#40: Remaking banking with AI

Today we're going to talk about highly regulated banking industry and how AI is changing that. Who are you?

Hey, I'm Manish Gupta, co founder of Corridor Platforms and whose main goal is how do you introduce AI and gen AI into regulated industries, specifically banking. I spent most of my career in banks thinking about how do you innovate in a banking environment, which is a very responsible environment and it's a very interesting problem to solve. So we started a company seven years ago and hoping to help banks across the world to really innovate and take advantage of the latest technologies.

Years ago I visited new bank down in Brazil which is now a public company and they built their own customer relationship management system to make the service of the bank better, which was an unregulated piece of the banking industry. Right. It didn't have to do with your money or what you're promising the customer or knowing your customer. All the things that your industry has to regulate. Is that where your first. Well, where are you starting with AI maybe is the best question. Are you thinking of service or are you thinking of actually going in and making the numbers better with AI? Because AI is not quite trustworthy yet, is it?

Yeah. So I'm no longer at a bank, so I am at a company that would facilitate a bank to start doing the kinds of things that you're talking about. And banks around the world are experimenting on various use cases, some of them internal, where they're trying to improve the efficiency of employees with copilot or how do you do better research, etc. Which have lower risks as you identify. But inherently, anytime banks do anything that is decisioning a customer or is facing a customer, almost everything that they do is fairly if not regulated by the government, is brand regulated because very difficult because we banks play in a world of trust and customers give them a lot of information. So if they're interacting with them, they have to be very careful in that they don't create any bias, they use the data properly. Every interaction is something that they would have no problem with if it is on the front page of the Wall Street Journal. So the issue is going to be how do we take the powers of genai, which have transformational abilities, but then make them available to be used at banks. Now when you come into the high roi, high impact use cases, most of them are facing the customer in some way, shape or form and therefore need to be very carefully moderated to be very responsible. And therein kind of lies the rub. What do you do to move from experimentation to production so that you can do the right thing by your customer, brand and regulators. Obviously.

Yeah, this is a new world. I mean, just on knowing your customer. Grok, for instance, an AI on X.com, if you ask it the right way, it'll give you my credit score. And it's pretty accurate. It's amazing, right? Because it can figure out by looking at all the data I've given X over the years and I've been on Twitter or X for 17 years, it can figure out, oh, paper. Based on all of his responses, his credit score is probably around 800. Right. And that's a new thing that a bank might want to use to further improve their credit scores. But now you bring in the bias of human beings. Right. If you don't have a good credit score, you know, and you're in the, in a bad neighborhood or whatever, you can get biases and you can get organizations to be biased against you in a bad way. We know this from the past, right? Supreme Court has made it so you're not allowed to bias your, your bank that way and, and redline neighborhoods. Right. People who live in a certain neighborhood go and get loans, for instance. Right. It's an interesting world.

I, you know what, where are you.

Focusing most of your efforts on with the banks? What, what are they trying to do right now to improve themselves, technology? And how are those early experiments or trials going and where are the challenges for what you're trying to do?

So I think banks already are adopting and adapting to generative AI because everybody sees the power of it. And a lot of them have created a fair amount of focus internally to explore how they could better customer journeys, how they could create efficiencies. Where we are really focusing on is how do you think through governance within a bank so that you can understand all of the risks, some of them very new. When it comes to generative AI like hallucination or the generator part, how do you measure them then? How do you then create a solution within the risk tolerance for that particular use and go through what we call second line of defense model, risk management, fair lending, bias checking and say, okay, we are okay to go into production? That process in banks has to be very rigorous and is also very quite manual so far. And that was okay when they were small data, reasonable models, reasonable insights. But when you're thinking about big data and somewhat opaque large LLMs, they are very, very powerful. But now you have to really figure out how to automate this. How do you make checks very instantaneous, how do you ensure what checks to do. So what we've built is a platform where if a bank is building a use case, how do you let teams kind of collaborate but also govern, test and measure and make it through all of the kind of required checks, then put it into production and then continuously monitor it. So think of Corridor platform as a responsible AI governance platform. And that's what we are trying to bring to the table because there are a lot of smart analytics people and a lot of tools and we've just announced a recent partnership with Google. A lot of banks are using LLMs from Google Cloud and many other hyperscalers. So there's a lot of availability. The question is how do you do it within the guardrails? And we want to be the guardrails that will facilitate it, which is really what's between the barrier, between unleashing the power to just experimenting with it.

