EP1020: From AI Promise to Performance

DN Prahlad, Founder and Chairman, Surya FinTech

As banks move beyond AI experimentation to enterprise-wide deployment, Puja Sharma speaks with DN Prahlad, Founder and Chairman, Surya FinTech, on why the industry must shift its focus from AI hype to measurable business outcomes, how agentic AI is reshaping financial services, the critical role of data readiness, and the strategic priorities bank boards should embrace to unlock AI's long-term value.

Transcript

SPEAKER_00 0:03

I'm Pooja Sharma of IBS Intelligence, and you're listening to the IBS IViews podcast. With me is DN Prilhad, founder and group chairperson Surya Fintech. Today we are exploring the gap between AI hype and real-world impact in banking, unpacking where AI is creating measurable value, how agent tech AI is reshaping financial services and what banks need to do to turn ambition into outcome. Welcome to the podcast, Mr. Prilad.

SPEAKER_01 0:32

Thank you.

unknown 0:33

Mr.

SPEAKER_00 0:34

Pralahad, there is a growing sense in the industry that AI's promise has outpaced its delivery. Do you agree?

SPEAKER_01 0:41

Yes, and now AI has promised a lot. And a lot more promises were read in by people, which was actually quite unreal. It's important for us to understand a few things about AI. It didn't begin yesterday. AI started in 1936. In fact, as a student, I built a speech recognition system, and one of the algorithms that I use is still in vogue today. So it's not something new, it's been there for a long time. Gen AI, which is the hot topic that everyone assumes is AI, is about three years old. The biggest problem is not with the tool sets of Gen AI or the promises that Gen AI has made. The problem is that the industry and the AI engineers have underestimated very badly the change management required. And in fact, have grossly overestimated the speed of transformation. So that gives the problem of promises not being met. Because the promise some promises are unrealistic and some cannot be implemented because we have grossly underestimated the effort. And equally important to understand, especially with banks and any uh you know financial institution, is that AI is rooted on good data. So the ones that are seeing any real ROI from AI are those that have invested already in data quality and good technical people as a matter of strategy for a while, not just because AI came in, they've been doing it as a business. Finally, I say that AI does not make a good bank or a bad bank. It is a tool set for good banks to become better. And it may not do much for the bad ones who have not invested in data, etc. So, therefore, it's a mixed bank. Some promises have been met, some were not being met. Yeah.

SPEAKER_00 2:45

You have developed a framework for identifying where AI creates genuine value. What does it look like?

SPEAKER_01 2:53

Four-pillar framework that we have thought of, you know. Well, we've been working with multiple customers on this issue. The first pillar I call the information asymmetry pillar. The person who needs information at a given time does not have it. This is a great use case for using AI. For example, I'll give you uh the case of a relationship manager not having access to the correct updated information in time to meet a client. So he may the information exists, but exists elsewhere and is not accessible. That's one type of example. Other type of use case would be a compliance officer or a board member who wants to refer to some regulations and it takes ages to get it. So, therefore, this information asymmetry is one good use case or set of use cases for use of AI because AI solves these problems in seconds. It's not some big D of document storage and data storage, etc. So that is the first pillar. The second pillar is the work should be repetitive and time intensive. For example, I've seen all this data is gotten from the system. Somebody sits creating a quarterly summary for investors for the board, involves working with a load of Excel's which are the output of regular systems. It's the same work every quarter, but file names have changed, and there are some small changes to these Excels that are being loaded. Now, this is repetitive and time-intensive work where use of AI will crossly bring down the time which is taken. And in fact, in some cases, we have noticed improves accuracy as well. And therefore, this is another use case for employment now, deployment of AI. The third pillar is the task is very valuable or the outcome is extremely valuable, but is it feasible without automation? For example, if you want to do a sentiment analysis based on a lot of customer calls, it's physically impossible for someone to hear all of them and mark out, you know, get the data saying A was happy, B was not happy, etc. etc. So, therefore, this would not be feasible at all without automation or AI. So, this is another type, the third pillar that I call it. The fourth pillar is more of a watchdog or a guardrail, if you will. The cost of employing AI for these should not make it more expensive than it would have normally become. Because the cost of AI today is by tokens and are not related to outcomes. It is like the beginning of the internet when we used to pay by megabytes of uploads and downloads and did not have any uh you know relation to the outcome of what we were doing. So, therefore, it's important to estimate the cost of AI as a guardrail whenever you're using these pillars. The fourth is really not a pillar, but some kind of a guardrail, and ensures that ROI is considering the variable cost that these AI is paying.

