Practical AI-informed decisioning in the bank

Key takeaways:
Reasonable expectations: AI-based decisioning in financial services leverages your existing tools and complements human judgment and intuition.

A task-based approach: Pinpoint the business challenge at hand and minimize risk with targeted AI-informed decisioning, such as pulling up AI-powered next-best product recommendations.

We’ve been here before: SAS’s Stephen Greer compares the mixed emotions around AI to the early days of cloud adoption. Benefits of cloud data storage were clear, but there was apprehension about security and other factors from major institutions.

TRANSCRIPT

Rachel Koning Beals: Whether your institution is an adopter or just an explorer, AI-informed decisioning in banking has arrived.

SAS’s Stephen Greer joins me, BAI Senior Editor Rachel Koning Beals, on the BAI Banking Strategies podcast to tell us how AI-informed decisioning is changing the bank as we know it. Let’s go deeper.

Stephen Greer is an industry consultant and financial services strategist with data analytics and AI solutions provider SAS. Over his career, he has developed a passion for helping banks and credit unions succeed at the intersection of technology and business strategy. He believes data and AI can be smartly and safely harnessed to solve industry challenges, maximize customer relationships and grow revenue. Welcome, Stephen.

Stephen Greer: Thank you for having me on.

Rachel Koning Beals: You bet. Stephen, you and I find ourselves connecting at a time that data informed banking, machine learning, AI, and especially generative AI, bring a lot of change, or at least the potential for change to financial services, and very quickly, with major market shifts even in just the past year or so. From the SAS vantage point, what’s changed the most and where do we as an industry sit today with these developments?

Stephen Greer: Yeah, I think that’s a really good question, and I’d almost say in a weird way, what’s changed the most, is we’re talking about just the last year to year and a half, is really the expectation for the future. And what I mean by that is that you had AI go mainstream in 2022 with ChatGPT. It’s almost been treated like a silver bullet that we’ve seen compared to other types of AI and predictive models. It’s the most user friendly. You can apply it to all these different problem areas. So what we’ve seen is this rush to launch a variety of different pilots, but in that process, there’s been this kind of slow realization that there’s a lot of potential here and some very clear productivity benefits. But you still have a lot of inherent limitations. And so we’re almost starting to see a reassessment of where these models are best suited, and kind of where some more traditional forms of AI might be a better option. So we’re seeing a resurgence of interest in AI, and particularly how you integrate that AI into operations for decisioning, and also a recognition that how we talk about it might change a little bit. We might try to step back and think about AI-based decisioning in a broader context, which takes all the tools into account that you might use to solve for various industry challenges.

Rachel Koning Beals: You started to talk a little bit about what we mean by AI-informed decisioning. But can we sum it up in a definition?

Stephen Greer: Yeah, definitely. So everyone wants to be an AI-driven organization, which means informing operational decision making. So if you think about it, every decision in financial services is the output of some combination of predictive models, decisioning, rules, human judgment and reasoning. So just like an algorithm on a computer, these decisions work by starting with an input and then ending with an output, and then analysis that takes place in between. And so sometimes you have that in between that is very simple and explainable. Sometimes it’s extremely complex, right? It could be mostly or all human. It could be entirely machine. And so when we think about this, institutions have so much input data, which holds the insights to enable better decisions, and the processing that happens in between then is how those insights are extracted and unlocked. And so when we think about AI-based decisioning, we’re thinking about using AI to gain considerable leverage, which complements human judgment and intuition.

So at SAS, our definition in financial services, AI then becomes essentially a system to support and accelerate human decisions and actions. You often hear us talk about this in terms of human in, on or out of the loop, depending on the decision or depending on the complexity of it.

Rachel Koning Beals: I hear the word automation, or at least the promise, sort of the hype of automation around this technology. True?

Stephen Greer: What we’ve seen be successful in financial services is just acknowledging that the complexity of putting AI into a decision flow. So oftentimes, we aren’t really advocating for full autonomy. We’re advocating for leverage, and usually with some of the more advanced models, like generative AI, right? That’s usually a human in or on the loop. And so this leverage, to kind of allow the employee to do more high-value work and offload some of the more monotonous or rote work that takes so much time and can add friction and be an impediment to getting a model or getting advanced analytics into a production state. And so when we talk about automation, that’s kind of how we view it, where we’ve seen institutions be successful building on that.

Rachel Koning Beals: Are there some more obvious departments in the bank or credit union where AI decisioning makes the most sense?

