IBS Intelligence Global FinTech Interviews
Go one-on-one with the innovators, disruptors, leaders, and decision-makers driving change in FinTech and financial services. IBS Intelligence delivers exclusive global interviews that uncover strategies, challenges, and the ideas powering the next wave of financial technology.
IBS Intelligence Global FinTech Interviews
EP1025: From Hype to ROI: Making AI Pay Off
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
This interview explores the current state of artificial intelligence within the banking sector, highlighting the growing gap between high corporate expectations and practical technological limitations. The source emphasizes that financial institutions must prioritize robust data foundations and "context engineering" over simply adopting popular language models. A strategic framework is proposed for identifying viable use cases, focusing on areas with high repetition and manageable error tolerances. While AI offers significant potential for operational efficiency and employee fatigue reduction, the text warns of critical risks regarding algorithmic bias, a lack of explainability, and high operational costs. Ultimately, the expert advises bank leadership to move beyond superficial hype by focusing on disciplined deployment and long-term skill development for their workforce.
So to start off today's deep dive, there is this um this really striking image we came across while prepping for the show. It perfectly captures, I think, the whole mood of the financial world right now.
SPEAKER_01Oh, yeah, the corporate art piece. It's quite the visual.
SPEAKER_00Right, exactly. So if you're listening, picture this. It's a sleek, metallic, robotic hand, and it's extending outward and just hovering right above its palm, is this glowing, bright blue digital globe.
SPEAKER_01Aaron Powell And right in the middle of that globe, taking up like all the real estate, is this massive dollar sign.
SPEAKER_00Aaron Ross Powell A huge dollar sign. It is a very deliberate piece of art. It's basically sending this message of, you know, a literal world of wealth being effortlessly handed over to us by machines.
SPEAKER_01Aaron Powell I mean, it is the visual definition of a promise. It completely captures what has been sold to boardrooms all over the world for well the last few years, really. Trevor Burrus, Jr.
SPEAKER_00But and this is where it gets interesting. When you step out of that boardroom and you actually talk to the executives who are, you know, tasked with implementing this stuff, the reality is entirely different. Totally Yeah. The banking sector is pouring billions into artificial intelligence. But there is this quiet, almost pervasive sense of disillusionment that's taking hold.
SPEAKER_01Aaron Powell Because that robotic hand with the glowing dollar sign, it just hasn't materialized.
SPEAKER_00No, it hasn't. So our mission today is to figure out why. We want to cut through all the buzzwords, the hype, and look at the incredibly unsexy groundwork that's actually required to generate a real return on investment with AI.
SPEAKER_01Aaron Powell Especially in uh a really high-stakes, highly regulated environment like banking.
SPEAKER_00Exactly. We're gonna look at the hidden frameworks for success that actually separate the banks that are seeing real gains from the ones that are, well, just burning cash.
SPEAKER_01And our foundation for this conversation is a really revealing interview. It's from the April 2026 edition of the IBSI FinTech Journal.
SPEAKER_00Right. It's titled From Hype to ROI: Making AI Pay Off.
SPEAKER_01Yeah. And it features this incredible in-depth discussion between assistant editor Puja Sharma and DN Pralad. He's the founder and group chairperson of Surya FinTech.
SPEAKER_00And Prahlad has this really unique vantage point, doesn't he?
SPEAKER_01He really does. I mean, he's watching this massive technology adoption cycle unfold across dozens of institutions all at once. And his main observation is that while while technology cycles are totally normal, the sheer velocity of this one has been incredibly destructive to strategic planning.
SPEAKER_00Let's trace that back to the root cause because we all live through the post-chat GPT boom. I mean, it was everywhere.
SPEAKER_01It was inescapable.
SPEAKER_00It became the fastest growing consumer product in history. And that unprecedented consumer adoption triggered what was basically a blind panic at the enterprise level.
SPEAKER_01Absolute panic. You had these boards of directors just demanding, you know, a comprehensive AI strategy by Friday.
SPEAKER_00By Friday. And, you know, when you rush a strategy based entirely on the fear of missing out, you create this massive divergence.
SPEAKER_01Yeah, prelude describes it as this huge gap between the capability curve and the expectation curve.
SPEAKER_00Can you break that down a bit?