Yeah, I mean I'm thinking through all the times I talked to people who are running banks, New bank is one of them. But there's different parts of the bank, right. There's the marketing part that does the website and decides on how the customer service is going to team is going to talk to customers. There's the, the credit scoring part of the bank. There's the knowing your customer part of the bank. Right. Because when you call the bank it knows that you're talking to it with a voice system and a defraud system. Right. And there's a whole bunch of things going on there. Which part of the bank should we focus in on for your business and and then talk about your tool and how it really helps the bank improve that one part of the banking system?

Yeah, I think Gen has applications throughout, but you're correct. Let's focus where I think the biggest impact is going to be, which is actually going to be in customer interactions of all kinds. Yeah, right. So whether a customer is calling about their product, their billing, their card, their statements, or they want to do something or get another product or a mortgage, the lot of calls are happening. And one of the issues that banks face is it's when you're interacting and there's an agent on the other side, the agent needs to be trained on a lot of things. Banking plus a lot of internal materials and policies and there's a lot of churn. So in general, it's been something that could always need improvement. Now with the ability for Gen AI to be taught to be almost like a perfect mind, you can really do many things, right? You can, if you can train it properly, you can ensure the customer gets the right answer really, really quickly. Two, you can ensure, you know, it's highly trained and it's always consistent. Right. And three, which is a byproduct you may not need as many agents, which is a very, very large efficiency save and cost save. So those use cases are really, I think some of the most high returning use cases and probably the most impactful and probably the most in need of understanding the risks and management. So if we could crack that, so we're really trying to focus on that, then if you crack that, you can do any of the internal ones because those are just little less risk and a little less testing. So that's where we are focusing on right now.

So you're really trying to help the banks compete with a new bank like down in Brazil. I mean, the reason their stock is doing pretty good and their company's doing pretty good is they built their own customer relationship management system, which means when a customer calls into the bank, they don't get into a phone tree. Right. The other banks still use an old phone tree. Oh, click one if you want to talk to real estate. Click two if you want to talk to, you know, auto person or something like that. That reduces customer satisfaction numbers.

That.

Right. And they didn't have that. They had a, a system that would listen to you talk and route you right to the right person so that you didn't have to be on hold and on a phone tree. And that really improved their customer satisfaction numbers. And that's why people, when I, when I visited there, they had six months waiting period to get their credit card for a lot of reasons. Because their customer service for dealing with the bank was so much better. What's the problem of trying to do that? Maybe go into the architecture of the. Because I assume you're building a whole orchestra of models to run these voice agents that are going to talk to customers and route them to the right thing or get the right answer. You're going to have to check that to make sure you're actually giving the right answer. Because if you give a customer a wrong answer, like, oh, your loan's approved and it wasn't, that's not good for anybody. And that gets you into regulatory problems. Or if you promise something or you're dealing with the investments, right. If you're starting to, if your AI is starting to tell you, oh, you should buy some Apple stock today, well, if Apple stock goes down tomorrow, you have a lawsuit. Right. You have to be very careful. And that's why the banking System's so regulated, right? Yeah. Give me a sense of what's underneath the covers and how is it. I assume you're building voice response agents then to answer the calls and stuff like that?

