Revolutionizing Customer Experience with AI Agents | Leading with Data Ep 53

Speaker A: Transition between a lot of business leaders when they move from gut based decision making to data based decision making. So mid 2000, I think a lot of companies evolved from gut based to database decision making using mathematical models to build a lot of scraping, uh, algorithms or scraping scripts. We kind of took our experts to build the first few scripts using copilot and then um, uh, the copilot learns along with you and it basically replicates your coding style a lot of what you do and obviously you get some help as you write the scripts as well. So after building the initial base scripts then we used to give it to a larger set of people so they can just, they have almost a ready script and have to make minimal changes. Customer feedback was based on threats saying that I will sue you or things like that. You can classify that topic and ensure that that order or that customer care department goes on hyper alert. And it's a very different kind of sentiment. This is to prevent mhm. These kind of orders then further going into litigation. Adoption is key. I think early adoption is key. You cannot say that okay, AI will either take away my job so I'll not code and things like that. So you have to learn the nuances of AI can make your code better. That is one thing. Even if they have 500 agents and tomorrow if I tell the person who's bought Salesforce that I can do your customer onboarding better, just that agent, I can plug in that agent where I have more details, I have more logic, I have better uh, algorithms and so on. So customer journey is a workflow. So if they've identified the customer journey, you probably need to define what journeys where agents will make a bigger impact. Campaign or campaign effectiveness or ensuring that campaign effectiveness is tracked. Maybe personalized campaign effectiveness is tracked and just that the agent turns out or churns out newer campaigns, maybe it's a lot more effective there. The other area that I see a lot of focus will be in um, usage of uh, AI generated audience, synthetic audience kind of thing. So that will be. That is uh, that entire market research industry is primed for disruption foreign.

Speaker B: Hi and welcome to this exciting episode of Leading with Data. Today I have with me Naveen Dhananjay who is currently the global CXM lead for Dentsu Global Services and he brings an immense experience of working with Fortune 100 companies and building their AI and Gen AI products. Naveen has also been an entrepreneur and he started one of the first companies in India in analytics space. And over time he has gained uh, immense experience and today that is what I'm looking to learn from him. Welcome to the show Naveen.

Speaker A: Thank you Kunal. It's a good opportunity to speak to you.

Speaker B: Sure. So Naveen, let's go back to how you got into analytics and data science and now AI and ML. So what has been your journey from your perspective?

Speaker A: Yeah, like you said, I think um, I happen to be part of an early wave of analytics as you would say. When I started Manthan, but much prior to that I used to work with enterprise business software companies like SAP where I worked a lot with the data, data modeling, data warehousing. And that's basically the foundation of where my analytics journey started.

Speaker B: Interesting.

Speaker A: And um, and we never called it analytics then but now everything that you do with reporting and data warehousing is also called analytics. So a lot of work in data started long before um I actually co founded the company. But since then uh, it's been a continuous um, journey with data and database technologies and then it's evolved into analytics, evolved into AI, evolving into agentic AI. So journey continues.

Speaker B: And in this journey what have been, let's say a few aha moments for you or what have been let's say significant milestones in the journey?

Speaker A: I think a few interesting things were the first time um, when I was certified as a data warehouse consultant and this is as long back as 1999, just before the Y2K thing on the SAP software in terms of um, how you could manage models or how you could manage data warehouse or how you could work um, with the data systems for decisioning.

Speaker B: And this would be cube based systems.

Speaker A: This is a cube based system. And so you normalize data, you bring data models and then you normalize tables, you create fact uh and dimension tables. There's learnings that I can't forget because of how it was taught to us then. Learnings have become very different now. But uh, so that fact and dimension tables are further um has been ingrained in my mind even now because in a lot of my interview questions I do ask model management at a basic level, at a data based management level I ask concepts of normalization and things like that. But that said that's one aha moment. There have been many aha um moments throughout the journey but um, other than that I think a couple of others I would like to highlight is um, when I was working um, um with pure play analytics companies transition between a lot of business leaders when they moved from gut based decision making to data based decision making. So mid 2000s I think a lot of Companies evolved from gut based to database decision making using mathematical models. I think that was another aha moment. We completed a lot of projects then. Then the other thing is uh, for in our journey with M, you know when we started um Manthan itself, basically um, how we used learnings from there to, to create insights for our clients in a particular vertical industry. In that case it was retail. That was another aha moment. And more importantly both if you remember um, the Amazon ad from about six, seven years back, it was called or the Cow where they used to talk about products one after the other. That count was done by Merkle uh at that point of time. And basically that led to a lot of other projects where we would do attributions for the products that we counted then and then classification of the products and then auto attributing that using AI. I think that was the genesis of where I transitioned from analytics to AI. Wow, that was back in 2016, 17 and the reaction from the audience, both our clients and these are clients like you know Amazon and others at that point of time, um gave that impetus to say okay this has legs, um obviously from transition from there to how analytics can scale and was for me or the teams that I built was relatively um, I would say straightforward because of the engineering background. But as an industry AI and analytics uh scaling required a lot of technological expertise. And going forward I see that barrier also going down with the kind of agent force technologies that is there.