SPEAKER_01Sure. So the capability curve is the actual mathematical reality of what an AI model can technically do right now. But the expectation curve, that's what a panicked executive assumes the AI can do just because they saw some flashy demo on social media.
SPEAKER_00Right. I was thinking about this, and it feels like um it feels like someone buying a million-dollar track ready sports car and then deciding to take it off-roading in a swamp.
SPEAKER_01That is a great analogy.
SPEAKER_00Right. Because the car is an absolute marvel of engineering.
SPEAKER_01Yeah.
SPEAKER_00But if you force it into an environment it was never built for, you're gonna sink.
SPEAKER_01You're gonna get stuck in the mud.
SPEAKER_00Exactly. Yeah. And when you inevitably sink into the mud, you don't blame your own poor judgment. You yell that the car is defective.
SPEAKER_01Yep. You blame the tech. So let's actually look under the hood of that sports car to see where it truly belongs. Pralod is extremely specific about where today's large language models excel.
SPEAKER_00Aaron Ross Powell Okay, so where is the highway for this car?
SPEAKER_01Aaron Ross Powell If you need to summarize a 50-page document or search through a massive archive, paraphrase complex text, translate languages, even generate standard computer code. The technology is genuinely exceptional.
SPEAKER_00Aaron Powell So in those environments, the AI is driving on a freshly paved highway.
SPEAKER_01Aaron Powell Exactly. It works beautifully.
SPEAKER_00But then what constitutes the swamp? Where does the AI get stuck?
SPEAKER_01It sinks the moment you demand sustained multi-step reasoning.
SPEAKER_00Aaron Powell Okay, why is that?
SPEAKER_01Aaron Ross Powell Well, these models, at their very core, are just probabilistic text generators. They predict the next most likely word based on patterns. Aaron Powell Right.
SPEAKER_00They aren't actually thinking.
SPEAKER_01No, they don't possess a working memory that can hold on to a rigid logical thread through 50 complicated, interdependent steps without eventually drifting off course.
SPEAKER_00Aaron Powell So they lose the plot.
SPEAKER_01Exactly. Plus, prelid notes, they struggle immensely with massive unstructured data sets that haven't been meticulously organized. But the biggest swamp of all, especially for the financial sector, is any task where absolute 100% precision is required.
SPEAKER_00Which, I mean, if you're dealing with mortgages or corporate mergers, that's essentially the entire job.
SPEAKER_01Right. And the banks that understood this mechanical limitation early on are the ones winning right now.
SPEAKER_00Because they didn't put the sports car in the swamp.
SPEAKER_01Exactly. They looked at the AI's actual capabilities and surgically matched them to specific, tolerant workflows. The banks experiencing all that disillusionment are the ones that tried to force a probabilistic text generator into rigid deterministic processes.
SPEAKER_00So if you're listening right now and you're leading a team or running a business, you really need a way to figure out if your current problem is a highway or a swamp. You can't just guess.
SPEAKER_01No, you can't. And thankfully, Preload offers a highly pragmatic four-condition framework for exactly this. It's basically a filter.
SPEAKER_00Okay, let's go through it.
SPEAKER_01So if a potential AI project doesn't meet these four conditions, you just shouldn't build it. Condition one is an information gap.
SPEAKER_00An information gap.
SPEAKER_01Yeah. You have to identify a scenario where a human decision maker is lacking the right information at the critical moment simply because it's too hard to find.
SPEAKER_00The text gives a really good example of this. It mentions a bank's relationship manager. They're on the phone with a highly valuable client, and the client asks a super specific question about a niche financial product.
SPEAKER_01Right. And the answer exists, but it's buried.
SPEAKER_00Exactly. It's on page 142 of some technical PDF that the manager hasn't looked at in two years. That right there is a massive, immediate information gap.
SPEAKER_01Perfect use case. Now the second condition is repetition. You want to look for highly compensated humans who are spending their days doing mechanical robotic work.
SPEAKER_00Like what specifically?
SPEAKER_01Like searching through thousands of legal contracts just to verify if a specific regulatory clause is present. It's absolutely necessary work, but it's a terrible use of human intellect.
