- Technology
AI helps banks of all sizes think bigger
- AI has potential to be a competition leveler in financial services.
Share
Lee Wetherington, Senior Director of Corporate Strategy at Jack Henry, joins Senior Editor Rachel Koning Beals on the BAI Banking Strategies Podcast to share his belief there’s a data and AI moment of reckoning in our industry, forced in part by the speed of gen AI curiosity. With this tech breakthrough, however, comes real opportunity for nimble financial services organizations.
Takeaways:
Don’t miss the other insights, methods and tools that can help your financial institution shape and evolve its AI strategy in the BAI Deep Dive: A practical approach to AI.
Transcript:
There are market equalizers inherent in AI. That means small, mid-sized and larger banks are all sitting at the table… Lee, welcome to the BAI Banking Strategies podcast.
Wow. So where to begin? A lot has changed in the last six months, 12 months even. And in fact, I’m hoping some of the things that we can cover will correct some of the assumptions some of us may carry from 12 months, 18 months ago that no longer stand. The bottom line for our industry relative to generative AI is that the popularity of generative AI, the promise of orders of magnitude, new efficiencies that can be realized and gained has suddenly put it front and center in a lot of different domains. But what it’s done in effect, it has backed our industry into finally getting serious and sober about data strategy because you can’t do anything meaningful with any kind of AI model, generative AI or otherwise, without having your data house in order. And that not only means what data you have to bring to bear, it means how quickly you can bring that data to bear.
And all of that gets to the question of whether or not you have a modern data regime and infrastructure that can lever your data in ways that are material both for back-office efficiencies as well as new product innovation, vis-a-vis your customers.
No surprise, AI is really a data story. And no one segment of our industry, big banks, small banks, has the data access they think they do, right? Explain for us…
I think that’s exactly right. So let’s break this into two parts. One is there was an understanding, if we go back 12 months, 18 months, 24 months, there was a thumbnail understanding that to play an AI, you had to have really, really large amounts of data with which to train these large language models that we read so much about and hear so much about. The problem with that is that if you’re not technical, you don’t understand that even the largest banks in the United States in and of themselves do not have enough data to train what we call a large language model.
In fact, the amount of data you need to train a large language model like a ChatGPT-4 or a Google Gemini, defies human comprehension and imagination. And in fact, those models have consumed so much publicly available non-proprietary data that they will run out of that kind of data by 2026, two years from now. Those large language models will have exhausted all non-proprietary publicly available data on the internet. So the question is quickly you get to, well, what does that mean? What does that mean for any bank, big or small? That’s one question.
The second part of that is a lot of banks assume that they have all financial data for their customers, and that’s the data that they will bring to bear to train and tune generative AI models for a host of different applications. That is a fundamentally inaccurate assumption. The average consumer in the United States has now relationships with between 15 and 20 different financial service providers. They have 14 financial apps on average on their smartphones. So their financial data is scattered across a whole host of providers and apps. And at best, even at the largest banks in the country, you might have maybe 20% maybe, maybe if you’re really lucky, 30% of the financial data on your customers, but the rest of it is scattered across this financially fragmented landscape.
And so when you start getting serious about AI and then you suddenly… That backs you into these considerations about data and you start doing an inventory, a lot of banks are realizing, “My gosh, I don’t even have the preponderance of financial data on my own customers that I thought I did. So where do I begin? What can I do? What is meaningful? Not to mention what is allowed.” We can have the regulatory and compliance conversation, but we’ve been backed into this larger conversation about data, modern data, modern data exchange, and of course, hovering over all of that is the CFPB’s new rule on privacy of financial data, the ways in which it’s exchanged. What we in shorthand refer to as open banking in the United States. That rule will be finalized this fall and in effect in 2025.
Two things to put a pin in, right? Regulations and open banking, we’re going to get there. I love it. Two things to talk about too. All this fragmentation, right? I think you and I talked recently about… You said, “Forget primacy. We have to talk about first app.”
This is a big blind spot in a lot of boardrooms and I’m in quite a few bank boardrooms year over year. And in a healthy percentage of those boardrooms, this is not even a known phenomena that financial fragmentation is a thing, that you think you have a primary checking account relationship, and in many instances, what you have is a zombie account. So money may come in from a direct deposit and then it immediately goes out and gets scattered across these different apps and/or third party providers or other financial service providers that that customer is using. So that’s one, recognize that that cow is out of the barn. That’s the first thing.