Speaker B: Uh just want to double click on the app. Amazing example. So can you elaborate for example what was the initial requirement and then you know uh, how you went about building that solution and finally the results and then the impact it created.

Speaker A: I would take uh a generalize it uh to any um retailer I would say so this is more like um, you know when a retailer onboards products uh on its sites it requires attributes.

Speaker B: Yeah.

Speaker A: So typical set of attributes are anywhere from three to four when they get from a catalog, if it is a branded entity or even if they have more, it's about 10. For a product to be searchable. For a product to be discoverable on Internet using all the search engines it requires an average of 30 or 40 attributes.

Speaker B: Got it.

Speaker A: So you can um. And it requires appropriate classification. So you can take an example of a sports uh shoe. So sports shoe can be in fashion, sports shoe can be in um under shoes. Sports shoe can be for uh fitness and well being. So there are different classifications of the same sport shoe. And if the trend is of fitness or trend is of some celebrity endorsing fitness, then the uh, sport shoes under fitness will sell more. And if your particular shoe is not under classified under that, the selling will not happen as long as if a competitor has had the same shoe in fitness, it will sell more of course. So based on a lot of outside in inputs like these, you would classify the shoe appropriately.

Speaker B: That's probably the reason why search at let's say site A or Amazon feels very different from the search on let's say competitor. Right.

Speaker A: And you have to continuously do this because trends change and trends have been changing significantly faster over the last several years. Right. So this was, this is one example how we can do it. And then the second one is assume that you've identified the problem. So we had our own methodologies in terms of based on the traffic to a site, uh, how much traffic comes to a particular site, uh, what kind of uh, products are more sellable and what kind of products are going away because your price elasticity is less. So you would give recommendations on positioning a product just about the product being there. The second is about if the product is being there and is searched, are you price competitive enough based on the competitor? And for both of that to happen, are ah, you attributing it enough? So from um, about 10 attributes, if you want to go to 30 attributes in a short span of time, one is to get a team of hundred people. I'm just taking a random example for like a couple thousand products. Otherwise you do it using some automation, using some um, um, models and using AI to augment that.

Speaker B: Interesting. And uh, these attributes, uh, so once you extract it from let's say either AI or human generated, then they go into the systems and then how does that feedback loop come in? So for example, whatever changes you did, its impact and how do you see whether that attribute was the right one to fit in?

Speaker A: That's a good question. And these things, this typically what the clients would ask you as well. Traditionally at least over the last um, 10 years, whatever systems we've built, we've built what is called as augmented system. When I say augmented systems, either they are augmented using humans or they are um, augmented using further systems that we will build. So if uh, I needed to impute say 10 attributes and eight got filled, first of all we need to check whether the eight are correct or not. And then you implement, impute the two remaining through human interference and um, um, intervention and then um, you kind of feed that back into the system. So the next time the model looks to impute, it'll ensure that at least those 10 are filled and then you worry about accuracy and so on and so forth.

Speaker B: Interesting.

Speaker A: So it evolves, always built uh, a learning system which is right now, I mean like, I mean you call it reinforced learning nowadays or you call it uh, augmented learning, even system. So that I think combination will continue. And the other interesting thing is unlike a pure product based company, we're always, at least I was always a part of product plus services companies that allowed us to have teams that complemented whatever engineering platforms we built. So that was very important.