Yeah. So the banks are. Because I just want to make sure we are a company that helps banks do it. So the banks are. And I think, I don't know when you went to Brazil, most of the modern banks or larger banks have already invested in things that you are talking about where when a call has come, it gets routed to the proper agent. But I think that can only be bettered so that they're routed more and more accurately to the right, right agent because they have a set of intent classifications that they can do. And then they say, we don't know. And it goes to a common agent who then does it. So you can improve that for sure. But now take it to the next step. And the agent is receiving the call, helping that agent as the person is speaking to give the right answers very, very quickly. That's what we call agent assist. Now you take it to the next part that even an agent is not needed. You can just ask and get the answers directly and therefore and get them accurately. That's what we call conversational ivr. And the banks are working through all of these steps because each one of them is creating better customer experience and better operational efficiency. And each one of them creates its own risk. When you're just creating an intent, there's only the risk of getting an intent wrong, which is not a big risk because then in the end you got to somebody who put you somewhere. Then if you are with an agent, your agent is in the middle. You're informing the agent what to do and what to say. And there does come a risk there because some agents will say exactly what they see from the help screen. So there is immediately that risk.

When you say, can you be clear that you're talking about a human agent that's answering the phone? Yeah, because there's AI agents. You're helping banks build a system for the human agent that's actually answering the calls to tell them, maybe even listen to that human agent in real time and be saying, hey, you just said something wrong. You know, don't do that. Or here's some things you really should remind this human that called when you're talking to him. And you know this, when you, when you deal with the banking system on a real estate loan or something like that, there's a lot of things that that agent has to remember to tell you. About your loan when they're working through it. Right. And they have to do it in the right order and they have to be very exact about it otherwise they're going to get in trouble. And your. The system that you're helping build can show them on the screen in front of them. Hey, Bing, you. You said this, you reminded them of this. Bing. Great, great job.

Right, that kind of thing. Exactly. So that would be what you call helping human agents, which is called human agent assist. And that itself, when you build that using a gen, you have to make sure you govern and manage what that model does. Test it thoroughly beforehand that it says the right thing, doesn't show information of the wrong person, does not do something wrong or create some bias in a recommended answer. And that. And now the banks are going to the next level too. Some are to say become a full agent. Right. So. And in that case, there's no human in the middle that could potentially correct the issue. So you go to the highest potential benefit and highest potential risk. So all of these are being worked on. And in each one of these you have to think about if you're starting to use Gen AI, it's potentially much better and easier to do using large LLMs and prompting and using rags. But then there's the inherent risk that comes with Gen AI to the extent it can hallucinate, to the extent it can give a wrong answer. So think of Jenny. I think of Gen AI as training a teenage mind. You can tell them what to do, but doesn't mean that they will always do that. Right. So then if there is that thought process which also is the power of Gen AI, then the question is when you have very low tolerance for risk, how do you create the model itself and test it so that it's acceptable? But then how do you also create guardrails and how do you test that whole system? So the idea is to have a platform like Corridor to kind of put all the components of these guardrails and what you're doing to control. Then test it thoroughly in a simulated environment. Even tested by simulating a conversation and trying to break it, get it through what is strict model risk management and fair lending criteria that regulators put. In fact in Genai. They haven't yet come out with all of it, but they certainly. We know what are the risks that are there and what are the new risks. Show the regulators you're on top of it. Then you're saying you're going into production and then now you're into doing it and monitoring it and so the platform that can govern and help test all of that is what we are. Banks are already being supported by a lot of hyperscalers and LLM builders and internal to kind of build the solution. The idea is, can you get comfortable to launch it? Yeah.

Well, I assume you have a pretty intricate system to take to the first meeting and say here's what we're doing to attack each of these problems. Because I bet this, this industry is very resistant to change. Right. It's my. And I'm hearing that from a lot of enterprise people who are taking AI into enterprise that oh, their first trials, you know, give some problems like hallucinations and everybody goes, I don't like this. Right. How are you. What, how are you. How are you changing the culture of the bank, you know, and getting everybody on board to get one of these systems, you know, all the way into production where it's actually talking to customers Because I bet that's a tough job.