The second thing, I do laugh when I get this question, is if you fill that blind spot, then the question that I get from certain board members is, “Well, Lee, how do we get them to shut down all of those other 15 to 19 relationships and bring it all home to the bank?” And that’s when I have to be as polite and diplomatic as I can and let them know that that’s delusional thinking. That cow is out of the barn. You’re not going to get anybody to give up any app that is offering them value in particular situations, circumstances in their money life.
The primary goal now is can you get to first app status? And what I mean by that, what we mean by that when we say that is can you be the app that makes sense of all of those other apps and/or service providers and the data that’s scattered in and among them by aggregating that in a secure, modern API-based, standardized way, back to the bank so that you solve for the fragmentation, you give the customer a comprehensive understanding of their money at a glance? You by the way, also completely eliminate inbound screen scraping, which is one of the things that the new CFPB rule is prescribing a timetable to end by the way.
And you want to do that. That’s the right thing to do, industry-wide and for your customers. You want to end inbound screen scraping because it brings with it a whole host of fraud and cyber security vulnerabilities that we all want to move beyond. But once you bring all of that data home, you now have that preponderance of data on your own customers so that you can not only understand definitively who they’re doing business with, for what, at what price points? So that can be a feedback loop into your decisions as a bank about what product or service we want to either build, buy or integrate next. But it also gives you everything you need to take advantage of generative AI and all of this is tied together, but these connections aren’t being made in the average boardroom and oftentimes in the leadership suite at banks across the country.
One quick point on that, you can tick the regulatory box there in a way, at least the CFPB box, but also, we’re talking about authenticity, we’re talking about personalization. Customers care, especially those younger customers, they’re being fed this through their streaming. This really does matter. So it’s not just thinking first app to be digital for the sake of digital. This really should pay off ultimately, right?
Yeah, pay off, let’s take that literally. So the nut to crack in banking is how can we suggest a next best product or service and actually be accurate in that recommendation? But the other part of that on the receiving end from the customer side is how do we expect customers to act on the next best product or service recommendation when we’re making that recommendation with only a view to 10% to 15% of their financial data that happens to be resident and native inside of the bank? So it’s a matter of financial confidence. We don’t have financial confidence. If you look at how financial health generally in the United States is trending, it’s not a pretty picture. 70% of adults in the United States are not financially healthy according to the Financial Health Network that does these yearly pulse checks on that data. And one of the fundamental, and again, largely hidden things is this financial fragmentation.
By the way, let’s talk about it for a second. How did we get to 14 financial apps on the average phone and 15 to 20 service providers? We got there out of a cumulative aggregate of spot conveniences over time. So for instance, Rachel, you and I may have had lunch somewhere and you bought lunch, and I’m like, “I got to pay you back Rachel. So hey Rachel, do you use Cash App?” And you say, “No, I use Venmo.” And I’m like, “Ah, let me download Venmo, right?” And so I download Venmo and I pay you. And then next person, it’s vice versa. It’s Cash App. So I got to make sure I’ve got Cash App. Do that with someone else and it’s Zelle, or maybe my kids prefer Cash App or whatever, whatever. So all of these spot conveniences in aggregate, cumulatively over time get us to this completely inconvenient fragmentation that doesn’t make it easy to understand where we are with our money.
Now, how does that track back to confidence? If you don’t know where you are with your money, you can’t really trust anybody’s recommendation for a next best product or service, much less act on it. So what we’re seeing is for those banks that are levering open banking rails to address and solve fragmentation and bring this data back home to the bank and to the customer, once that customer sees a reliable and comprehensive view of their money, they’re two to three times more likely to act on a next best product or service recommendation from the bank. Now you’re cooking with gas, now you’ve got something that’s driving revenue, that’s deepening relationship. And that’s really the end game from a financial standpoint, from a viability standpoint for banks to connect these dots, bring the data home and provide better, more accurate, more relevant and more personalized service recommendations and product recommendations to their customers.
So we’re helping banks bring those customers home, we’re helping them with data and let’s help them get that data in a smart way. It’s not always the large language models that dominate the headlines. They’re small language models. There’s a difference and that difference can be a big deal for smaller and mid-sized banks. So Lee, walk us through that.
