Is India Behind in the AI Race? Policybazaar’s Chief Data Scientist Explains

Hi, I'm Rich Pava um editorial lead with BW business world exchange for media and martekchai.com and I am at global fintech fest here in Mumbai and I have with me uh Santosh but he's the chief data scientist at policybazar.com. So I I understand that you're a data scientist but I want to start with a an umbrella question. What do you as a somebody who you know sees data every day millions of data points uh what do you think about the Indian economy right now solely in terms of what the data suggests? If I'm looking at the data and broadly what we see in our numbers and so on, the numbers are have been consistently going up. Okay. >> And I don't think that has slowed down in terms of the whether it's motor insurance or uh you know uh health insurance or the term insurance. It's not necessarily slowed down. It's sort of you know been along the expected lines as such >> but but you know we have been talking about financial security of Indian families right Indian households and a lot of people say that families were more secure loans were lesser when you know let's say 10 years before 20 years before the economy is growing but do you think Indian households are also are becoming more secure financially I can only answer that with respect to basically the financial literacy >> I I think there's still a long way uh for India in terms of the financial literacy Right? I mean people have misconceptions about you know how financially secure they are or what do they need to become or secure their family and so on. So there's still some way to go to achieve that financial result. >> Interesting. Right. >> Uh you talked about data modeling right and using artificial intelligence a layer of putting that as a layer on data and then actually predicting customer behavior. Right. >> Now I have a question. Um comes to my mind as a very interesting recipe. Uh I was scrolling my Instagram the other day and I got an advertisement while scrolling my stories of a car which was about 20 lakh rupees and me as a target customer and after six more stories I get an advertisement of a car which was 60 lakh rupees and with me as a target customer. Right. So one of the marketers has gotten the data out. Right now we are living in an era where intelligence has you know moved leaps and bounds but data prediction models are still lagging behind. Why do you think that is the case? >> It is predominantly lagging behind because of the quality of data that is being fed to the models. and fed to the model when I say if you have the wrong set of data and you train the model on that >> then obviously the model is going to create a bias >> and that is the main reason why you see all of this happening so any data which is present with us will have to be kind of checked for veracity and the you know accuracy and so on right and the enrichment of that data becomes very very essential it has to be unbiased when you're training the model only then the model will actually start predicting something which is useful >> uh now you have also said in your previous interviews that >> good questions yes you have you have said in your previous interviews that you know AI handle more than half of policy as customer chats now right uh has this improved customer satisfaction what are the what is the data that you are seeing here because you know Indian consumer is you largely knows what they want to do right and all of us get a lot of calls from different companies right u but as a you know if you put the layer of artificial intelligence on top of these customer AI agents are talking to customers how has that changed customer satisfaction experience >> so what we've done and consciously is to actually keep the customer as the focus. You don't want to kind of, you know, uh, get the customer annoyed because he's been talking to an AI agent. The AI >> agent's role is to understand and solve whatever possible in whatever possible ways he could solve a customer's problem. >> And the role of an AI agent is to be available 24/7 >> and you know whenever the customer needs it, right? It is so you could think of the AI agent as a digital twin of the of a human adviser, right? which means mostly the easy things that you know people come there to ask and so on and so forth it's able to solve you know for example if he says I don't I want to have a soft copy it is it's able to give him that answer if he's got very reasonably simple questions around the policy the benefits etc it's able to answer it doesn't need the you know human to be there on the call at all times >> but it should also be able to hand over the call when there is a complexity and when there needs to be an interaction between two humans >> right that's so what we've seen is the the chats whether it's through voice or whether it's through uh you know text text chats So it's more than 50% it's possibly about 60% now. Um now that number has slightly gone up if you I mean compared to last year but the satisfaction pretty much has remained the same which is a which is what it is telling us is it has not impacted the customer in a wrong way >> usage of a customer would still want to speak to somebody who is more has more human behavioral like a voice or a tone or empathy right have you seen that issue >> in some customers largely what happens is let's say when I buy a policy and I want to get issuance of my policy I would want that issuance you know I want the problem to be solved M >> what we have seen in the last 3 or 4 months at least is that people who like the customers who booked the policy with us are okay talking to agents as long as but there are people for certain kind of complex problems particularly payment related issues refund cancellation any kind of complex issues that's when the uh we actually deliberately transfer it to the human >> now you have also written about the blackbox problem in AI right what do you mean by that >> uh the explanability of of AI >> okay >> right so like when I say explainability like it can give you u suggestions or it can say something which is not