Speaker B: Interesting. And in this journey, uh, specifically over the last, let's say couple of years. Right, so Genai has come to foray uh what was your aha moment with generative AI technologies and was it you know, with ChatGPT, was it pre ChatGPT or was there any particular moment which you remember which made you really think about the possibilities and the impact it can create?

Speaker A: So we um, had this, we used to uh, when we initially built our cognitive system we called it jarvis. So there was this moment when we built this module that said Jarvis can write, which basically wrote content for a lot of um, E commerce products.

Speaker B: And this was pre chatgpt.

Speaker A: This was pre chatgpt. So that is a pre Chat GPT world. And then um, again it was a combination of what um, uh, the algorithm would do and then obviously you would correct it. But the fact that you could come so close to writing it had uh, a lot of promise. And of course um, ChatGPT can do significantly better than what we did that time. But then that's where that evolution enabled at least personally for me to understand what the power of AI could do. And obviously with uh, large language models, the other thing that has happened in a lot of these algorithms that used to be, I mean as they released beta versions and all of that, we also took learnings from a lot of these models, say image classification. The models were released earlier. We used to take some of these models and um, you know, uh, augment it with R learning or train it with our set of images so that it is more contextualized to our problem statement. Fine tuning. And then again you call it, you know, um, augmented learning of the models as well. So basically we used to take out a few layers of that model and train it with our layers so that it can classify our kind of models better.

Speaker B: These were deep learning based transfer learning was the term which was used.

Speaker A: So a lot of transfer learning is what we did.

Speaker B: Yeah.

Speaker A: So at least for the models that we implemented, why I'm saying that we augmented that or we ensured uh, we did transfer learning is the context of our problem was then far superior to what you would get in any commercial AI system that existed and that proved to be the differentiator wherever we.

Speaker B: So you would add a few more

Speaker A: layers of remove it probably remove it,

Speaker B: then add uh, interesting. Uh, so this was on the deep learning models and then what was let's say the initial m. Few uh gen AI based applications and how did your let's say innovation and then just changing the customer experience evolve from that point to where we are today.

Speaker A: So I'll take uh 3 examples uh on that front um, especially one is uh, maybe an internal um, ah code based problem, uh, like a customer feedback problem. And the third one I'll use the um, example of um, how we do it uh, real time for marketing. So the first one was in a typical coding exercise. I mean you, all of you have heard of Copilot. So Copilot has this capability uh, of I uh mean it's a generative AI based um, uh toolkit coding system. But um, we used to scrape a lot of data to build a lot of scraping uh algorithms or scraping scripts. We kind of took our experts to build the first few scripts using Copilot. And then um, uh the copilot learns along with you and it basically replicates your coding style a lot of what you do and obviously you get some help as you write the scripts as well. So after building the initial base scripts then we used to give it to a larger set of people so they can just, they have almost a ready script and have to make minimal changes.

Speaker B: So that becomes a template for the

Speaker A: larger template and as the larger team develops that they in turn make it more efficient. So what typically would have taken say as an example a team of um, 100 people. We managed to do it in like a fifth of the uh team. So using about 15 to 20 people were able to generate this. And this was a fairly large project that required at least we were on this project for almost two years and this was for one of the Fortune and clients and we were able to significantly add value both internally and to

Speaker B: what the client wanted and to the business. It would mean for example let's say say if we have a holiday season coming up, something which used to take 100 people working overnight during the season could be done in 20% of the

Speaker A: time, very little time. Yeah and it's the efficiency here clearly. The other one is um, say topic uh modeling in customer feedback. So there is a couple of Things in a lot of customer feedback systems. And I'm going to generalize a little bit here rather than going to specifics of the example, but you will get a gist of it is that um, if customer M feedback was based on threats saying that I will sue you or things like that, you can classify that topic and ensure that that order or that customer care department goes on hyper alert. And it's a very different kind of sentiment. But this is to prevent, HM these kind of orders then further going into litigation. And we have tangible proof that we've prevented millions of dollars of litigation based on something like that.

Speaker B: This case you were using uh, essentially a large language model to understand the feedback in a more nuanced way and essentially then taking action. So there was no generation as such.