SPEAKER_00 6:05

The concept of AI agent is gaining traction in financial services. How real is this and how should banks prepare?

SPEAKER_01 6:13

Okay, this is extremely real. I mean, they you know, for all the promises not met, story, yeah, is here to stay. Like I said, it's only three years old. The LLM or the Jan AI, as we call it, is only three years old, and there's a long way to go to mature in multiple ways, both in the business sense and in the technology sense. For example, I understand that Morgan Stanley and JP Morgan have deployed agents for research synthesis and contract review. So this is quite relevant. Let me describe what is an agent. An agent is a goal-oriented uh entity. This needs precise instructions, relevant knowledge, and access to the right tools. The whole three things put together, it's called the context engineering. There's a new type of engineering that we have to deal with these days. The models, the LLMs, are themselves commoditized to a large extent because you know we can use multiple LLMs to achieve the same task. So without much difference. So they become commoditized, and you're not going to see a place where the fact that one LLM has done something better than the other one is going to be having a massive difference, except in specialized niches. So now what should the banks do? That's a wonderful question. Any financial services company should do. They have to go and invest in this context engineering as it were. That is knowledge management, tool integration, and orchestration layers. But we also have to understand that AI is not foolproof. In the sense, whatever we do, AI is a copilot. We are in the co-pilot regime. We are not in an autopilot regime. There's physical AI which is in the autopilot regime, but the Gen AI that we are talking about is in the copilot regime. This big difference has to be understood. And especially when you go back to the fourth pillar, you have to add the cost of the expense of the human in the loop as well into the uh into these cars. So, I mean, to summarize, context engineering is where banks should be investing.

SPEAKER_00 8:31

Data is often cited as a biggest barrier to AI success. Is that a solvable problem or an inherent constraint?

SPEAKER_01 8:41

Of course, I mean, if we created the problem, it's solvable, right? So problem as one of our own making. And obviously, it's a solvable problem. Question is how much does it cost to solve the problem? You know, BA's foundation, as I mentioned earlier, is data. And the data value chain is that one which goes up in value as we go further down the chain. So the chain that I have conceptualized consists of capture, which is stage one, harmonization, which is stage two, stage three is aggregation, and finally, for agree out of anal aggregation, you get analytics, and finally, what we call knowledge. That is the ultimate Himalaya for the data value chain. Now, if you look at the institutions, most of them have got this capture right. You know, we all do transaction banking or transactions with all these institutions, mutual funds, whatever you have, brokerage systems, etc., etc. All of them have got the capture right. Otherwise, they won't be in business, correct? Now next step is harmonization. The this being inconsistent. As are the other paths, you know, aggregation and finally knowledge. Teaching knowledge is extremely difficult. To get to the knowledge level in the data value chain, there needs to be a battle plan. It's not a simple, you know, saying that I'll turn a switch and it comes on. No, it's a battle that has to be fought. It may take a couple of years to get there. So there has to be a massive, well-considered and well-funded plan. Now, compounding the problems that I've described earlier is the data infrastructure. They're old transaction banking systems which are here to stay. And barely functional, you know, almost failed data warehouses, which is a it's a fact of life because we store data much more cheaply today, and much better formats exist for storage and retrieval of data. So this also renders the cost of modernization extremely high. That strengthens my case for having a well-considered and well-funded and well-orchestrated battle plan to get the data right. An institution that understands all these values that the data value chain brings and considers data quality as an investment is more likely to unlock genuine AI value than those that don't. That I think will be the big difference. So you're absolutely correct. Data is a big barrier. Lack of aggregated data, lack of data that can lead to knowledge is the barrier in most uh institutions.