Yeah. So that's our somewhat usp, right. Because we spent all our lives, me and my co founder. My co founder is Ash Gupta who was chief risk officer of American Express for a few decades. I spent a lot of time just thinking about this problem inside. How do you create the change management? How do you create the governance and banks, you know, banks are slow for sometimes a very good reason because people are trusting you with a lot of their data and money. So you want to make sure. And then obviously when you're in that system, you also get bureaucratic. A lot of checking the checkers and it can get out of hand. But it's true, getting change into banks is very, very difficult. And it is makes it even more difficult because there are multiple groups touching the problem and a lot of manual. So automation is key. Showing them that you can now automate a lot of the things which required manual intervention, which cause inefficiencies and time is key. Showing them that you can actually devise and standardize these tests and iterate on them so that you can get comfortable. So putting a lot of thought into what are those tests? You know, we have one of our investors and partners is Oliver Wyman who has a lot of regulatory or governance experts in AI and gen AI. So we've collaborated with them to think through what are the risks, how do you test them and what are these have not been mandated yet, so we've included that. So I'm happy to show you part of the platform if you want to show that.

Please do. I love demos. I like to see what you're working on.

Let me, you can share screen. I was, okay, so I'm going to share my screen and hopefully just go directly to the platform. Okay, so you are seeing the platform, right? So this is Gen God X, which.

Is, you can, you can minimize us if you want.

Okay, there you go. So yeah, this is gengardex. It's a digital shared platform. The idea is if you're building and thinking through some of these solutions, how would you kind of take it through governance? Right. So one, when you're entering the space, anybody who enters the space in admin here is got a role so they can see and do only what they are meant to do. So they could either be people who are putting the use case and models in place and there's somebody else who's an approver and all of that is role governed and managed in a very auditable manner.

And what role are you logged in? Because are you logged in as an executive or a frontline employee or.