Speaker A: But it was not a large language, almost like a small language model of sorts. But we were using it in the context of what data we had. So we had collected a lot of data in the past. So that is where uh, it was a third case is. I mean again it falls in between both these elements of what Gen AI can do and what uh, large language models can do is um, we are all used to seeing a lot of these OTT serials and things like that. So for one of the media houses we even looked at um, TV serials and based on the theme or the sentiment of the serial we could then uh, pitch in the ads. Ah. And uh, the same ads could be different for different audiences watching it. And once the audiences were again identified based on what their preferences were. So it is like a three way thing. So first of all you are talking to a client, then you have your audiences and you are adding attributes to both audiences and what is happening real time in terms of themes and mixing both to generate the relevant or generate the relevant uh, content ad as it may be.

Speaker B: So uh, just to put it in business context, while let's say IPL match is going on, a person sitting in Chennai would see an ad for that team and probably a different dish if it is let's say Zomato Swiggy versus let's say someone sitting in Delhi and then all of this happening in real time.

Speaker A: In real time. Yeah.

Speaker B: Interesting, interesting. And uh, uh in terms of impact was there, I mean for something like

Speaker A: this here we have classified, they have a tangible impact as well in terms of increased impressions, increased viewership, all of that. So there are tangible metrics based on which we further augmented a lot of these attributes that I'm talking about. So this didn't happen day One, all the three things that I happened in three different stages. But uh, it's still going on. And uh, there are tangible metrics that we are doing.

Speaker B: And today when you step back a bit, right? And in terms of just reflecting on the amount of changes which have happened in the domain, uh, how do you see the life of a practitioner and a leader changing? And how should their, let's say, thinking framework adopt with the kind of changes happening in the background?

Speaker A: I think, at least in the AI world, I think you can learn a lot from AI itself. I would use a lot of generative AI to learn to be my teacher of sorts. Uh, I have a friend of mine who um, started something on his own. He kind of uses uh, uses uh, Genai in two things. He uses it for a lot of his legal contracts. He uses Genai himself and he's senior to me and he's using it at this stage. I thought it was pretty cool. I mean he's kind of optimized the prompt so much to use it in that fashion.

Speaker B: In that fashion. Not an idea where you would not want any hallucinations and absolutely no.

Speaker A: I've done pilots, in this case even for our company. But this guy using it real time is amazing. I was starting like, then I'm behind. So you can, some of these guys, when they use it, uh, use it very differently. So that's uh, you know, one thing, um, maybe the AI itself will teach you a lot more. The second is that adoption is key. I think early adoption is key. You cannot say that, okay, AI will either take away my job, so I'll not code and things like that. So you have to learn the nuances of AI can make your code better. That is one thing. The other thing is, is just the journey that you are in. So you will just use a lot of these systems to make things better. So I think um, the learning mindset will be key for you to adopt and then add value in whatever form you can.

Speaker B: So going, uh, this is almost what we were discussing, uh, before we started recording, having a childlike mindset and how uh, Gen Alpha can actually adopt to new interfaces. Uh, very interesting, uh, looking a bit forward and projecting into future. Right, uh, AI agents are going, uh, the way, uh, they are developing and the way industry is forecasting. I think 2025 is now being said that this would be the year of uh, AI agents. Uh, so can you share your perspective on uh, what agents can enable, what kind of applications can they do and what sort of projects or problems can we solve? With these agents which uh, would have been let's say theoretical constructs a few years back.

Speaker A: I'll take simple example because um, some of this and everybody and anybody is talking about AI agents. Right. And I do believe that agents will make a difference. That's one. A lot of enterprise software companies are also reimagining how their functioning will work because of uh, agents coming to the phone for like large companies like Salesforce itself are promising over uh, over I think 500 plus agents using the agent force and things like that. But what agents can do like in a simple form if you are opening, if you are the front desk person in a bank where you go to open an account from simple things like that to when they ask you your name and details and um, a little bit of whatever documentation uh is needed and get an account to be open is a simple task that we would have automated in the past. Many companies would have done that but in a conversational manner. Any agent can do it so making it more personal and also more context. More context. So the same uh, agent or bot can talk to you slightly differently, talk to me slightly differently. But in the end of the day it is seeking just for inputs and closing the thing. Now once you open an account it can hand it over to the investment guy to make it where the investment guy bakes on basis on your portfolio versus my portfolio. Suggest different things is another reasoning. Agent is another agent that can do uh this work.

Speaker B: Mhm.

