GenAI vs Agentic AI: Key Skills Powering the Future of Work

Uh, we will start, uh, we'll start, uh, today's session shortly. Uh, till the participants join in, could you quickly confirm if I'm audible? You can type a yes in the chat box. All right, perfect. Thank you so much for confirming. We'll begin right now. Good evening, everyone, and welcome to today's insightful and hands-on webinar. We're excited to invite you to a high-energy and revealing session on the topic GenAI versus Agentic AI, key skills powering the future of work. And to guide us through this essential topic, we're honored to welcome our expert speaker, Rushikesh Maharware. Rushikesh is the founder of Vidwata and currently serves as vice president of data science at Motilal Oswal. With over 13-plus years of experience spanning data science, performance test testing, application monitoring, NLP, and, uh, corporate training, he has worked with leading organizations such as Capgemini and Renault-Nissan. A passionate educator and mentor for top edtech platforms, Rushikesh is widely recognized for his expertise in machine learning, natural language processing, Python, and AI problem-solving. To help us out with the webinar today, we have Shilpa Das. Shilpa is a career growth specialist at Great Learning. She's a passionate leader with 8-plus years of experience in career counseling, academic advising, and career enhancement. Her expertise lies in driving strategic initiatives and guiding data science and AI into this yes in career growth. So, without further ado, let's get started. Over to you, Rushikesh and Shilpa. Hey, thanks a lot, Shuchi. Let me just, uh, start. Great, great. So, yeah, hi, everyone. This is Rushikesh here. I hope this to before we get started, quick check, are you guys able to hear me and can you see the screen? Can you just quickly put your feedback over the chat. It will be very helpful. So, can you guys hear me? Awesome. Great. Great, guys. Thanks a lot for your feedback. Uh I'll get started. Shilpa, you want to start or uh Please take the lead. Yeah, sure. Thank you very much. — Yeah. Thank you. Okay, great. So, guys, welcome to the session on a very interesting topic, that is Gen AI and the Agentic AI. Now, the whole current world is running behind the AI. There's so much of things are happening. Every day you keep seeing there's a new model is coming. This company is doing with the AI. something or the other model. Everywhere there are just AI and AI going on. And on top of it, the underlying layer, which is powering all of this technology, is this Gen AI and the Agentic AI. Okay? So, we'll discuss around it, like what is it, what is the importance of it, what kind of use cases, why it is needed, how do you learn it. So, all those things we are going to cover. Before we get started, let me just give a very brief again about me. The team has already discussed about myself, but my name is Rushikesh Mewada, and I have around 14 years of experience in the overall IT industry. Initial career of mine has been more of in the performance testing kind of So, more of I was in the testing, uh performance, production monitoring, and all those things. We used to look into a lot of data also to monitor the production systems. And around '15, '16, '17, I was more looking into, okay, how do I transition myself into something different from the testing. Testing was still good. I was liking the field. lot of automation capabilities were there, but then this data science as the sexiest job of the 21st century, that was something which was running behind me. The way you have AI going on everywhere right now. So, the last decade was more around the data science, machine learning, and all those things. And getting into that field was slightly tricky because the math and all those things were required. And I used to keep doing some online courses and always trying to transition, but every year I was trying to postpone. Okay, I'll do this course and that course, I'll do this course. So, then I was looking again for some more mental based to courses. And around '17 '18, Great Learning had some very good course at that point, which was business intelligence and business analytics. See, at that point, the overall landscape was very different. So, we had R as a more priority language, machine learning was more prioritized. Deep learning and all those were not there. It was more of a classroom training for 1 year. That just gave me a very good mental led learning part. So, sometimes

it helps you to transition much faster because you have more structured learning and all those things. So, it helps you to transition faster. So, again, after that also it took some time for me to get a job, but then I was able to transition into some AI roles. I led I was in the leadership roles. The last year I was working as a vice president of data science and all those things. Now, we are currently working into education consulting and working with various companies in implementing AI based strategies. So, that's a bit about me. Now, let's get started on the topic. Okay? So, now the first thing the topic that we have been dealing today, like why does it matter? So, as I told you, after the advent of ChatGPT, okay, the moment the ChatGPT came in, before to that, the whole world didn't know much about AI, what was happening. We were dealing with this kind of technology from '17, '18 in the background. We were trying to work on the systems, but those systems were slightly tricky to deploy and use. So, only some people had control over it. But OpenAI made a big change. They just transitioned everything behind a API layer. Okay? So, now these models are running in the backend. These models are nothing but some algorithm, you can say, which does a data mining and give you some answer based on the query. So, earlier we had to host our own model. But then, OpenAI created API layer and just allowed everyone to have