This is probably the best unknown news for your listeners because roughly 12, 14 months ago, all of the research, most of the analysts out there predicting what generative AI’s impact on our industry would be, were saying the following thing. This is going to exacerbate the competitive gap between the biggest banks in the country, who have of course, the most data and the most data to train and/or tune generative AI models versus every other bank in the country that doesn’t have nearly the amount of data that a top five bank or a tier one bank would have and be able to bring to bear. And so a lot of people are still… Because they’re not tracking the evolution of not just AI but AI development, AI tooling, what’s called Software 2.0, they don’t understand what is trending and evolving and changing about the performance of different language models.
We only hear about large language models, but there’s an entirely different set of the language models that are very, very small relative to the larger language models that we read a lot about. So for instance, Google has partnered with a small language model called Mistral, which is an AI company out of France, is causing big stirs in our industry, but across other industries as well because it is able to create models that are 60, 70, 80 times smaller than a large language model, yet is able to perform in ways that are on par with the outputs of large language models. Why does it matter that suddenly we’re seeing for instance, right now, of the top six performing language models in the world, two of them are small, these small language models? Why does that matter for our industry?
Well, it matters because it has turned upside down this presumption that AI will extend big banks’ large mover advantage, their competitive differentiation. When actually, the way this is playing out is just the opposite because these smaller language models are much more cheap to run and they’re also easier to contain and secure in terms of whatever data you may want to control going into them or whatever outputs are coming out of them. That is front and center in a hyper-regulated industry like ours.
And in fact, the prediction by CB Insights and others now is that for hyper-regulated industries like banking, like financial services, we’re going to see a heavy bias towards small language models that can perform really well at fairly predictable and repetitive tasks. But also, these small language models can run locally on device. They can run very easily sequestered and contained in a given cloud instance that the bank has full control over. In other words, the bank doesn’t have to worry as much about PPI data and PCI data and other IP type data finding its way into these large language models that the bank does not own or control in terms of what goes in and what goes out. That’s a nightmare scenario and that’s why a lot of banks are still waiting. We’re all waiting for very clear guidance on what’s okay and what’s not okay in terms of the types of data and the types of models that can be used to do different things to bring efficiencies both to the back office, but also new and wonderful personalized services in the self-service mode through digital channels and mobile apps.
You mentioned back office there. Let’s talk a little bit about that because I think probably for cost considerations, probably for regulatory considerations, maybe the best AI bang for your buck right now may be spent in the back office, unit economics matter. Can you talk a little bit about that, budget considerations, priorities?
Absolutely. So we just completed at Jack Henry what we call our annual strategy benchmark. In this year’s survey, we surveyed 127 CEOs. And by the way, this is always really interesting. We survey both the CEOs of banks and we survey the CEOs of credit unions, the same set of questions so we can compare and contrast priorities and things that are going on. I would tell you, one thing that your audience here may be interested in is that credit unions historically, including this year based on this survey data we just got back, have always placed an outsized strategic priority on data, analytics, machine learning, and now AI. And the reason why credit unions have always historically done that is twofold. One, credit unions on average, even the average credit union is still smaller than the average bank in the United States. So credit unions have always seen data and analytics and automation as a leveler between themselves and just banks generally of all sizes.
But more importantly, credit unions have always had a retail focus, which is to say they churn out a lot more loan volume at smaller dollar amounts because of their retail focus. And the only way to do that larger volume of smaller loans to consumer members in the credit unions’ case is they’ve had to rely a lot on automation in the past. And so they’re very comfortable and fluent with analytics and now these new waves and new orders of efficiency that can be realized.
One thing that is in common between the bank CEOs that answered our survey and the credit union CEOs is we saw this jump in the strategic prioritization of efficiency rolling into 2024, and that is tied to their number one priority, which is still deposits, gathering new deposits, protecting deposits that you have, trying to stave off more deposit attrition. But the way credit unions see doing that, and the way banks realize that in order to survive, they’ve got to prioritize in the short to midterm is efficiency. They’ve seen all this pressure on cost of funds, on net interest margin, on net interest income. So now they’re like, “How do we get more bang for our buck based on existing technology or new technologies we can put into place?” And generative AI is the answer to most of those needs and priorities around realizing efficiencies in the back office. So it is the number two most strategic priority for both banks and credit unions by the way in 2024 and 2025. And that’s driving a lot of this interest.
