van Dyke | May 15, 2023

On today’s episode, we look at the various angles of generative AI in banking. In our "Headlines" segment, we discuss regulators' urgency to install oversight over the rapidly advancing technology. In "Story by Numbers," we examine the percentage of executives that are not aware of how fast generative AI is coming, as well as discuss an old number that has been around the finance world for a while about human customer service agents versus bots. And in "For Argument's Sake," we take sides in the debate over whether generative AI will raise the bar or level the playing field for small and regional banks. Tune in to the discussion with host Rob Rubin, our analyst Eleni Digalaki, and vice president of content strategy Dan Van Dyke.

Subscribe to “The Banking & Payments Show” podcast on Apple Podcasts, Spotify, Podbean or wherever you listen to podcasts. Follow us on Instagram

Episode Transcript:

Rob Rubin:

Hello everyone and welcome to the Banking and Payment Show, a Behind the Numbers podcast from eMarketer. Today is May 16th, I'm Rob Rubin, GM of Financial Services and your host today. If you enjoy this podcast, please give us a five star rating and subscribe. The title of today's episode, Will Generative AI Save Banks? And I invited back Head of Financial Services Research, Eleni Digalaki and Banking and Payment Show veteran Dan Van Dyke, VP of Content Strategy at Insider Intelligence and our point person on all things generative AI. Hi Eleni, how you doing?

Eleni Digalaki:

Hi, Rob.

Rob Rubin:

Hey Dan, how you doing? So happy to have you here today.

Dan Van Dyke:

Hey Rob, doing well. Thanks for having me back.

Rob Rubin:

Yes, nice to have you back, Dan. Dan, before we get into it, we introduce quickfire sessions maybe four or five episodes ago and I realized that you haven't actually done it.

Dan Van Dyke:

Its been a while.

Rob Rubin:

Yeah.

Dan Van Dyke:

Who has the top spot for these podcasts at this point? Is it Eleni? Is it Tiffany? Is it me?

Rob Rubin:

I think it's between Eleni and Tiffany.

Dan Van Dyke:

Man. Okay.

Rob Rubin:

All right. We'll have you back more.

Dan Van Dyke:

So what is it?

Rob Rubin:

So the first one is, where do you live?

Dan Van Dyke:

Williamsburg, Brooklyn.

Rob Rubin:

All right. Have you ever had a good experience with a chatbot?

Dan Van Dyke:

It's a hard one. So if we're calling ChatGPT a chatbot, absolutely. If we're talking about chatbots like IVRs, like calling your bank, sometimes. So I found one that pointed out that a gym was double charging me and I thought that was pretty cool. So I'll say yes, but the majority of time it's a negative experience

Rob Rubin:

Yeah, for now. And why is ChatGPT bad at math?

Dan Van Dyke:

That's a great question. A lot of people talk about ChatGPT like it's supercharged, T-9. So remembering back in the days when you were texting and it was auto complete, there still is that technology today.

Rob Rubin:

Yeah.

Dan Van Dyke:

ChatGPT's basic reasoning engine sits on top of predicting the next word in a sequence of words. And so that optimizes it for things that have to do with grammar and language. It means that functions like math that are super rigid and where there is one correct answer really doesn't lend itself to what ChatGPT does best and that's why it's bad at math.

Rob Rubin:

Oh, it sounds like if it needs more than one answer, it'll be good at politics. Here's an oddball question, but you know that I like to collect data points.

Dan Van Dyke:

Yeah.

Rob Rubin:

Do you prefer to eat cake with a fork or a spoon?

Dan Van Dyke:

I really want to see the spreadsheet where you collect these data points. So I prefer to eat cake with a fork, unless it has ice cream with it and then I'm definitely a spoon guy.

Rob Rubin:

There you go. I'm with you, it depends on what kind of cake.

Dan Van Dyke:

Yeah, that's true.

Rob Rubin:

Well, that was fun and we have a ton to cover, so let's get right to the headlines.

In the headlines, we chat about a top story as it relates to our episode. For the headline today I want to go back to an article we published May 3rd, CFPB and Regulators Unite to Tackle AI Oversight. It talks about a joint statement put out by the CFPB and the FTC and the Civil Rights Division of the US Department of Justice and the US Equal Employment Opportunity Commission. The US Regulators' joint approach shows the scale of AI's impact on financial services and other industries, the potential problems it poses and the speed with which watchdogs are really grappling with introducing stronger oversight. So my first question is really why the urgency?