access to these systems without deploying their own. The whole world got a easier access to the AI. Now, everyone has in their pocket a complete running AI which you can just use as a complete second brain, your personal assistant, lot of things. You have a complete sales agent. People are just implementing so many different kind of technologies and I would guarantee you that over the next 5 years, the way things are going to be, we cannot predict and lot of transformation is going to happen. The way we interact with the system, for example, keyboard, mouse, we do click, click over the websites and all. All of those things are going to change a lot and the way we interact with the whole world is going to drastically change. I'm giving 5 years, but even 2 to 3 years, lot of transformation is going to happen. So, that's how these technologies are right now running on. Okay? So, just give me a second before we go. I just need to change the overlay into different screen. Okay? So, just give me a quick minute. Okay? And then, we will continue. Just a second. Let me share it now. Perfect. Okay, I hope you guys can see the screen. I may need just a bit. Just again, can you guys see my screen? Please put yes over the chat. That will be helpful. Otherwise, I'll be keep talking. Yeah, perfect. Thanks a lot, buddy. Great. Now, why are we talking about this enterprise? Every company nowadays is just looking for integrating here. You know, already heard about companies are trying to do layouts and all because they see some value with the AI. Okay, see some are real and some are in a different way, but most of the companies are trying to use AI and they see the potential where the existing people can be reduced. It doesn't mean that they just want to reduce the quantum of people. It's just that they're looking for people who are more advanced on these things. Okay, the people who are not able to learn upskill, then they feel that they're going to get the company down. So, those people are getting discarded so that this company can bring in new people because so much of transformation is happening. So, every enterprise is trying to see the use cases and implement the use cases around the AI and we're going beyond just the chatbots. Now, the systems, which we call AI agents, cannot just chat with you. They can actually connect with the surrounding digital interfaces like it's your database, it's your CRM, it could be all of these different things or APIs or your local knowledge bases, documents, all of this data the systems can access and work on that and it can just answer any questions for you. This technology which is called as agentic AI is just doing a lot of different impact on the overall world. Automation demand is increasing across every single industry. Okay? I currently don't know any single sector or industry where AI is not making any impact. Okay?

If you know any, let me know, but I currently don't know any industry or sector where AI is not useful. It's useful in every single domain and it's really making an impact in every single domain. Only one difference I would say is right now people don't have a good awareness about what is AI and they are not able to identify very well where it is useful and where it is not. Sometimes they just apply to use case that it is not good for and they think we are AI is not useful, but it is not the case. It's useful in some applications and use cases in a lot more better way as compared to others. So, when we understand AI, that part becomes much more clear and we actually bring in the business value. Okay? Businesses need AI that can execute and not just generate the output. Agentic AI reduces the dependency on the manual workflows, people trying to interact as I say. We just dealing with websites as click click. All of those things are going to change. It's going to more chat driven, more voice driven. So, systems are going to take a good amount of control as compared to what we humans do. Humans are going to more of a mediators or decision makers to some level or managers. Okay? Supervisors. And then competitive advantage now depends on the intelligent automation which we can achieve with this systems. So, that's a huge differentiation which the systems are creating. That's why this topic matters a lot. Now, if we let's just have a quick idea around what is JNN what is agentic AI, okay, before we go into it, can someone tell me what is agentic AI over the chat? And there will be no right and wrong responses. Okay, there is no right and wrong. Please try whatever you know about it. Okay, I just want some honest answers. Okay, and I hope everyone must have heard about agentic AI. So, what is your vision or what's your perspective right now about it? Okay. Good. So, Anubhav says agentic AI is just a structured information that an LLM or AI follows to complete a task. Okay, very good Anubhav. I would say it is very close. You have got it very close. Good. Jude says agentic AI is a bot that helps us provide the output that we require in a more customized and fit-for-purpose output. Okay, very good Jude. You're also very close. Yeah, that's good. Very good. Great guys. So, now let's see what it is. You guys have already given some angle to it. Okay, let me just see one more here. Vishal says agentic AI will have one or more LLM, tools, and memory along with the ability to execute the processes. Okay, very good. Upnish says human desire requirements, data in professional way. Okay, good. I think you all are together trying to summarize the whole thing about the Gen AI. Okay, let me now just try to summarize all of it together. See, one thing if you're using ChatGPT, okay, I think nowadays there are a lot of changes. But do you imagine the ChatGPT for 2 years back? 2 years back when we were using ChatGPT, if you try to put some instruction or a prompt, ChatGPT can only help you generate it. Now, I can ask it, "Can you help