SPEAKER_00 11:19

If you had 30 minutes with banks, board, what three points would you make?

SPEAKER_01 11:24

Yeah, generally I get 15 minutes, so it'll be great to have 30 minutes with any board. AI, first of all, is a tool. It's not a banking strategy. Banks have to lend money, banks have to uh you know, get money from the people. That's not going to change because of AI. Uh, I will go to a bank that gives me more money as deposits, which is of given reasonable quality. Uh, if two banks are of the same quality, I'd go to the one that pays more, and I'll take loans from the one that gives me uh less interest rate, right? I don't mind waiting one day forever. So, therefore, AI is a tool, it's not a strategy at all. That's something that boards have to understand. And all AI investments, like any other investment, has to have a relation to an outcome. Because you know, there's no point in just saying I'm AI driven. You know, maybe there was a time in life when people said we have electricity, and that would have been, or people thought that's a competitive advantage, but sadly it didn't turn out to be because everyone had electricity. So, therefore, it's it's how you use the tool, not the tool itself, that becomes relevant. Again, I go back to the data question. The models are not your competitive edge. The context that you create out of the data and the processes that are there in your institution and outside of your institution is the competitive edge. For example, of you may ask me, what do you mean by knowledge of the process? For example, there are probably 30 odd ways of calculating a discounted cash flow. But then you may be using a certain method in certain context. And the model itself will have all the 36, so it will pick any one of them. But if you want to pick the particular one that you want, you would have to then go and say, I want this to be the way to calculate DCF. That is, in terms of you know what we call skills, you have to write the skill into the uh your your context. So, therefore, data is the competitive edge, not the models themselves. That is something that boards need to understand. And we have to convey to the boards, I would convey to the boards that the success of the outcome. So, what makes a massive difference to the outcome will be context engineering. That is, get the data fine, get the data to an aggregated level, get all your processes and uniqueness mapped out, and the processes which are outside that you want to bring in, make that a part of your skill set which you are feeding into the LLM. And this context will be your intellectual property, and the context will be what makes a huge difference to your outcomes. Don't worry about doing research on 100 models. The model is a machine, right? It's a machine like any other machine. We don't go around calculating, you know, whether Mac this was better than Mac that or Windows this or Windows that. We we say, all right, you know, we take it as a commoditized uh option for us. And similarly, the models are not the big deal. The context that you build is the one that makes the difference to your outcome. Then there's a question of deployment strategy. Where do you go and deploy this AI? It's not an easy question to answer because most people tend to think that the best way of using AI is customer-facing application. It is typically the chat board that you find on every other website. And these are at best unsatisfactory, terrible in the worst case. So these are not the ones that we should be deploying in the initial stage, though competitive pressures may ask you to deploy one. First, if I were the bank's board, I would focus on the getting the internal applications built first. Whatever inference engines we need to get. For instance, I talked about quarterly summation, these are things that we should focus on. Learn from these experiences. And finally, when there's enough context knowledge, enough data which is harmonized, then we can move to thinking of customer-facing applications. And lastly, you know, it's a very uh non-intuitive uh message that I want to convey to the board. You have to invest in people. The way people work is going to change drastically over the next five to ten years because AI will have a role to play. Let's go back to the internet example. When the internet came in, the way people worked became a lot different. So if there were people who could not adjust to the internet, they would be out of place today. So that is the way the AI chain train also will be. That is, we have to invest in people today who are going to be comfortable using AI, comfortable saying that yeah, there are times it does hallucinate. Because hallucination, the way I look at it, is a feature, it's not a bug. Yeah, we can have another podcast on that some other day. So you need to invest in good people and you need to invest in talent to the new era. That about sums up my message. And uh yeah, seriously, it will not take more than 15 minutes, I guess.

SPEAKER_00 16:41

DN Pralhad, founder and group chairperson, Surya FinTech.