I'm logged in as right now as an admin who has all rights. Right. So if you go here and there are roles, there are different things. I'm logged in as an admin, but in an analyst they could be doing only read and write and not approve. And there is an approval workflow which says if an analyst is doing it and he wants to get something approved, who does it go to for approval? And then there would be a role for an approver who would have approval rights. This is a supervisor. And then you can also control what data and what is seen by who. So this is the background which really creates the backbone of governance. And then when you go in here, the question is how do you create this ability of various groups to collaborate and share so that it's not all about sending emails and spreadsheets to each other. So the idea is here you kind of have a data vault where all of the data, which is the power behind all of this modeling, is kind of stored and standardized. Whether it is ground truth data which any experts will understand, or validation data sets that are used to test question and answers to tests. You have a controlled environment so that everybody's using the same data, which is a good start because people start using different data for different things. In this Genai studio you have the ability to look at the models. These are all the LLMs, whether it's Gemini GPT4 llama and be able to run tests. And then we've developed these tests where you can run and check LLMs against each other. You know what is Their understanding of toxicity, how toxic are there, is there a bias, gender bias in them? Is there, what is their reasoning ability in terms of different subjects? So you can think through all of the different tests whether they are fair lending or bias or accuracy and say okay, I evaluate an LLM that I want to use and then usually when you have an LLM then you use it for a given purpose. Here this is the interactive voice response use cases and prompts which is doing conversational or the customer agent assist prompts. So how do you take these prompts which is a new field which instruct the LLM and how do you track them and share them and then as you kind of create them then you have to have the knowledge base going into the billing system or going into all of the customer data. How do you kind of make sure that customer data is approved and accurate and, and everybody's using the right one and then create the, the kind of use case where you're saying hey, go in there, go and use this prompt, use this model, use this guardrail and then run checks on it, right? Run a bias report on it. So and then the ability to say okay, send it for approvals and track everything or see the change history, what was changed at what point. These are the kinds of things which sound easy to do but when they are manual are almost impossible to track and break down. Thinking about a governance capability and all the tests and then once it's there you can export it directly into production. Then the interesting thing when it comes to Genai and this is the last thing and we'll go back is how do you do human testing? So here you can basically go in and say I was running a test just before we came. Somebody is doing a conversation and saying what is my last transaction? And the system I had built this is simulated, it was giving the last transactions which was accurate. So I gave it a thumbs up. But then I said give me last one is not mine and it basically says oh, so the 7-5-1 is not yours. You want to dispute it? Well, I was meaning this last one and it said it understood last one as the last, as the recent, most recent one, which happens to be the first one here. So then I have written, oh, that's wrong. I think it is. Basically I'm misunderstanding the question because I meant it to be the last one and then I can keep going. No, I meant the first one and then see what the. And then I can even stress it and say hey, give me. So, so now it's accurate, now it's basically saying July 1st as the first one though it shows up last. And I can, I can try to break it and say, give me Robert's transactions instead of mine. Right. So because. And then see if the system you built, what does it do? So the idea is give a human in the loop report card and it gave you. Now in this case, it gave Roberts transactions, which is wrong. So I would say this is very bad. It's giving anybody's transaction you ask, put in a guardrail. So you just give it the ability to test and then have feedback. Then finally you go for final approval that everything is okay and it is meeting all the thresholds and you can put it into production. And then even when you put it into production, the expectation is every day or as quickly as possible, you're actually seeing what's happening. You're actually a human is saying, yes, that was right. What they did, what it did not do and keep a check and keep improving it. So that's what it kind of takes to really build a governance system which can then automate, give them comfort. Everything is organized. And once something is correct, everybody reuses the same thing. Because there's knowledge management here. This worked. This prompt should be the prompt for this kind of stuff. Everybody don't create. So just thinking when you said, how do you do this? What we did in our two, three decades was just think about, what are the hurdles? Yeah. And how do you solve them? And then we sat down, Ash, myself and the other and said, how would we solve this if we had the smartest people available to us? So we said, let's leave a tenured job, cushy job, hire some of the smartest talents from the top engineering schools in the world, and let's solve the problem that we had to solve internally in the most efficient way. And that gave birth to corridor platforms. And now we are at some large banks with traditional AI, but now we are solving the problem with Gen AI.

Yeah, that's really real interesting. It shows the kind of software effort you have to make before you even go in for the first sales call. Basically because you have to be able to prove that this system works and works in this highly regulated environment. And an environment where a lot of business risk and a lot of resistance to change. Right?

Yeah. That is what you need to do. So this is not where you kind of go and talk about it before in this industry. It's more like we've been doing it for 30 years in a very reputed firm and therefore we can take our knowledge and replicate it and apply it and know exactly what. That's, I think, the reason of our success. But more importantly, that's what's needed.

Yeah.

To kind of do this and get it through, you know, getting banks through. The other big problem that banks have is a lot of the times they want to do everything themselves. So this is where this is a new change, which I also took a long time. These, the advancements are happening so fast that it is best you adopt them. But then it's also best that these kind of fintechs like ours are really focused on this problem and are keeping up with what are the new risks, what are the new tests, and we're giving it to them. They should still understand them and obviously approve them. And therefore this partnership with Banks can keep them running as fast as the advancements. I think there is no other way. And that's why we thought there was a huge need. There's no way banks can keep up. They need to do what they do.

Well, I can. I'm watching the screens right now with AI industry gone by and every few minutes there's something new. I mean, it's really hard to keep up with the model changes and all the new innovations that are coming along. I bet that you're doing a lot of logging underneath the system as well. So that if either an AI agent says to the customer approves something, or a human agent does, you're logging everything and logging everything. Everybody who sees that, I bet. Right. Or everything that touches that.