me generate a LinkedIn post related content? " It can generate it. I can ask it, "Okay, help me draft an email for a client. I want to send them this information. " And it can help me transform it into a nice-looking content. I can ask it, "Okay, help me define this documentation for a code. " Or help me create a code. It'll give you file by file code that, "Okay, you create this file, you create this file. " So, the thing is Gen AI tries to do generation. As the name itself says, it got popular with the Gen AI maybe because it was good in generation. It is able to generate certain stuff. Okay? So, it requires a continuous human prompting. Executes a single-step task very well. Cannot independently decide what is to be done next. It performs one step and then waits for your next instruction. Then no ownership of multi-step processes. If certain thing requires different steps, for example, I want to do some analysis in the database, okay, for a user. I want to understand what kind of revenue that like top 10 users with the revenue. Now, if you have to

use this data, maybe sitting in some databases, some Excel, and some multiple places. Now, the Gen AI cannot do that. But Gen AI, if I ask it that, "Okay, see, this is my database, these are my columns, and I want you to help me create a query. " It will do that. So, Gen AI, if you're focused on generating things, but it cannot take any kind of an action. That okay, I generated a query, but what to do now? Till I do not say this query is working or not, it will not be able to give me the next step. No ownership of the multi-step process. Stops when the response is delivered. Cannot execute transactions or operations. Acts within its conversational interface only. Limited to the output generation. Now, as you can see in this one of the challenges, we are only limited to the window. We are not able to go beyond it. Now, let's see what is agentic AI and how it is different. See, agentic AI, uh agentic AI is derived more from, I would say, the way we humans work. Okay? But, first of all, AI agent. Now, just a little analogy that I'm giving you. See, the agents we have, for example, uh there is a ticket booking agent. Okay, there's a travel booking agent. Or some other kind of agents that we usually interact with. So, what we generally do is a task, we try to assign it to them. And they then try to take care of all the activity. And that agent will be I'm talking about more of a human agent. So, if a travel agent, I ask him that okay, see, I want to go from this location to this location. Can you help me book a ticket? I may not care which bus you book. I may just want to go. You just book a bus. You book a train. Or you just give me the whole detail what will be the flight cost, bus train cost. I'll give you just a feedback what to do, and you just book it for me. I will not do all of this. You help me do this. Same way, agents also work to take our task and delegate it. Okay? We just allow it a task, and it will be able to take care of that whole task for us. Now, this is not only about the generation. Okay? Here, it is not like this agent is only just going one activity and every time trying to take a permission for you from you. The agent is now independent. He will do his own activity. He's expert now. Okay, he's a expert to him whom you are trying to achieve give the task and that expert will achieve all the task with all the resources that he has access to. Now, resources are very important. A travel agent may have this different website links or the agent access to ILS to PTC or some web booking Sorry, the train booking websites and other things. Now, he has this resources access. That you may not have or some other agent may not have. So, all of these agents are specialized. One agent is good in only certain thing. Another other things. Okay? So, agents now are not just generating things, but they are not trying to take things beyond. That just go to the later and can actually take some actions in the real world. So, can also initiate multi-step workflows. So, now when I say agent, now this agent is not coming back to you but taking a lot of decision by himself. He will try to look for 10 websites for the railway booking, for the bus booking, for the flight. He will just do the whole exploration and then come back to you. So, it's a multi-step process that's happening and then it is giving you the response. Break down the complex goal into smaller subtasks. So, what's really happening is the overall larger task. One of the thing that we humans have is called as problem-solving skill. Okay? Now, the problem-solving skill is all about we take a big problem statement, we try to break it into smaller parts and then we try to achieve one by one. That gives us more control. Okay? Uh Guys, can you hear me? Just a quick check. Yeah, you guys can hear me, right? Perfect. Thanks a lot, everyone. Thanks for the feedback. Great. So, now we were saying that agents are not just a generator, but they can actually take some step action in the real world. Okay? Now, in the real world, I would say that anything it could be digital. Now, digital means it could be either a API, it could be a database, it could be some folder structure, it could be your command terminal terminal. These are all digital things the agent can access and perform with. Okay? It can decide what is the best next action by itself. It's smart. Adapts plan based on the changing condition. I think many of you must have seen a situation if you're using GitHub Copilot or Cloud Code or any of them. It may happen that it will do a wrong step and it will automatically say that