I would mention this too, if you get outside of our industry and you just ask CEOs of all types of companies in the United States how important generative AI is this year, 40% of them say, “We’ve got to make a move using generative AI this year because we can’t afford to be left behind competitively if anybody else in our industry is using these same tools to achieve unit economics that we can no longer compete with.” So this is the tension. In our industry, we’re always sort of waiting for regulators to give us safe harbor about what we can and can’t do, but there will be progressive banks and there are progressive banks out there that are willing to push the envelope to be first to those new efficiencies to give them the competitive advantage that will be difficult to close the gap on if you’re just sitting on your hands waiting for a green light from the CFPB or anybody else for that matter.
Regulation is always tricky. Do you self-regulate? Do you try to be ahead and anticipate it’s a global market in a lot of ways? And we flicked a little bit here at regulations. You mentioned small language models as a self-contained way to answer regulatory questions. How do you advise clients wholesale? Should our industry try to be ahead of the game when it comes to regulation in this fast-moving world right now?
I think there’s a very clear short-term, near-term path. And that path is one, focusing generative AI capabilities in the back office that don’t directly touch or impact the customer. And let me give you the best example. In fact, if you look at the top three use cases for generative AI in banking, the number one use case is a generative AI assist capability that is assisting humans at the bank, people who are customer service representatives and others inside the bank that are fielding and resolving moments of need that come through the digital channel or the mobile app. So for instance, if you think about a common example, you open up your mobile app, your balance is nowhere near what you expected it to be, and you look down at the top of the transaction list and you see two transactions that you don’t recognize and now you’re having a mini panic attack thinking, “Oh my gosh, I’ve been had, this is fraud.”
And so in the best mobile banking apps today, you have what I would call an augmented chat capability, which is to say at a touch of a button, I can be in connection with a real live and not only a real live human being, but the right human being at the bank to solve this moment of need I have and this moment of fear around this fraud. But not only that. I can tap, this is augmented chat, I can tap the two transactions that I don’t recognize and attach them to this live authenticated secured chat thread with the right person in the bank. Now, what’s going on in the background is that this augmented chat capability is triaging the type of language that’s used in that initial reach out from that mobile app. And if the word fraud is in there, well, it’s going to connect you automatically with someone in the Fraud Department. It’s not going to bounce you around in the bank. The person in the Fraud Department immediately can see the two questionable transactions and can immediately reply back.
Now, before that, what’s happening is now a generative AI assistant on the bank side is looking at this all in real-time. And when this query comes in about these two fraudulent transactions, there’s literally in real-time, the generative AI assistant is saying, “Here is the answer that historically has proven to be the best, most accurate and definitive answer to this question.” But then whether that reply goes out is based on the decision of the person at the bank, that service representative or that person in the Fraud Department. They can go, “Oh, yes, that’s right.” Tap, yes, it goes out as is. Or”, “You know what? No, I want to change this one little thing.” They’re going to cure that answer. They’re going to curate that answer, is a better way to put it.
Moreover, here’s the beautiful part, the creativity in this, they’ll be able to right-click that generative AI suggested response and they can say, “Make this more friendly or make this more formal,” or we know that this is a native Spanish speaker. “So from here on out, all of my responses, make them Spanish, even though I’m seeing them in English on the bank side, and keep it that way for the life of this exchange.” You can do crazy things. You can say, “Oh no, you know what? This is actually my personal friend that loves hip hop, so let’s have fun here. Put this in the voice of Snoop Dogg and send that back.”
So my point here is that generative AI assist capability is going to turbocharge the personal service in real time that an average bank can render cheaper, faster, better, and at higher quality than has ever been possible before. And that, in my mind is going to inaugurate a new golden era for relationship banking models in the United States. So a lot of people think, “Well, generative AI, ultimately, once we get the green light from the regulator, generative AI is just going to do all the work and answer everything, and it’s all self-serve from there on out.” First of all, you don’t want to do that with a pure play generative AI model because a lot of people don’t understand that generative AI is a statistical form of AI. In other words, you can put the same input into it over and over again and get different outputs each time.
That’s not kosher in a customer-facing application right now. You have to put so many guardrails around it before it would ever meet a regulator’s muster or an examiner’s muster. So for the short term, back to your question, what am I recommending? First application, first use case is get generative AI assist to turbocharge your ability to service, to solve and to make product recommendations even because General AI can do all of that, but it has to be curated by a human being at the bank because it’s the human beings who are going to have liability vis-a-vis the regulator or the examiner. You can’t just unleash it yet raw to the customer.