explainable and as to how it actually arrived at that answer. >> And so any large language model is predominantly a black box, right? You given certain inputs on certain prompts and so on, it is it's forced to answer, but it is it won't tell you how it got to that reasoning. >> Yeah. >> Right. So although they've gotten better with the with the language models and so but it's still predominantly a black box. So now how do you make that explainable? That's I think very important to build trust as well amongst if you're going to use AI in the future as well. And you talked about LLM and that takes me to another interesting question and I was talking to a CTO of a very large company L& tech Ashish Kushu. He told me something very interesting. He said whatever we are seeing in terms of LLMs today is just automation under the garb of artificial intelligence. Now why he says that is even LM work on a preset of data. Right? If I'm asking a question, it has a certain set of data which it's using to answer that question. It's not really using the reasoning intelligence that it should. Where do you think how far are we from a model where AI will actually use real intelligence and not just be trained on certain data sets? Because as as an Indian consu consumer sitting in India, it is probably showing me answers trained on data sets in the United States of America or United Kingdom. It is not really grasping real Indian intelligence or Indian context. How far are we from that? And this is an umbrella question I know but what's what's your view on? >> There are there are two parts to it. I think your question has two parts to it. one is getting I'll come to the first question where how far are we from that real intelligence and so on I think that question if it refers to the art I mean artificial general intelligence then I think we are some way away we are not anywhere close that's my view at least >> but astra has been launched >> I know but AGI but again I think the real you know intelligence is somewhere away whether it is 3 years away whether it's 5 years away we'll wait and watch very difficult to answer that question >> now coming to the Indian context and how >> sorry we're at AGI so there was an an experiment done by meta where it uh led to AGI models start conversing to each other and if I'm if I'm correct they started talking about the end of humanity because they didn't like humanity and trying to get nuclear codes out of the US president. This was four five years ago and they had to immediately uh turn the conversational models down. So to the viewers who are watching it we are still not there at as we are speaking whatever you are using is just automation under the G of artificial intelligence. It's not artificial intelligence but yes we will hear more fun from the expert. Yes. So >> uh automation under the garb of intelligence. >> Well so it depends on how you look at it. >> So the way I look at it is There are certain tasks that we do as humans which the older not older the current set of technologies like typical software development could not achieve because those are deterministic nature. The current set of you know AI models help us to do a bit of that right. So hence we find a lot of automation that is happening. There are certain tasks which with the current set of LLMs are done much better. for example, reading documents, extracting information, even extracting, you know, information from calls. Those are the kind of things it does quite well, right? As compared to basically dealing with numbers, looking at numbers and you know, making sense of numbers. LLMs are not necessarily great at doing those those kind of stuff. >> Now, this another very interesting question that comes to my mind. There might come a time when AI will probably start predicting diseases in humans, right? In terms of looking at certain features, looking at certain reports, it might be able to predict, right? How does that how will it impact the insurance sector? uh because if AI starts predicting disease patterns, if AI starts predicting what will happen to a certain person 10 years from now, how will I I I know it's a very u futuristic question but uh if you can just throw some light on it >> it's very difficult to say uh >> there are already certain thoughts if I'm I mean it's very early days I guess but there are already certain thoughts which are moving in this direction >> which are moving in this direction for example and uh AI starts reading your face and basically based on that would I actually kind of skip your medical verification for that matter, right? Um trying to make issuance faster, trying to make life easier for the customer. That's the initial direction. But it this what you're essentially saying has more implications. For example, >> if AI says that >> I I will get cancer 20 years down the line, how will it impact insurance? Those are very difficult questions to answer and it is it is possible to do it. I mean in the sense AI is already doing it >> and there was there was a use case where this lady wasn't able the doctors weren't able to find out a problem with the lady and and the LLM she went to helped it and she was saved. >> Correct. very positive you know cases there have been very bad cases as well. So I think you have to take it with a pinch of salt right now although I think there are thoughts going in that direction as to how do I ease the customer onboarding >> with the help of AI. uh what has AI made the easiest at policy bazar if you talk about 5 years ago from now on now >> there have been multiple sort of initiative that we did with the help of AI which has helped the customer for example um there is something called as a break-in vehicle insurance wherein you know your your uh policy has expired and you want to renew it now let's say it expired yesterday and you want to renew the the the regulations say that all the insurance mandated that you have to record a video of your car and send it to them >> right now what what helps is they they will I they will look at the video and they'll issue right and they look at multiple 40 50 parameters and they'll