Correct. You hit the nail on the head. Because when you create a platform like this, your main thing is anything that's changing is fully recorded so that you can learn from it. Then any answer that's coming, any data that's coming is getting logged. Because it's a constant improvement thought process. Every time you see something or this change made this benefit or this improvement, then you can easily record it. And it's not manual and all over the place. When there were simpler models, it was easy to do. Now the idea is to constantly give it feedback loops with humans in it. In banking, I still believe humans still have to be able to judge in the end. But make it so that everything is done and you have all the checking going on and then go on. So yes, the very important part of our platform is logging every step and also any learning so that it can be done that. Yeah.

What are you learning by building this kind of company? Because the AI is changing so fast, I bet you're being surprised by the new models that are coming along what are you learning by building this company?

Yeah, it's been a phenomenal ride. Obviously it has also been one of the most painful journeys to become an entrepreneur at our stage of life. Not that I'm old, I'm fairly young at heart, but I think the amount of learning is what keeps you fully excited. And there are small things when you're working with clients or you're working with your partners like Google Cloud and experts there you're really understanding the power of these LLMs. And I am convinced, as I see more and more before I was slightly skeptical that they have huge ability and they're becoming better, faster than we can even imagine.

Everybody is bringing new models in, the.

Next models and smarter models and cheaper and they become cheaper and cheaper. So now then the question is how do you kind of use them? And there are nuances to this which are very, very cool. Like, you know, I got amazed that when you are telling an LLM, hey, give it examples to categorize things into abc, if somebody says this then it's this kind of, call this, this kind of, they give you examples and then you basically said do it and they're doing it. And then suddenly we found a lot of the mistakes were when we hadn't exactly classified the question and what we had said. If you don't know, in the prompt you say if you don't know, you can say I don't know. But unbeknownst to me, at least the LLM mind, that's why I call it a mind, felt like if I say I don't know, it's not a good thing. So adding one line in there saying hey, if you don't, if you say I don't know, there is no penalty. That's just one line. Suddenly the model improves to a phenomenal degree because now every time it didn't know it was trying to guess, it doesn't guess anymore. So you there, this is new kind of understanding how to use and create the use cases, understanding the risks. And obviously there's lot of risks like prompt injections. You can tell the mind forget you heard this, forget you did this and do that. So there's a lot of nuance advantage in this. Just giving some examples will suddenly give you a fantastic model. You don't need the big whole AI modeling team but now you have to figure out how to control it. So I think this is going to be a fast learning exercise and hopefully platforms like us will give the regulated companies the comfort to learn and iterate so that they can use it. Because if they don't, you're right. Some of the other banks who figure it out, the landscape will change because it's a digital world. Nobody is saying anymore to Robert, if I don't like this bank, I can't move. Before you couldn't move. Now you move like this.

Yeah, right.

So this is going to be a key transformation and people who can adopt and adapt that. I think the banking landscape will change very fast.

Yeah, it's, it's amazing. I could talk to you for an hour about prompt tips, you know, because you, you already gave me one. Right. I didn't know that, that you can sort of tell the system, tell me if you don't know. Right. It's okay. What other kinds of things that have you found that's, that's interesting about LLMs as, as you and your teams have worked with them to tune them up and give them, to get them to give you better answers?