There’s also, in terms of financial data analysis, and by the way, there’s even been a lot of investment in what’s called synthetic data generation. So for a lot of these smaller banks who were thinking, “Well, we don’t have enough data to train or tune a generative AI model, where are we going to get that data?” Now, we’ve already talked about one of those things, which is open banking, get the rest of your customer’s data back to the bank. Now, with that preponderance of data, you can train generative AI models.
But another thing that generative AI can generate is data sets that look exactly like very small real data sets that you tell a generative AI model to look like or to reproduce. So think about that for a minute. I’m a small little community bank. I only have a tiny little set of data. It’s not enough to really train a model. I can feed that to a generative AI and say, “Make a data set that is 1,000, 10,000, 1,000,000 times bigger than this data set, but in its properties as a data set is exactly identical to this real live little data set I gave the model in the first place.” And now you’ve got a set of data, it’s dummy data. It doesn’t have PPI in it. It doesn’t have PCI in it. You don’t have to worry about any of those things, but now you’ve got a large quantity of data that an aspect and profile mirrors your real data, and you can use that to train or tune models for a variety of different purposes.
Now, I think that based on what we’re seeing with the improved performance of small language models, that will make having to synthesize a bunch of data moot for most small banks. That’s my take at the moment in terms of where we are with the evolution of all of these tools.
You almost went there. I’m going to push you just a little bit. So based on everything you just said, and in addition to it, the fact that cloud storage, cloud computing is also getting ever more competitive, thus cheaper for smaller banks too, that’s going to help too, that third party vendor cost component, throw that in too, and all the things you just said. As we wrap up, is it fair to call AI the great leveler right now in banking among all the market sizes of banks, or does that oversimplify it? Are we at that point, is it the great leveler right now? I’m putting you on the spot.
You are putting me on the spot, and I’m happy to stand on the spot. I’ll put down this. I do think it’s going to be the great leveler, the way the economics is evolving. So you mentioned that, and I’m glad you mentioned that, Rachel, that the big cloud providers, which also happen to be the big AI providers. Why? Because you have data in those public clouds. So you’re talking about Google with GCP and now with Gemini, which is the first commercial-grade multimodal generative AI, which means a single AI can create text, video, audio. You go on down the line, you’ve got Google, you’ve got Microsoft with its OpenAI investment and partnership. And by the way, it’s really interesting development. Just in the last week or so, Microsoft has hired basically everybody at Inflection AI to come over and begin working at Microsoft. If it was a way of acquiring a company without acquiring the company, that is a fascinating thing, and I won’t go any further on that. And then you’ve got Amazon.
So Google, Microsoft and Amazon are at war with each other in terms of being the public cloud provider of choice, not just among financial institutions, among everybody. What are they using to fight those wars between each other? They’re fighting using the tooling that you get if you come to their particular cloud platform. So this includes AI tools. This includes a brand new way of even doing software development that is large-language model-centric. It’s called Software 2.0, and at the moment, Google’s got everybody beat at the moment because Google created the transformer technologies that got us to the large-language models that we’re talking about. They’ve been applying these AI tools and development methodologies to their own product sets for years they’ve been doing this. Google had a ChatGPT internally back in 2018 and just was uncertain about what it would mean if they released it to the public. When OpenAI opened that Pandora’s box, now all the gloves are off and they’re at war with each other.
The bottom line for the banks who are listening is that that means the cost of compute is plunging. The cost of data storage is plunging. Access to AI tools and AI models is skyrocketing. AI as a service is making AI in its most sophisticated forms accessible and affordable for the smallest of financial institutions, and this is why, based on all of those trends, I’m confident that this is going to end up being a leveler and is not necessarily going to be the exacerbator of the large mover advantage of the top five banks in the country. I’m excited. I’ll tell you this, Rachel, I’m excited to be alive to see how this unfolds. This is a really interesting moment in history to occupy and be around for.
Well, I’m so glad I put you on the spot because what an answer, Lee. So glad we got you in this moment in time. Really fascinating. Lots for banks to think about. Jack Henry’s Lee Wetherington, thanks for joining us today. Great to have you.
Thank you, Rachel.
Become a member to unlock exclusive content, connect with industry experts, and gain access to valuable resources. If your employer is an institutional member, activate your ProSight membership benefits with a simple email address.