Dan Van Dyke:

That's a great question. And I think one phrase that stood out to me from the press release speaks to the why behind the urgency, which is that they use the specific phrasing of companies that market themselves as using AI. And what is the significance of that? It's that every company, it seems like overnight has some sort of AI product or are claiming that AI is behind their product. And that can be a smoke screen for a lot of tactics or a lot of practices that are anti-consumer.

So some specific examples of those that were called out were the black box algorithms that are behind a lot of credit decisions, things like dark patterns. So if you've ever tried to sign up for a service and then suddenly you find that you've signed up for a couple of auxiliary services, that might be what's known as a dark pattern. But all these things that companies can use AI for that could have some really negative impacts from the customer perspective, those are becoming much more prevalent as companies in reality or in perception are saying that, "I am now using AI," and that's why the CFPB and all these other government agencies are moving so quickly to regulate the space.

Eleni Digalaki:

So I totally agree with Dan. The risks are significant both when it comes to abusing the tech, but biases, misinformation or all the things ChatGPT is known to not be getting it quite right. But also the tech is developing faster than many other technologies have in the past like smartphones, so they need to be moving fast on this and they're aware. The difference between ChatGPT 4 and ChatGPT 3.5 is mind-blowing, so it's really moving very rapidly. But I'll also say that moving fast is as important to regulators as it is to banks. A lot of banks won't touch any of the more sensitive use cases until they get more clarity from their regulator. So I see that as good news for the industry, especially right now with all those failures we've seen with banks, financial institutions don't want to jeopardize consumer trust. People are just ready to change a bank in no time if they get a red signal. I think moving fast is important.

Dan Van Dyke:

Eleni, you brought up such an important point that the scale at which AI is developing is so unprecedented. And as an example of that, I actually wrote my first report on AI and banking and payments when I joined this company, and that was in 2017. And since then, Moore's Law for Natural Language Processing is that the scale of large language models, so think ChatGPT, the size of them has 10X every year. And that means that when I wrote about AI and banking and payments models, then were 1,000,000th of the size that they are currently, which is just incredible in putting in perspective how quickly this has really develops. If it feels overnight, it's because it really has 10X just over the past year.

Rob Rubin:

Clearly generative AI is equal parts scary and exciting. And the thing that I think we're talking about here is the criminals seem to be as excited as everyone else about the potential, which is why the regulators are paying attention and hopefully that's why the banks are waiting for the regulators before they jump right into it. In our next segment, Story by Numbers. Let's look at the upside of generative AI digging into some of the early use cases. In Story by Numbers, we pick a number or two that helps us dig into the episode's topic.

And today for Story by Numbers, the first number that I want to focus on is 65%. And that's the percentage of executives surveyed that believe generative AI will have a big impact on their organization in the next three to five years. This was a KPMG survey data point, which was cited in a Wall Street Journal article from May 2nd, Goldman Sachs CIO tests generative AI. We will put a link in the show notes. For both of you, the thing that stands out to me is that 65% of executives don't know how fast generative AI is coming. Three to five years seems overly conservative to me. Do you agree with that?

Dan Van Dyke:

I do. There's that often repeated adage that people tend to overestimate what can be achieved in the next two years and underestimate what can be achieved in the next 10. I think all that logic goes out the window when you're talking about being pegged to the growth curve that we just described. Moore's Law for NLP is that the models are 10Xing every year. And so I feel like the impacts are already being felt. If you look at a lot of companies saying that they're actively piloting the tech, you have IBM saying that it's going to shift its hiring practices to account for where it sees the technology going. This is happening right now, it's not a theoretical thing that will happen over the next three to five years. But Eleni, what's your take?

Eleni Digalaki:

Yeah, I don't disagree, but I'm also thinking of the word big impact there. I think many banks are already experimenting with the technology, they're doing a number of pilots. But from that to having a big and significant impact at the organization level, it does require some time. Five years is definitely off though, I think it's going to be much faster.

Rob Rubin:

Maybe two to three because also what could be a big impact might not necessarily be what faces their customers. It could be some of the things that they're doing behind the scenes that really streamline some of their processes. Wouldn't you agree?

Eleni Digalaki:

Yeah, absolutely. I agree with that.

Rob Rubin:

So to me, those are the things that we might see as impactful. Some of the things which I think are scary are credit decisioning, providing any kind of guidance or advice using AI, that might take more time. What do you think of that?