Speaker A: And that is another important thing of how handovers can happen. Now these two workflows, if it can be integrated by a third agent that manages the workflow is where the differentiation starts coming. And more importantly it might even go and tell you that okay, suggest investments for Kunal up to say 1cr until he and he reaches or uh a breakthrough of like 12 or 13% you can give it a goal. So then it will work on your portfolio and continuously work till you get a 12 or 13% return to that extent. Now that, now how real is that? I don't know whether because you can always work with your uh, broker or somebody like that to get to that 12 or 13% thing. But these three steps done by an agent is real actually. So that's something that uh, people are building and uh, to what extent it will succeed and how effective it will be is for time to tell. But now if you can imagine this there is a whole number of um use um cases or enterprise software that will change. And I, I firmly believe that uh Salesforce is doing the right thing. Because the entire CRM uh package in itself I think is bound to be disrupted. So now imagine, I mean if they don't have agents and even if they have find it agents and tomorrow if I tell the person who's bought Salesforce that I can do your customer onboarding better, just that agent I can plug in that agent where I have more details, I have more logic, I have better algorithms and so on so forth.

Speaker B: And it's running 247 objectively.

Speaker A: So there is no doubt that uh, there are certain areas or certain software that will see a complete disruption on all this. You can even like do very uh, different kind of objectives. Uh like saying that okay, um, if you're going to can say that replicate uh, replicate Bill Gates investment portfolio for me. Okay now how it does that is all of these guys file you know different kinds of returns. So they take from that returns in terms of what they, what their investments are and it can replicate that as a matter. And then there are advisors sitting in between M and advisors can always have what you want. Uh, um, they are, they will advise you. But if you write that logic, some basic logic as to and people are have written algorithmic trading packages and things like that. All of them can fit into reasoning models within an AI agent. It builds on and on. I mean there's no end to where it can.

Speaker B: And so let's say for a enterprise which is, which has been reading and which has been you know uh, seeing all the action, uh they want to probably do a pilot or start with uh using this technology. What uh. How would you say that they should think about this and what should some be some of the let's say best practices which they should think about before starting this journey.

Speaker A: See in a lot of uh these cases there are open source um frameworks available. They basically do the same, they use the same um layers below in terms of large language models or LangChain and stuff like that. So I would say that at least from an enterprise software I think you need to get um, your liabilities legal covered first because agents can run riot. So that having done that I think if you are basic um, uh if it is um a cleaner enterprise version I would go for. If it is, even if it is a pilot I would go more for enterprise software, enterprise versions which is a lot more responsible. Which is inside your firewall of sorts or which is in a dedicated cloud mhm. Where you can experiment with it a little more uh clearly or a little more objectively. And um, within that um experiments I think the Timeline for this getting an agent up and running is very quick. It's not going to be that far. So in about two to four weeks time you can definitely get one agent working for you to scale. I think um, I would step back and see what workflows. A lot of companies have journeys identified. We talked a lot about customer journey. So customer journey is a workflow. So if they've identified the customer journey you probably need to define what journeys where agents will make a bigger impact. Say campaign or campaign effectiveness or ensuring that a campaign effectiveness is tracked, maybe personalized campaign effectiveness is tracked and just that the agent turns out or churns out newer campaigns features in the generative way. Maybe it's a lot more effective there, things like that. So I think I would, at an enterprise scale I would probably study the workflow, study the foundational layer and replicate that in the agents that will be more effective to start with before jumping on to multiple agents. Because the real value is in having a uh, series of agents work with each other. And more importantly I would never take human out of the loop of course so agents will help me, help me make it more effective. But it can humans in the loop will add value to.

Speaker B: So two follow up questions, right? One, when a system like this is working where there are multiple agents, each of them having their own, let's say objective how do you optimize and how do you monitor these? Because any of these agents can go wrong or hallucinate or uh, can you know, even if it is a slight destruction in the system, it can actually uh, uh accelerate. So that's the first question. Uh, and then maybe we'll just go through that.

Speaker A: So this for even true for AI and AI models as well. So I think even when I spoke last time one of the things I am recommending is that you um, measurement of AI implementation FAA is one thing but measurement and you know what would be called quality of AI checks is important and you want to measure the metrics uh first and there were different measuring methods that have been told and it is like right from. I mean you need to adopt one and ensure that you have the right measurement for AI. Likewise for agents. I think um, I think it's evolving still but uh, to get the right framework of how you will measure the effectiveness of an AI agent will be very crucial for the enterprise.