SPEAKER_00Aaron Powell So the AI takes over the repetition. Then we have the third condition, viability, which I interpret this as asking: is this task fundamentally impossible for a human workforce to accomplish on their own?
SPEAKER_01Aaron Powell Exactly. Let's take customer sentiment analysis across a major global bank. If you want to know the emotional state of your customers based on every single phone call, email, and chat interaction that happens in a 24-hour period, you literally cannot hire enough humans to listen, read, and categorize all of that data in real time.
SPEAKER_00Aaron Powell Right. The scale is just too huge.
SPEAKER_01Yeah, the task is only viable with AI.
SPEAKER_00Aaron Powell So the first three information gap, repetition, viability, they make total sense. But the fourth condition in Pralide's framework is where I suspect a lot of leaders get tripped up.
SPEAKER_01Oh, absolutely.
SPEAKER_00He calls it error tolerance. And look, if you are listening to this on your commute right now, you might be gripping the steering wheel thinking, I absolutely do not want my bank experimenting with error tolerance on my checking account. Trevor Burrus, Jr.
SPEAKER_01It sounds incredibly risky, right? Until you understand this concept of the blast radius.
SPEAKER_00Aaron Powell The Blast Rus. I love that term.
SPEAKER_01Yeah, prelot is basically forcing executives to answer a very uncomfortable question. What actually happens to our business when the AI inevitably makes a mistake?
SPEAKER_00Aaron Powell Because it will make a mistake.
SPEAKER_01It will. But if the cost of the error is low and the error is instantly detectable by a human reviewer, you have a great AI use case.
SPEAKER_00Aaron Powell Okay, let's say the AI summarizes an internal hour-long strategy meeting, but it misinterprets some minor bullet point. Trevor Burrus, Jr.
SPEAKER_01Right. A manager reads the summary, spots the error right away, deletes the sentence, and sends it out.
SPEAKER_00Aaron Powell The blast radius is practically zero. No one gets hurt.
SPEAKER_01Exactly. Now compare that to a scenario where an AI is autonomously approving or denying small business loans. Oh boy. Yeah. If it hallucinates a metric and misprices the risk on a thousand loans, or if it misses a subtle pattern in a fraud detection sweep, the blast radius is catastrophic.
SPEAKER_00Aaron Powell You just can't deploy a standard AI model into a zero-tolerance environment like that.
SPEAKER_01Aaron Powell No, you absolutely cannot. It requires a radically different, highly supervised, and a much more expensive architecture.
SPEAKER_00Aaron Powell Okay, so let's say we've used the framework. We found the Holy Grail, a task with an information gap. It's highly repetitive, it's viable, and the blast radius of a mistake is totally manageable.
SPEAKER_01Aaron Powell Sounds perfect.
SPEAKER_00The team is ready to build it. But the moment they plug the AI into their internal systems, the entire project just falls flat. Why does that happen?
SPEAKER_01Because they forgot to look in their own basement. Right. They slam headfirst into what Pra Lod calls the data value chain. And this is honestly the great irony of the modern banking industry. Trevor Burrus, Jr.
SPEAKER_00Because they have so much data. Trevor Burrus, Jr.
SPEAKER_01Mountains of it. Core banking records, decades of emails, regulatory filings, CRM logs. They are incredibly data rich.
SPEAKER_00Aaron Powell But they are tragically insight poor.
SPEAKER_01Exactly.
SPEAKER_00I was trying to visualize this, and I think of it like a restaurant owner who decides to hire this world-class Michelin star chef to turn their failing business around.
SPEAKER_01Aaron Powell Okay, and the chef is the AI in this scenario.
SPEAKER_00Aaron Ross Powell Right. The chef is the AI. So the chef arrives ready to create a masterpiece. But when they open the door to the kitchen pantry, they find absolute chaos.
SPEAKER_01Oh man.
SPEAKER_00The cans are rusted, the labels are all peeled off, half the ingredients expired in 2018, and everything is just dumped in a pile on the floor.
SPEAKER_01Aaron Ross Powell That is exactly what enterprise data looks like. And let's expand on why that chef fails in this scenario.
SPEAKER_00Yeah, please do.
SPEAKER_01Because unlike a human chef who actually has the common sense to, you know, smell a piece of rotting meat and throw it in the trash, an AI model is just a mathematical pattern matcher. Trevor Burrus, Jr.