Eleni Digalaki:

Yeah, I agree. And I think ChatGPT could very well take chatbots and personalized recommendations to the next level. But it goes back to that clarity and having the regulatory guardrails needed to move in that direction. They might not be allowed to give recommendations using AI at all, but I think there's a lot of opportunity for example there to have those chatbots, those tools face the employees internally and help them with their recommendations and personalizing service on the background.

Rob Rubin:

Let's pivot to this front end. And for my second number, I'm going to use a number which is basically a banking myth, I think, because it's been around forever. And it's old, I knew about it, but I think it was a Deloitte study, but it might not have been. And I'm just saying this because I don't know the source, but it's a number. And it's that the cost of a bank customer service call with a human is $4 and the cost of a conversational bot is 10 to 15 cents a call. So now you know why it's hard to speak to a human. Based on our mobile banking benchmark surveys, bank chatbots to date haven't even been close to comparing to a human experience. So when do you think generative AI is going to help banks deliver a human quality customer service experience at a fraction of the cost of an actual human?

Dan Van Dyke:

That's a great question. And first speak to that stat that you cited. I used to work at this company Javelin and we had almost that exact same stat, but it was about mobile banking for depositing a check versus using a customer service support rep. So my guardrails immediately went up when I heard that stat. But anyways, getting back to your question, when is it going to reach the level of support that you get from a human? Well, I think a human can do a few things. One is they can understand the meaning behind what you're asking and second is they can take action.

So for that first portion, helping to understands the meaning, I think that we can see a pretty significant impact over a short period of time. So an example of that would be a chatbot offered by Capital One has to think about, what are the thousands of ways that you could ask for a balance if you're a bank customer? And now ChatGPT can help figure out how to code in all those different ways that a customer might ask for something that simple.

And think about that across every different feature. So it can help take a query, point you to the feature. When it comes to actually executing complex things or taking action on your behalf, that's where I think we'll need a lot of guardrails in place to make sure that architecture like ChatGPT is actually not used for that. Because it's a little bit too creative, it's a little bit too unhinged in some cases, if you look at some of the conversations that Bard or Bing have had with different reporters, it's really not appropriate to be customer facing.

Rob Rubin:

Right. At least now.

Dan Van Dyke:

Yeah, at least for now. And the more powerful use cases will be, as Eleni mentioned, employee facing. So it still might positively impact the customer experience, but there has to be a guardrail of a human for now where hallucinations, misinformation, potential algorithmic bias are still major problems with this technology despite its sky-high potential.

Rob Rubin:

And also I think that part of it is how well is it trained?

Eleni Digalaki:

Yeah, that's a good point.

Rob Rubin:

So a bank's ability to train it to handle all these things. You mentioned Capital One. Capital One's a big bank and we're going to talk about this next. But training it for all these scenarios is a challenge, no?

Dan Van Dyke:

Huge challenge. And what Bloomberg did I think is really exciting. So Bloomberg released its own transformer model. It trains on a data set that it had, which is designed specifically for financial tasks. I think the potential to release models that are fine-tuned to optimized trains all for these tasks can have a really transformative impact in terms of improving the quality of tools that are customer facing. But that's a little bit further off into the future.

Eleni Digalaki:

Yeah, it is a challenge, but it's also the direction banks should be taken because by limiting the pool of training documentation available to a chatbot, you can ensure greater accuracy and you solve for some of the problems with ChatGPT's great imagination.

Rob Rubin:

Okay. In our report on generative AI and banking, we covered nine different bank use cases for generative AI, but today we really only had a chance to cover a few with great potential, but also obviously not with some unforeseen challenges like training, for example, and proper training and making sure that the agent isn't making stuff up. In our final segment, For Argument's Sake. We're going to argue nicely about whether generative AI raises the bar or levels the playing field for small banks and credit unions and regional banks. And for background, a few episodes ago, I posited that within 10 years, 95% of banks and credit unions today will cease to exist. And my perspective is that smaller institutions won't be able to keep up with consumers expectations for digital experiences. So for this segment, I'm going to just throw out this question and we can have a three-way conversation about it. How can generative AI help level the playing field for small institutions?

Eleni Digalaki:

I think theoretically it makes it easier. There's a lot of discussion around how ChatGPT and generative AI democratizes the technology and how it's accessible to small and larger players alike. So tech-wise, it's definitely easier. To me, culturally, it is a massive shift, it's not going to be easy. And I'm also thinking about it more broadly when I look at existing AI use cases, not limiting it to generative AI. When it comes to community banks, you see a minority of them have deployed AI, the majority might be planning to do. The picture is the opposite when you look at big banks. So in my mind, there's no good reason why the existing AI gap wouldn't be maintained when it comes to generative AI. And I think one could argue even that it will increase because the rapid advancement of the generative AI tech offers more opportunities to the people that are actually already ahead.