Speaker B: And uh, what's a good way to keep the human in the loop? I know there is a broad spectrum but uh, you know let's say going back to the campaign example right now These agents are throwing out campaigns which are, let's say sending thousands or billions of communications. Do you do sample testing? Do you see that? You know, here are the guardrails and anything outside of these get checked. So what's a good way to think about keeping the human in the loop in problems like this?

Speaker A: I think uh, again it depends. Obviously you don't want an agent to just set an objective function and kind of let loose. I think somewhere you'll have to um, ah, do a dipstick, especially with agent influence and if the effectiveness is not there, maybe pause and take a look at it. I think that's the change management that the industry will go through. There are, there are published best practices but unless I think a lot of this gets into production, this area of research will probably emerge or best practices will emerge. We do have a few in the sense that uh, uh, as we go with um, um, multiple agents in a workflow, what you can do in a workflow and plus how elements of reasoning will work in a workflow. And mostly elements of reasoning are limited to small language models or you know, even smaller micro language models for a particular decision making. I think it will make it more effective as it may not go as rogue as it could possibly do.

Speaker B: So still uh, essentially experiment a lot before scaling up, make sure that the uh, uh, guardrails and these workflows are in place and then kind of uh, interesting. And uh, for someone who is just entering you know their career and there is so much flux happening in technology and you know, for example a few years back I would have said learn Python, learn machine learning algorithms, take a project and work. Uh, that template almost feels like it doesn't fit. Well now uh, if let's say there is someone who is coming out of college and uh, uh, wanting to be part of this action, what's your recommendation to people like those and how would you uh, advise them to think about their.

Speaker A: You still need to get your fundamentals there because knowing Python is still necessary if you want to kind of augment the Python code that an agent is writing. So you're not going far away from that. But I would say augment your learning with a lot of what the AI world is providing. I would just uh, and I would say that um, you know, if you want to be multidisciplinary, one is about recording, uh, I mean learning about what agents will do or uh, what you can do with agent. But the second thing is how you can help in the consumption which will mean that articulation of what an agent will do both. From a technology and business perspective, both are important. And then there is a foundational layer which is at a data layer or a cloud layer. I think that's an important technology as well. You can configure the technology using an agent or using AI, but you need to know the fundamentals of how the technology works. Otherwise you will um, end up uh, you know, uh, consuming a lot more cloud dollars without knowing how to optimize the cloud.

Speaker B: So essentially still focus on fundamentals. Make sure uh, you know, you understand the coding language, the math at least at a decent level and then the infra things and then kind of I

Speaker A: think everything is linked with. I mean nowadays the compute has become so easy to get. I think some of these are very fundamental to how you will operationalize even in AI infrastructure.

Speaker B: Interesting. And just taking a uh, you know, very uh, wild guess. So let's say if you had to share a few examples which you would love to see getting solved with agents or any projects where you think this could be a huge impact in next two, three years or whatever time frame you want to choose, what would be some of those?

Speaker A: One is campaigns for sure. Uh, I think it's a big area of impact and I think I would love to kind of see the entire campaign journey being uh, serviced by multiple agents. Whether it is running of the campaigns, whether it is generating of the campaigns, which is a lot of visual design um, and creative measuring um, of the campaign. Uh, so all of that is one area. The other area that I see a lot of um, focus will be in um, usage of AI generated audience, synthetic audience kind of thing. So that will be, that is uh, that entire market research industry is primed for disruption and um, usage of syndicated research, usage of fast research using syndicated uh, synthetic data or AI generated data I think is uh, is going to create a significant disruption in that industry. So at least these two examples is

Speaker B: what I would say and uh, uh, just double clicking on that. Right. So we have gone from a world where you know, you could say if there is a photograph or image, this has happened to a place where you no longer can trust uh, uh, if you're seeing image M or now even video, uh, do you see the same challenge coming up with data specifically with as let's say the use and tools of synthetic data become more prolific and uh, uh. So, so how do you see this space evolving? Synthetic data generation, usage and then again responsibility in terms of how do you use those.