Stephen Greer: Yeah, it’s kind of everywhere. Maybe it’s worth putting in context what we mean when we talk about AI. One thing we’ve seen, that we tend to gravitate towards and where the hype is right now, that’s generative AI. It’s kind of what everyone wants to talk about. It’s what you’re seeing in the news, and it’s kind of when you think of AI, you think of what’s the latest technology that’s been released? But AI is a broad umbrella, right? And there are a lot of different disciplines, and there are a lot of different techniques that are underneath that, and each of those techniques are optimized to solve for specific problem sets.

So maybe we could put it into maybe the context of an onboarding journey. So let’s say you start with marketing automation. You’re calculating turn propensity or next-best action for a specific segment, and then you’re serving up an offer with AI as the underpinning for that decision. A customer might apply for that offer, and then a credit risk assessment is made using AI, right? Decisions are made on the pricing and the loan amount behind the scenes, or what the offer is that you’re going to be giving that specific customer. After that, you have things like customer due diligence and other fraud checks that require decisioning, the sanction screening, KYC requirements. Is this application fraud? Do we need to trigger additional controls here? Then you’ve got after-funding and provisioning. You’ve got the ongoing servicing and monitoring, so protecting early warnings for maybe a stressed borrower, and deciding on different treatments for collections if they go into default and delinquency, and onto all the authorizing upsells, credit line increases, ongoing transaction monitoring. So just within that journey itself, you can see all the different areas in which AI could play a part and where you could amplify human-based decisioning through advanced AI models.

Rachel Koning Beals: In a way with that onboarding example, you’ve sort of answered this, but it does make me wonder, can this migration, this adoption of AI often be task based, or sort of product specific? Or should banks think more whole-of-enterprise when they’re approaching AI?

Stephen Greer: Yeah. I mean, it kind of depends on the organization. We’re seeing a lot of task-based approaches that might look at a discrete process and then explore ways to optimize it, and so you can get the benefit, and this lowers risks. It might be quicker to achieve. It might get closer to the business problem, which, in the case of a lot of generative models, for instance, like LLMs (large-language models), is increasingly where they’re doing the best. Then you have kind of the constraint of getting closer to the business problem itself, rather than some kind of foundational model that can boil the ocean. And so in that instance, task-based approaches more specific to a process make the most sense.

But then we also have a lot of institutions undergoing larger transformations, where we see a couple of our customers have consolidated decision engines, right, or consolidated systems into a single decision engine. And so this might create a more cohesive approach, where you get some scalability benefits. You might have more governance and control over the outcome of how AI is being used across a variety of different lines of business or processes. And so it kind of varies depending on who you’re talking about and what kind of problem they’re trying to solve.

Rachel Koning Beals: You know, we’re sort of in early days here, especially with Gen AI. Have you guys as consultants already been able to help clients deal with some pain points with pivots? Because, for instance, what they’ve market tested has changed minds. I’m interested in knowing, are there sort of sophomore year market tests going on out there? What stage are we at?

Stephen Greer:  In some ways, depending on how familiar they are with the space, there’s pretty well-established challenges that we have today, and you deal with hallucinations – that’s a big one. Output performance and the size of the context window, or the input data kind of requires some pre-processing and other things that go into it. There are concerns around interpretability and explainability of the results, whether or not bias is being mitigated. Regulators are especially concerned here. There are all these concerns around cost and speed of the different models … which might limit their scalability or use in kind of real-time context. I would say, one of the biggest pain points with these models, you kind of have this creativity that we aren’t used to, and a new type of problem set that they’re trying to solve for. So if you think about like traditional machine learning or traditional AI, usually it’s to solve a specific task where the output is fairly well defined. What your goal is, what you want to predict, is fairly well defined. But with generative models, you have a different approach, where they’re learning the distribution of the data it’s trained on and how it’s structured, so it can reproduce something that looks similarly structured to the data that it was trained on, but it doesn’t follow directly from it the way it would in traditional machine learning context. So you get this kind of low level judgment about the output, and it kind of appears like creativity, which is interesting for us. We aren’t really used to that kind of creativity inherent to these models, and so I think that’s impacting the expectations and how they’re being attempted to put into production.

But really, where we’re at now is very early days, a lot of really cool demos. There’s incredible potential in the future, but we aren’t seeing anything substantial yet. We aren’t seeing that kind of killer use case.

Rachel Koning Beals: I mean, that’s kind of a good segue, in a way, to your advice on which stakeholders and when to potentially get involved when pitching AI decisioning. It strikes me we are talking about potentially everything from shareholders to board members to your call center leads to your marketing heads. It’s a long list, potentially, right?