There are many such things. I'll give you one more example. Then we can have a more private, private conversation. But you know, it's the, the LLMs itself are becoming smarter and smarter. Before I was of the only about a year ago, you know, things are moving so fast. I just say a year ago where I thought people like banks would have to build their LLM using their own data because that's the safest way to do it. But now large companies like Google, Cloud, etc. Know that and, and they are building models responsibly, but they're also giving models lot of information about general things, but enough information about banking that you could just ask them a question and they most likely would answer it correctly. And that is surprising. Right. So you can now ask a model, hey, you should be a fair lending expert. People don't even know what that is. But the model, because of what it's been traded on, knows. And please, if you hear something which has got an inherent input which may create bias, remove that from the next step. So you could even say something like that, which is surprising to me. Now you're talking to a model in writing. Just behave when you hear something before you go to the next step. Remove anything. And you have to say, this is your Persona. You're a fair lending expert. So now the model is focusing on all its knowledge that it knows about fair lending, which we didn't know they would know. But then suddenly it's removing all of the references to gender and race and making it out of the utterance and then say now go to the next Step without it and give it. Tell us what it is, what is the answer? So now you've created another so. And I didn't know you could do that. So I'm learning. But these are just tricks that modelers will learn. And then there's the whole thought process of LLMs being checkers. So another part we are doing in our governance platform is humans don't have to check. Another LLM could be IU are the best checking guy. So you can automate the checking. So you know there is endless potential. There is no, it's just like, it's like a, what do you call it when you do that mirror after mirror, you see like the, you just, you can do LLM into LLM kind of things. And then the whole part is how can you make sure you govern and safe and check and be comfortable and you can unleash. And that most exciting part, which I feel like our main mandate is this just to remove the barriers, right? Yeah.

How many different components are running? If you were running this at a major bank, at JP Morgan or Wells Fargo or something like that, how many different components, how many different models are actually running underneath to check each other's answers or do various different tasks? Because one model might be better at doing the fair lending, you know, Persona, and another model might be better at doing math or something like that.

Right? Yeah, I think like banks are experimenting and they're different models, but they're by use type. So let's say you want to do one uses code generation or code transformation, taking old code and COBOL and SAS and moving it to Python. And there's a set of pipeline that does that. There's one around looking at documents and summarizing them and giving the right answer, I think. So by use case you have different pipelines. And I think the thought process that you're asking for is, is pipelines and use cases are also got different blocks. A pipeline, what we say, which is a concept is you have a foundational model, which is an LLM and you have a choice of many or you can build your own, you can fine tune it, but you have a model, you have a prompt which is instructing the model which really becomes the one that you control. Now before you used to build a model, then you have what you call knowledge basis which says hey, instructed, but here is the information from our bank that should be used, which we call rags. And then we have guardrails, which is, hey, if it ever said something profane, check it against a certain set of words which is not LLM or another LLM is going to check. So you have these various components and you stitch them together for a given use case. Right. Because you're prompting it to do something and as soon as you do and then the checks for that right now are, can be done by LLM but also by metrics is what you have to really understand what is the bias check? How do you check this pipeline for bias? Right. And basically it's interesting in Genai checking is also question and answers. So you basically ask the question and see the answer and you, you know, you ask a question and you add it's a woman asking this question, it's a male asking this question and see if there's any difference in answer the same question. And then how do you then evaluate that? And then humans should check that always. I think in the end in banking I would never remove. It's always about how a human can use the tools to finally get fully comfortable and then say do some human in the loop checking. Yes. So I think, I don't, I don't think that answered your questions but I think it did. It did, you know complexities and I think banks are experimenting with a whole lot of them.

Well, thank, thank you so much for spending a little bit of your time today taking me into a place where I usually don't talk about, you know, banking. I'm not a banker and you know, every once in a while I talk with somebody in the industry or have to deal with it and so it's really interesting to see how AI is starting to change the banking system.

But I thought with you. Yeah, which is important. I talk about banking but nothing about what I say or our platform is banking specific because at this point whether it's healthcare, insurance or even when you're talking regular interactions, consumers really want to know that their data is being used for the right purpose and is not being misused. So I think this level of governance eventually especially because it's Genai based will be has to be prevalent everywhere. Right. Whether you are trying to do something wrong on a social media or a Gen AI agent is doing something in a normal conversation but is biasing it. These are all very relevant. We are focused on banking and then we'll expand to health care insurance where it the mistakes could be, you know, very, very sensitive to what impact that they have on a human. But because it's all about data being used. You see that same issue in social media. You saw there was an app where they were talking to the person and eventually that person was kind of suicidal. So is that not going to be regulated in the end? How could you. So I think eventually we are hoping the platforms like ours will help many industries and if we make it easier and simpler, why not do the check before you put it out? Right. So that's, I think that's the goal. That's why we are excited about what we are doing because it is the enabler. But enabling is going to be very critical in Genai. So even though it feels like it's one spot focused area, we feel it is what will really get this all going and accepted by the world. Scoble. Can I ask a question, Arjun?