SPEAKER_00It doesn't have common sense.
SPEAKER_01It has zero common sense. It assumes the data you feed it represents reality. So if your bank's data exists in these isolated silos, if the formatting is wildly inconsistent, and if there are contradictory records for the exact same customer, the AI just absorbs it all? It confidently absorbs all of that garbage, and then it will cook you a terrible meal, plate it beautifully, and present it to you as absolute fact.
SPEAKER_00And then the executive team is going to look at this terrible meal, blame the Michelin star chef, and say, well, I guess AI just doesn't work for us.
SPEAKER_01When the reality is they just refuse to do the unsexy work of harmonizing their data.
SPEAKER_00Right.
SPEAKER_01Proud Laud argues that executives spend way, way too much time in board meetings debating which shiny new AI model to purchase. Before that conversation even happens, the entire organization needs to stop and ask, is our underlying data ready for any model at all?
SPEAKER_00No amount of sophisticated AI can compensate for garbage data.
SPEAKER_01None at all.
SPEAKER_00Okay, so let's assume the bank actually takes that advice. They halt the flashy pilot projects, they go into the basement and they clean the pantry. The data is suddenly pristine, unified, harmonized. Right. How do they actually get the AI to cook the meal? Because according to the source material, just having clean data and a good model still isn't enough to beat the competition.
SPEAKER_01No, it's not. And this leads to a really fascinating realization about the current state of artificial intelligence. Which is the big, famous foundational models that dominate all the headlines, the ones developed by the major tech giants, they are largely commoditized now.
SPEAKER_00That is such a startling fact to process. The very technology causing this global frenzy is basically standard issue.
SPEAKER_01Yeah. A massive legacy bank with a trillion dollars in assets has access to the exact same underlying intelligence as like a five-person startup working out of a garage.
SPEAKER_00Wow. So the model itself cannot be your competitive moat.
SPEAKER_01Exactly. What separates a frustrating, mediocre AI deployment from a truly transformational one is this specialized practice known as context engineering.
SPEAKER_00Context engineering.
SPEAKER_01Yes. Pralod has this brilliant way of summarizing it. He says, the model is the engine, but context is the fuel. That is a very accurate way to visualize it, yeah. Context engineering is basically the art of tightly constraining the AI's probabilistic nature. You are giving it the specific parameters of reality it is allowed to operate within.
SPEAKER_00Okay, can we get a practical example?
SPEAKER_01Yeah, the text use is a great one. Tasking an AI to extract complex financial metrics from technical SEC filings.
SPEAKER_00Which is a highly tedious task that junior analysts usually spend hundreds of hours doing.
SPEAKER_01Right. Now, if you take a standard commoditized AI model, hand it a 300-page SEC filing, and simply prompt it to extract the revenue and growth metrics, it will almost certainly fail.
SPEAKER_00Aaron Powell Because it's just guessing.
SPEAKER_01Exactly. It might pull the wrong number from some obscure footnote, or it might just hallucinate a plausible sounding figure entirely.
SPEAKER_00So how does context engineering fix that?
SPEAKER_01Well, you don't just give it the document, you inject a strict verification layer.
SPEAKER_00Okay.
SPEAKER_01You provide the AI with a coded definition of what actually constitutes revenue in this specific context. You instruct it on exactly which sections of the document to prioritize and which to ignore.
SPEAKER_00Oh, I see.
SPEAKER_01And most importantly, you mandate that it must cite the exact sentence where it found the data. And you give it access to computational tools so it actually does the math itself rather than just guessing the next logical number.
SPEAKER_00Wow. So you are essentially surrounding the model with this incredibly detailed instruction manual that forces it to be precise.
SPEAKER_01Aaron Powell You are forcing the AI to show its work. And banks that master this specific process are going to squeeze significantly more value out of the exact same models that everyone else is using.
SPEAKER_00Aaron Powell That is a huge differentiator. So, okay, once a bank has done all this, they found the right use case, they have clean data and beautiful context engineering, they face the final hurdle.
SPEAKER_01The ROI dashboard.
SPEAKER_00Exactly. The executive team wants to see the return on investment. They want the dashboard showing exactly how much faster everyone is working now.