Speaker A: So I'll uh, separate out synthetic data and synthetic Audience, Sure. So synthetic data and how you would impute synthetic data will only add to what has been done so far. And I'm sure during your times as a data analyst and over the years we've used methods to impute data. So this will only help is what I feel Now Synthetic audience is uh, an altogether different uh world. It is evolving, uh, it is not yet being trusted. But I think uh, it is a means where faster research can be done on one thing. Um, there is a lot of test and learn that can be done in a much faster iterative manner. Annette Science scale using synthesis. So there is a lot of intermediate companies or marketplaces that provide real audiences today they have to pivot to this world otherwise they'll become irrelevant very soon.

Speaker B: Right. And then obviously the cost structures and the possibilities become immense and uh, it may be linked. So you know for example when Google was building DeepMind, uh after a lot of uh, iterations it started coming up with its own strategy which people had not seen. The famous move against Lee Sedol, uh, that that could also happen with synthetic data. Right. So researchers, so things which are unlocked or not there in ah, human knowledge as of now are not as prevalent, could start coming out very, very powerful uh, in terms of uh, your own team and evolving these people. So at Merkle and then sue, how do you kind of uh, nurture talent and on an ongoing basis upscale and be at the abreast of what's happening in the industry.

Speaker A: So couple of things I think from uh being abreast in the industry we do have experts that have built their expertise over many many years and they've gone through many generational, you know, uh, uh, analytics as I would say, whether it is data warehousing or mathematical models or uh, machine learning models. And now AI, um, we do participate in a lot of forums, um India and in the US where we get those inputs. Ah, we have um, both external and internal training curriculum in those areas and we create um, we gamify a lot of that. We have like promptathons for people to learn a lot about AI. Uh, we do internships where we kind of um, at least the topics on the internships are a lot more to do with future tech related areas. So I think the teams or the mentors in some of those projects learn a lot more in those areas. And of course there is a broader vision we put together like every quarter like this last few quarters we have what is called as everybody needs to have one AI initiative whether it is internal related internal optimization or customer Facing or like it is disruptive. So we've classified into three or four areas and they need to classify that and we will probably uh monitor it and in some cases even fund it. So and this is broad based so we've kind of um, kind of uh given this out to almost uh 2,3 skip levels which covers 200 trillion people and that basically represents broad 5,000 people

Speaker B: audience and you uh we discussed about the promptathon um as well. Can you share what.

Speaker A: I think we'll uh try and uh see if we can get a demo out for you. But we've done this at ah scale even for our leadership team where we have kind of tried to solve problems using uh, it's almost like solving puzzles or riddles using some of those uh uh prompts. It's an interesting way of learning. We have got prompt battles where we have different teams fighting it out. Uh we've got um, we've continuously um you know evolved that curriculum so that a larger audience participates in that m. We've got uh an AI month that was there where through the month the different team showcased a bunch of things that they've done using AI or potentially want to do using AI and uh, so a lot of these helps bring about awareness. And then of course we've invested or we are at least you know every year we are planning to invest in AI training so we budget dollars accordingly.

Speaker B: Interesting. So Naveen, double clicking on that uh career advice piece uh and you know there are people who may or may not come from coding background or data and uh analytics background. So what would be your recommendation to let's say people from these different backgrounds and how should they think about it uh in terms of upskilling and their learning path of sorts. If

Speaker A: this is something that um we had highly encouraged uh within the kind of people we hired we always encourage diversity in hiring from this perspective also. So we've hired um like uh one of um to give you one very interesting example, one of our best person uh who does attribution was actually hired from Coimbatore Airport because while buying speech we realized how well she spoke and how well she attributed the product there. And she's been in the ecosystem for many years now I think so. But we use this example to see how people like them, people with a break and who have come so when they've taken a break and at least um a lot of uh, when we welcomed um ah a lot of people with bricks the technology that they worked in the past was very different than the technology that was there right now.

Speaker B: Yeah.

Speaker A: But they had the uh, background of building systems, testing it very thoroughly during their times at that level. Nowadays we don't test it as thoroughly. We use that trends to see how they can contribute both at a technology level or at a mental level. And some of those have been very effective. Then there are a whole lot of people who have done the education in computer science or IT but have then gone on to do other things. Um, like one of the folks we recruited was ah, a badminton coach who was done his bachelor's uh, in computers and then you know he had uh, um we had hired him and he did extremely well because of his nature of coaching many, many seniors uh, in the industry, uh, in the game actually he did very well in terms of ensuring that he articulated the report and the insights from the report very well to the client of course.