issue >> typically this would have taken two to three days even before the video if the quality of the video is good or bad or so on today I mean if the the moment a customer uploads a video I can actually tell him whether the video was correct how many parameters are correct and if everything was correct we could actually potentially get that issued in the next 5 minutes >> it's a significant uplift for the customer >> interesting yeah >> right so that that's one pay as you drive now all he has to do is basically take the take his uh you know camera in the full camera around the vehicle and certain three or four points like you photoometer, charging number etc etc the three or four points >> the issuance can be in just four five minutes >> there's nothing more that is required so imagine the customer experience it's also now the we've seen a decent amount of uplift in the customer service side which when it comes to basically issuance again it is not the journey is not done it's still early days but we do see that you know signs of that customer you know kind of >> have you created your own SLMs also >> we are in the process so we have some SLM so there basically I can't say that we have created a SLM for all of policy we are in the process of doing that. However, for certain type of use cases, we use something called as a fine-tuned model, right? So those are also SLMs, but they they do maybe two or three tasks very well. So those kind of models are already in place. Now, if I'm building SLM broadly for answering all the questions, yeah, we are in the process of doing that. >> It also comes back to your other question where can it understand the context of India, Indian insurance and so on that possibly is the >> Yeah. So the SLM answer solves that problem. >> I'll come back to that question again. any behavioral patterns, any insights that you have for the consumer data that you have at policy based [laughter] but I have given you another thing to go back to your >> I took the answer for that like the two >> the two use cases >> use cases we have to stick it together some >> interesting what next for cons >> oh there's still a long way to go I mean see ultimately if you look at it >> what are you working on immediately the next five six months >> by by and large okay the first thing that I think from a technology perspective that SLM is an important one um both from a cost perspective I mean that is cost is an important factor. The second one is actually understanding our contracts, our data and so on. That's one big thing that we are working on. Right? The the the single biggest uh area of focus is the customer service. How do I improve customer service? Again, it is not one project, it's not two projects, it's a series of projects. Like for example, I I will I'll have to have a voice assistant. Not because they're going to replace somebody human, but it is to supplement the uh the our own human advisor. So maybe you know the voice agent can actually answer a certain set of questions even the chat bots. Then you know can can the internal bots can can they help in you know basically streaming streamlining the entire operations for example talk AI can automate that process it's only going to help the customer. M >> now my final question this is the most serious question I have had Santo not been a chief data scientist >> what would he have been >> nothing else I guess >> nothing else >> nothing else I've been doing yes I've been doing this for the last 20 odd years >> no but some people say that they would have been a golfer or a singer or that way so what maybe what other line of um work not work or hobby would you have loved to >> so uh my hobbies keep changing every five years but what has stayed with me apart from my love for mathematics as such AI is also is mathematics um is actually wildlife environment and wildlife. So I do photography, I do some bit of conservation and here and there. >> Tell me also tell me something now. Um India is a country, right? We were not late to the AI race, right? We might have been led to various other uh you know races globally but everybody had the same starting point when it came to the AI race. we still somehow lagged behind, right? We haven't really produced an open eye of ours or a a deepseek of ours, right? Yes, we have GBD again which is not B2C, right? It's more of a B2B platform. Where have where have we lagged? Is it the computational possibilities? Is the data possibility? Uh is is that the lack of grit that people that founders in India have? Where have we lacked in terms of actually giving the world an MLM? So I think the you you kind of answered it yourself in the sense that we've lacked in the foundational part of the AI race which means we don't have any foundational model which has been built in India nor are we kind of controlling the compute right possibly the the data centers and so on yeah that may be still there um it's still okay I mean so think about it right the hardware is not with us even the software is not with us at least as of now but it's I I think we're still early in the in the race in the sense while we may have been lagging behind there we we might be able to catch might be able to catch up. >> I think on that note we end this interview. Thank you so much for giving your time. I think it's some very interesting insights. >> There's a lot of interesting things that India could go ahead for example no other country has like 25 30 languages. >> Yes. No other country has 1.4 billion people also. India sitting on just all kind of data. >> All kind of data and there could be many innovation that can come even with the even if you don't own the you know >> and I think that makes your job even exciting as a data scientist. I mean I'd love to you know and as a somebody who loves mathematics you would love to sit on numbers sit on data every day. Interesting. It's interesting to see people who love their job. Uh to all those who are watching this interview, I hope you gained something from it. Uh Santosh was very very good and insightful with his answers. Thank you so much for giving us nice