Rob Rubin:

Eleni, you raised the issue of training and the challenges. And I think that that's ultimately going to be hard because I don't think that banks are very good today at leveraging their first party data to begin with and I think it's all over the map and unorganized relatively. So for smaller banks, for banks that aren't the biggest ones, it might be really hard. However, on the other side of the equation, they could take advantage of fintechs who might be able to bring this capability in. Dan, do you think that that's their saving grace is to work with fintechs?

Dan Van Dyke:

Yeah, I think that's one of them. I think underlying too, the unit economics for a small bank thanks to Durban are a little bit advantageous. And if you imagine in a future where technology as emote becomes less important with big banks or banks in general in the US spend about $88 billion a year on internal and client facing technology. And what if generative AI was able to level the playing field in terms of being able to stand up incredible client facing experiences with less resources? I think that then what big banks and small banks compete on is the financials.

And we're already seeing pressure on the margins with companies like Apple in partnership with Goldman Sachs offering incredible savings rates relative to big banks. I think smaller institutions will be able to capture more interchange can be able to rely on that underlying unit economics benefit to stand up to bigger banks and to be able to partner with fintechs to offer better client facing interactions and solutions. And I'm excited to see that play out. It won't be all small banks capturing the lead over the big banks, but I think you will see some examples of smaller banks punching outside of their weight class by using generative AI to accelerate their technology offerings.

Rob Rubin:

I wonder if we're going to see a tech company emerge that buys a small bank and is really good at it. In other words, creates an experience which is one that people remember, one that people strive to have. And that's the challenge that I have with smaller institutions is they're not generally good at building these digital experiences. And while I see the advantages from a programming capability is they can leverage a tool to develop for them as opposed to humans having to do it. But at the same time, I don't see how these smaller institutions are going to be able to pivot their cultures to being more digital. Is there opportunity for a small institution to break through as opposed to, can smaller institutions survive? Is there an opportunity here for smaller institutions to break through and actually have an impact?

Eleni Digalaki:

I think yes, because as we said, the tech does level the playing field, so the potential is there. I think it's going to be bad news for the majority of community banks, but yes, there is opportunity for those that act smart. I agree partnering with fintechs will probably be a very smart way to go about it, for example, to break through. And I also think it's important to maintain the advantages that they have as they go about doing it. Because yes, maybe they're not known for offering the best digital experiences, but they are known for understanding what their communities need and being ingrained into these communities. So how do you bring those two things together?

Rob Rubin:

So I think if I was to sum this up, I still think everybody still agrees with me that in 10 years, 95% of the banks that exist today might still be gone. And I want to point out there are 10,000 banks and credit unions in the US, so when I say 95%, I still think there's 500 banks in the US in 10 years, that's a lot. I mean, if you look at how many banks are in the UK or saving societies, not very many, right?

Eleni Digalaki:

Yeah, it doesn't compare. It's a much more concentrated market in the UK.

Rob Rubin:

Right. And it's been concentrating for years already, so that law of things compounding 10X each year I think it might apply to the rate which banks go away. And I think it has a lot to do with technology, not necessarily AI, but the writing's been on the wall for a while. I want to say though, again, I think that there's this opportunity for a breakthrough and some of these smaller institutions will break through and really I think put pressure on some of the bigger institutions to start to deliver some better experiences as well. So we'll see what's going to go on. I want to first say thank you to Dan and Eleni for joining us today. Thanks guys.

Eleni Digalaki:

Thank you, Rob. Thanks for having us.

Dan Van Dyke:

Thank you. Always a pleasure.

Rob Rubin:

Yeah, it was a lot of fun. And I hope, Dan, that we can have you back more and more as time goes on.

Dan Van Dyke:

Would love to.

Rob Rubin:

I want to thank everyone for listening to the Banking and Payment Show, an eMarketer podcast, sponsored by TransUnion. I want to thank our editor, Todd. In today's episode, we referenced our new report on generative AI in financial services, an article that we published May 3rd on generative AI and how regulators are jumping in. And an article from the Wall Street Journal on how Goldman is already in front on leveraging generative AI. We'll put links to all these things in the show notes. Our next episode is on May 30th and you'll not want to miss it. See you then. Bye everyone.