SPEAKER_01Aaron Powell And this is where we hate what the source material describes as the productivity paradox. And the data here is just completely counterintuitive.
SPEAKER_00It really is. Can you explain what happens?
SPEAKER_01It is genuinely perplexing when you first look at the research. So Pralod's site study is showing very clear wins in certain areas. Like AI deployment and customer support led to a measurable 14% boost in issue resolution.
SPEAKER_00That's pretty solid.
SPEAKER_01Yeah. The agents were simply closing more tickets per hour. And interestingly, the largest productivity gains were actually seen in the least experienced staff. The AI elevated the baseline performance of the whole team.
SPEAKER_00Okay, so that makes sense. But then the text points to a completely different study focusing on open source software development, which is highly complex technical knowledge work.
SPEAKER_01Right.
SPEAKER_00And in that environment, giving developers AI tools actually increased the amount of time it took them to complete their tasks. It slowed them down.
SPEAKER_01Yes. And here is the truly fascinating detail of that software study. Despite taking longer to finish their work, the developers subjectively believed the AI had made them faster and more productive.
SPEAKER_00We really need to wrestle with this paradox, because if I am a chief financial officer looking at that data, I am seeing red flags everywhere.
SPEAKER_01Oh, absolutely.
SPEAKER_00Why on earth are we paying millions of dollars for software licenses that actually increase the time it takes our engineers to build products? Are we just deploying these incredibly expensive placebos that make workers feel good while dragging down actual output?
SPEAKER_01I mean, the initial reaction is always to assume the tool is failing. But Prelod argues that purely measuring time to completion misses a massive, invisible variable in modern knowledge work. What's the very introduces this concept called fatigue reduction.
SPEAKER_00Fatigue reduction. Okay, let's break down how that applies to the developers who took longer but felt faster.
SPEAKER_01Think about the cognitive load of a software developer. A huge portion of their day is spent on draining low-level mental friction.
SPEAKER_00Like looking up syntax.
SPEAKER_01Exactly. Looking up syntax, formatting boilerplate code, searching for a specific library. It's the mental equivalent of walking through deep snow. It's exhausting. Right. AI acts like a snowplow. It clears away all that mechanical drudgery. Now the AI might generate complex code that the developer then has to spend significant time rigorously reviewing and debugging, which is why the total time on task increases.
SPEAKER_00But the nature of the time spent is fundamentally changed.
SPEAKER_01Precisely. The developer isn't totally exhausted by 2.00 PM from writing boilerplate. They have reserved their deep cognitive energy for the high-level logic and the intricate problem solving.
SPEAKER_00So the AI absorbs the fatigue of the mechanical work, leaving the human with the mental bandwidth required for actual judgment.
SPEAKER_01It doesn't necessarily speed up the clock, but it drastically improves the endurance and the cognitive state of the worker. It stops their brain from turning to mush.
SPEAKER_00That makes perfect sense.
SPEAKER_01And as Pralad notes, fatigue reduction does not show up neatly on a standard corporate spreadsheet measuring tickets closed per hour. But it is a completely transformational benefit for a company's workforce.
SPEAKER_00However, even with these transformational benefits, the source material is very clear-eyed about the existential risks of deploying AI at scale.
SPEAKER_01Yeah, we can't ignore those.
SPEAKER_00Prelod outlines three massive hurdles. What's the first one?
SPEAKER_01The first major hurdle involves the historical skew of data. AI models learn by ingesting vast amounts of historical information. Right. In the financial sector, historical data regarding who is approved for credit, who received mortgages, how neighborhoods were valued. All of that is heavily skewed by past human decisions and societal structures.
SPEAKER_00Aaron Powell And if you feed 30 years of that historical lending data into a pattern matching machine, the AI doesn't understand context or history.
SPEAKER_01Aaron Powell No, it just sees a mathematical correlation.
SPEAKER_00Aaron Powell So it will seamlessly replicate and potentially amplify those historical biases and denial rates.
SPEAKER_01Aaron Powell Exactly. And from a purely business and regulatory standpoint, this is a massive liability. Regulators will not accept, well, the algorithm did it as an excuse for discriminatory lending practices.
SPEAKER_00Aaron Powell Which leads directly into the second critical risk: explainability. Trevor Burrus, Jr.