Yeah, you just jumped in a. Is my, my friend who runs a PR company a great person? What's going on? Good, good.

So Manish, I have a question. So you talked before about first of all the need for adapting, for banks to adapt to a new future. And I guess my question to you is what would you say to the bank that is afraid to take the next step to utilize Gen AI? What would you say to them to convince them that they need to take that step, that it's a necessary step? What I would say, I've been in the banking industry for my life and you should be afraid if you don't take that step because most likely it's an existential issue. Like I said, banks are digital and no longer have captive clients or members and people will move to where the best outcomes are, where they get their best answers quickly. And I think this will happen pretty fast. The other thing I would say is banks feel like everything will be slow. The way this technology and speed with the speed is moving, the people who are moving with it fast enough will take the best clients.

Right.

We've already seen that a little bit in other places. And so I think the landscape within five years, which is a very small amount of time, and how bankers think will be very different and nobody is there forever. You know that in the dao most companies don't exist. That existed but now I think banks and other regulated entities all have to change. So that's what I would say to them. The other thing I would say to them is the main difficulty that banks will have having done innovation in banks is they are by nature fairly siloed between data analytics, governance technology and something like a Gen AI solution. Kind of cuts across a lot and customer management and marketing. And so there is this whole thought process of how to work together, better collaborate. So some of the tools are collaborative in our platform but there's also a lot of ownership and politics which comes in the way. And so I think this is a leadership issue of truly figuring out how do you cut across silos and create the ability to drive change where each of the people would have to be involved to kind of drive. So. And that's a much more complex problems. I think that's what's going to stop. So in fact, I think for the first time, smaller banks may actually have the advantage. Right. Because the beauty of they can make decisions faster and the LLM is already there. It's not like you need huge modeling talent so that you don't have the ability to compete with the big guys. As long as you can be nimble, get some talent to kind of think through the pipeline, which. And the beauty of Gen AI is I think it democratizes AI. Traditionally I was difficult because you had needed a lot of talent to really understand and build the model. Now the model is built, it's difficult to build models and therefore the models are being built by the best in the world. Now how do you apply it, tune it, prompt, engineer it? That's not a, that's a talent issue, but not a scale of talent issue which was an issue. So it's all about nimbleness, leadership, cutting across. And traditionally that's very difficult in large companies, bureaucratic companies, companies where, you know, discipline was very important. So it has its positives. So I think that is going to be a very, very important challenge and we'll see. We would love to help people to kind of do that.

This kind of system can really let a bank add on new, new features that are really cool. I mean, back to the new bank I started with, they, they had a iPhone app that had a lot of really cool features that the other banks weren't using back in the early days. Right. It took a while for the older banks to react to the trend of the customer being on a mobile phone and all that and rebuild all their systems to be able to do that.

Wow.

It's been a pleasure talking to you. I could go for one. Thank you. Where did we learn more about you and your business?

Well, we have, you know, Scorriador platforms and the new business is new platform solution is called GenguardX. You can easily just Google it. I'm Manish Gupta and you can easily on LinkedIn reach me or you can reach ils. We have, we just announced the partnership with Google Cloud and launch of GenguardX also on social media. So totally available and you can reach Robert and hopefully he can direct it.

I'm on X usually once in a while. Other Things, Facebook and LinkedIn. Right. And Threads and Blue sky and all the other places. So thank you so much for having an interesting conversation and giving us a deep dive into what's happening in the banking system with AI and how it's changing the world, So I really appreciate it.