Stephen Greer: So at the end of the day, if you’re in financial services, especially a bank, you’re in the risk and trust business, and that’s a lens that you apply across everything that you do in the industry. And for good reason, right? Because you’re handling customers’ money. When we think about this space, and we think about AI-based decisioning, our posture has always been, and we think this should be the posture of the industry as a whole, has always been a reflection of the industries we operate in. And so we work with very highly regulated, highly risk focused [institutions], and they approach very complex problem solving, and the consequences are really high, right? There’s low tolerance for failure here. And so we need to trust that when we put the AI models into production, [they] are fair and ethical and that controls and governance structures are in place so that we can mitigate any unwanted or adverse impacts of model drift or bias data that might creep into the training and all of those things.

So we’ve seen in conversations like that for our clients, at the major concern a risk of putting a model in production and making sure that it’s not biased, so that it’s not drifting and causing harm to their customers. That represents a significant risk for an organization. That is always in the back of our mind. It’s in the back of the minds of our clients. And so we’ve tried, to that extent, to embed those kinds of capabilities across the life cycle, to maintain that trustworthiness and that ethical AI standpoint.

Rachel Koning Beals:  Is this an industry shift that discriminates by bank size, only for the biggest operators, or is it more of a leveler, potentially?

Stephen Greer: I think it used to be the case that when you were smaller, you were at a significant disadvantage. I mean, it definitely still is, I think as far as just having the resources, the scalability, the kind of heft to throw money at tech and to attract the talent needed to, like, build internally, and all these things that a Tier 1 might do. But I think we’re in a significant period of change here where there’s a democratization of what was previously kind of opaque processes. You had real steep learning curves for data science, for instance, data science initiatives and extensive investment requirements if you wanted to get benefit out of it. We’re at this place where the accessibility of a lot of these tools keeps getting better and has really never been better. So traditionally, right, if you’re small, you’re running up against tech constraints, but we’re seeing increasingly that these tools allow an organization to kind of punch above their weight. And for us, this is incredibly exciting, because it means that all of a sudden, you’re democratizing Innovation, and you’re gaining parity industry wide, and opening up possibilities for all of these different institutions of various sizes to now engage in an area where it might have been prohibitive 10 years ago.

Rachel Koning Beals: Any comparable tech adoption of recent memory that might help us understand all these mixed feelings around AI, anything come to mind?

Stephen Greer: I kind of think of this as… similar to cloud computing, in the early days of the cloud. When cloud computing first came onto the scene, the benefits were clear. Or at least they were easy to describe. But there was some apprehension from a lot of major institutions. You had concerns around multi-tenancy, like if a security vulnerability could cascade into, you know, another tenant. There were questions around where data was residing, how encryption took place…how you could demonstrate regulatory compliance.

I remember it was a huge issue when if you were procuring a vendor who was deployed on the cloud, a lot of the processes were seen as somewhat of a black box.

Rachel Koning Beals: We talked a little bit about automation, but listen, there’s sort of these other, bigger, existential almost, questions around AI and the future of banking, and banking is getting increasingly unique all the time. Long gone are the days really, of just walking into one branch dealing with one banker. So people ask very good questions about the future the bank, should AI replace more humans? What’s your take?

Stephen Greer: It’s a good question, and it’s something that’s been top of mind for a lot of people, thinking about this and thinking about the potential. And there’s been a lot of, I think, fear here. My take is that I think if you look at where we are today in many different domains of work – and this view is shared with a lot of people across tech – there’s demand for services which far outpace the supply. You’ve got IT teams that are overwhelmed by different feature requests. You’ve got migration efforts that are monopolizing schedules of release cycles. You think about how much COBOL still exists in financial services today, a 60-year-old language, right? Fraud investigators and compliance officers are often inundated with different cases, and the traditional response has been to just throw more bodies at the problem, rather than to increase the efficiency there when there’s kind of a regulatory requirement. [Regulation and compliance] is another area that we’ve seen a lot of issues where you kind of did the fire drill for the amount of effort required to bring an institution up to compliance, scan documents, review policies, all of these things, right? So there’s so much noise and there’s not enough signal, and there’s so much demand where there’s not enough supply.

So when I think about this technology, you really think about the leverage to offload a lot of what we’d call menial work, to spend more time with higher-value work. And ultimately we feel like this will be the outcome… will be more work going live, more models put into production, more time on things that are actually going to make a difference rather than things that, from a day-to-day basis, kind of bog you down.

Rachel Koning Beals: I think that’s a great place to stop, because it gives us all a lot to think about. We can all be better at our jobs. What does banking look like today and tomorrow? Stephen, thanks for joining us. We’d love to have you back on the BAI Banking Strategies podcast.

Stephen Greer: I enjoyed it.