SPEAKER_01The black box problem.
SPEAKER_00Aaron Ross Powell Right. Because if an experienced human loan officer denies an application, a regulator can sit down with them. And the officer can point to specific debt-to-income ratios or credit history flags to justify the decision.
SPEAKER_01Trevor Burrus But advanced neural networks do not think in a way that is easily translatable to human logic. They distribute billions of mathematical weights across a vast web.
SPEAKER_00Aaron Powell So when an AI denies a loan, asking it why it did so is incredibly difficult.
SPEAKER_01Almost impossible sometimes. And financial regulators are increasingly demanding absolute transparency. They require a clear, step-by-step audit trail for every single automated decision.
SPEAKER_00So if your model cannot mathematically explain its logic to a regulator's satisfaction, you are legally barred from using it for those decisions.
SPEAKER_01You can have the smartest AI in the world, but if it's a black box, it's just a really expensive paperweight in the banking sector.
SPEAKER_00And speaking of expensive, the third risk Pralad highlights is much more immediate to the bottom line, the sheer cost of the technology.
SPEAKER_01Yeah. Running inference, which is the actual computational process of an AI generating an answer, requires an unbelievable amount of specialized hardware and electricity.
SPEAKER_00It is magnitudes more expensive than running traditional cloud software.
SPEAKER_01But right now, it feels cheap.
SPEAKER_00Because every company is buying these $20 a month subscriptions.
SPEAKER_01Exactly. But that is an illusion. Pralid warns that the true costs of this compute power are currently being heavily subsidized by the massive tech companies providing the foundational models.
SPEAKER_00Why are they doing that?
SPEAKER_01Because they are absorbing the financial hit in order to grab market share. It's a land grab. They want every enterprise hooked on their specific ecosystem.
SPEAKER_00Oh, it's the classic Silicon Valley playbook. Give the technology away at an artificially low price until the customer builds their entire infrastructure around it and then aggressively raise the price to cover the actual costs.
SPEAKER_01Exactly. And we have no idea when those subsidies will evaporate. Therefore, every single AI project a bank greenlights needs to undergo a rigorous cost-benefit analysis based on the true cost of compute, not the artificially subsidized price they are paying today.
SPEAKER_00Right. You have to plan for reality. Which brings us to a clear playbook for the leaders navigating all this.
SPEAKER_01The CEO playbook.
SPEAKER_00Yeah. If you are feeling the pressure to show immediate ROI on AI, Pralad's advice is to forcefully resist the temptation to launch flashy, highly visible pilot programs just to get the board off your back.
SPEAKER_01The path to actual value requires discipline. It means focusing strictly on operational efficiency in the background.
SPEAKER_00Finding those unglamorous use cases where the blast radius of an error is minimal.
SPEAKER_01Yes, and most importantly, it means doing the grueling, unsexy work of cleaning and harmonizing your foundational data so that AI actually has something true to learn from.
SPEAKER_00You also have to invest heavily in the people who will be using it.
SPEAKER_01Right.
SPEAKER_00Your workforce needs to understand the mechanics of context engineering and how to treat AI as a flawed but powerful collaborator, not just an oracle they type questions into.
SPEAKER_01Right. The organizations that will survive this hype cycle and emerge stronger are the ones deploying the technology thoughtfully against very specific, constrained problems.
SPEAKER_00And you know, the concept of fatigue reduction we talked about earlier, that leaves us with a really compelling thought to consider about our own careers.
SPEAKER_01I think so too.
SPEAKER_00If we assume AI will eventually succeed in acting as a sponge, absorbing all the draining, repetitive cognitive tasks that make up so much of our daily jobs, what remains?
SPEAKER_01What are the uniquely human skills like deep strategic thinking, complex empathy, the unprogrammable creativity that an AI simply cannot replicate through pattern matching?
SPEAKER_00Those are the specific muscles you should be exercising right now to ensure you remain invaluable in a workplace where the mechanical thinking has been completely outsourced.
SPEAKER_01It's a fascinating shift. We are building machines to handle the processing, which means we are going to be judged entirely on our judgment.
SPEAKER_00Perfectly said. Thank you for joining us on this deep dive. We'll catch you next time.