Speaker B: Yeah.

Speaker A: And then obviously he's put in hard work, he's kind of moved on from. It was a good stepping stone for him. So what I'm saying is like there are many paths to that success. You just need to have a foot in, in the door either using technology or one of the competency that you have learned. And I would say that the kind of competency that is changing and we talked a whole lot on Genai or GPT. I think there is, those are very good instructional mediums for you to pick up on any new technology today. M. I would just recommend that you like work hard on that and then go deeper into whatever areas you want to do basically whether it is just in the, in areas of analytics, whether it is just a reporting kind of uh, area that you want to be good with or if you want to be good with um, data or databases or cloud. I think right from what you can do on a um, basic cloud infrastructure like um, AWS or Google or Azure plus what you can do over and about that using something like Snowflake, um, or how you could become better. So that learning curve, as long as people can follow, I think it will be an enriching journey for them because all of these areas that I'm talking about, they pay very well in the market. They're still short of a lot of talent in some of these areas.

Speaker B: Interesting. So essentially reflect on your strengths and identify the area where you would want to go deeper and establish yourself in that field. Great, thanks. Thanks. Uh, Navi, towards the end of our episodes we do a rapid fire question with our guests and the idea is to just know you better. So uh, any particular book which had A huge influence on you and which you would recommend to people.

Speaker A: I think. Um, um, there are a couple of books I'm um, writing the writing I'm reading. I just read a book recently called um, how to jump the S Curves.

Speaker B: Okay.

Speaker A: It was written by X Accenture guys. In terms of how you like. You know, when we talked about this um, uh, entire concept of uh, analytics and then machine learning and then AI and, and then agent. So each of them have some S curve. So basically as one ends the previous one has already started.

Speaker B: Right.

Speaker A: So how do you identify those trends and stay at top of that S curve? So if you plot the S curve based on time and impact, you kind of realize that. So something similar to that is what it's. If you plot it in retrospective also you realize how impactful it could be. So that I think is a good framework for me to leverage. So that's one of the recent reach that I would say.

Speaker B: And then does it leave you with some frameworks? For example, you know, typically it's difficult to identify that curve at the start. Uh so in analytics and data science was one term and then machine learning started picking up. So does that give that framework?

Speaker A: I think more than identifying thing it tells you that if you're on an SQL, how do you um, build on that and stay on top of the scope? I think the key to be it is to stay on top of an Escobar curve.

Speaker B: Interesting. It's called navigating the S curve.

Speaker A: Yeah. Jumping the S curves.

Speaker B: Interesting.

Speaker A: Navigating and jumping. I think I forget what. I'll give you the name.

Speaker B: Okay. And uh, any particular tool or app with AI based app which you have started um, using lot more in the last few years?

Speaker A: I think I use uh, the AI based search a lot more because I kind of superplexity. Yeah. So I think I use that a lot more. I mean I do searches based on some images or um, some hobbies that I pursue either like um, I typically use it. I mean I was trying to help um, some of the conservationists with um, forest mapping. So for that trying to use something like that.

Speaker B: Okay.

Speaker A: Um, so that is, that is one area that I've specifically used AI a lot more on the image side of things.

Speaker B: Wow. And you mentioned that you hire a lot of interns and ask them to almost pursue what they feel excited about. So any application which uh, you know really stood out for you and then was very different than what you would usually see.

Speaker A: I mentioned earlier in the thing uh, called um, you know, when we trained our cognitive computing system to write. That was one such case and thing. In fact, early days of AI and image, we also had, um, uh, you know, uh, dressing up of this models using AI so they could, like when the catalog came out, instead of real models, the models and how AI could make them wear a particular dress.

Speaker B: Right.

Speaker A: So those two were pretty much interesting stuff that, uh, some of our, uh, folks worked on.

Speaker B: Great, great. Thanks a lot, Naveen, for sharing your insights, your perspective, and, uh, you know, thanks for setting this up and taking time out for it. Thank you.

Speaker A: Thank you. Thank you, Kunal. It was, uh, really enjoyable.

Speaker B: Thank you. Thanks a lot.