"Okay, sorry, I did a mistake. Let me correct it. " So, this kind of things the systems adapt on themselves. Works with the minimal human supervision. Again, if I have to give you the example of a train agent, okay? So, the training agent will be mostly taking all the analysis that, "Okay, from this place to this place, there is this many trains. There are this many buses. " He will do all that analysis and only finally give you the result. He will not ask every single step that, "Okay, this is a train. Do you want to go with this? this? " So, there's minimal human supervision and most of the things will be done by agent by itself. Orchestrates a task across multiple platforms. Now, this agents, okay? They can have access to multiple systems and they can communicate the way the humans do. try to interact with the multiple systems, they can also perform multiple steps. Monitor the program and adjust the execution. This is already told you and delivers measurable business results. Okay? I use agent on a continuous basis in my complete day-to-day life. My code work, my deployments, uh any customers analysis, or most of the things all go through agents. I'm giving my GitHub access to my agent. I'm giving my Azure So, I don't do any Azure-based deployment nowadays over the UI. All of it I completely ask my agent to take care of it. My database I'm giving complete access to my agent. I can ask it any question, it'll be able to go through the tables, create a query, run the query, and give me the output. So, this is what agents can perform. Okay? Now, how does this agents really work? So, first of all, they are based on the goal. Okay? They come up with the goal, and those goals are broken down first into smaller tasks. Now, if you're giving it a larger goal, it'll try to break it into more smaller parts. See, why this goal-driven structure is created is if you try to give it a larger problem statement, the agents may not perform very well. See, even the humans also work in a very similar way in the real world. So, if there's a larger project, it's not like a larger project is assigned to one single person to perform in a company. The task is divided between the multiple experts. You try to say that, "Okay, now software development project, do you need a project manager who will take care of the project management aspect. There'll be a software engineer, there'll be a QA engineer, there's a DevOps engineer. Each of them are going to perform some different steps. So, the whole business objective is broken down into smaller parts. Select the tools and APIs required for execution. As I told you, agents require access to the external digital world. This digital world is nothing but accessible through to a good extent the APIs. Okay, but APIs are only one thing. We call it tools. When we implement agents the name that is generally given a technical to this connector. Okay, so we can even call it connector to the agent. We can include different connector. We can say that okay, database is my connector. Azure is my connector or my CRM is my connector. So, you can assign different connectors to the agent, which we call it as a tool. So, this tools could also be API driven. Monitor the outcome and adjust plans dynamically. Okay, now in many of the situations what happens is my when I'm asking it to work on the database, sometimes it hallucinates. Okay. But when it runs the query, it get to see that okay, this query didn't run because I hallucinated. It now sees okay, I did something wrong. Let me see again what happened. It'll again recheck, again recreate the whole thing, and then work through it. Okay, I did a complete Kubernetes deployment using agents. I didn't know much of a Kubernetes to some level. I knew only basics, but one of a project I wanted to deploy, it just creates a complete end-to-end deployment using the agents. It fails. Okay, it fails multiple times. It took some 2 hours for the agent to finish, but it fails, again sees the error, again make the changes, does the deployment, checks the logs, sees the error, again make the changes, again goes back to like this. The systems are adaptive in nature. Operates with minimal human supervision. This is very important. Okay, now see, many times we call it autonomous. Now, autonomous is basically saying that the systems are completely independent. They will do the whole activity. But still the systems cannot compete fully with the human. Lot of repetitive tasks they have already taken up. But still our human

supervision is required when things are complex. Okay? Most of the basic stuff the systems can do it fully automatically, but if the thing goes complex, then the systems cannot perform very well. For example, nowadays I'm working on a larger project that I'm using agents. But then every step I need to provide a proper input saying what needs to be changed, how needs to be changed. All of this input are human we need to provide then only it will work in a better way. I cannot give it a open task. It will just keep searching, keep looking, and will not be able to achieve some new things. So, human input still requires a great margin. But I would say in terms of coding development earlier the system like this 1 year back the system used to be more of a interns. But now they have become a junior software engineer. Okay? Not just a intern, but I would say 2 to 3 years of software engineer. This is what the systems have become. And they can implement things much more faster. The things that we took 6 months back some 4 5 years uh behind those things now can be implemented within just some 2 to 3 weeks. Okay? So, that is the growth which we have with respect to this agents nowadays. I would say but that all of this gives to the enterprise. Okay? Okay, all the enterprises are running behind revenue as a one of the important thing. So, how can we improve the revenue? And to improve the revenue one of the very important thing is how do you reduce the repeatable things? Okay? If something is repeatable, how do we automate it so that those things can start working in automated way. So, automation plays a very important role. So, automates like the agentic AI automates the repetitive administration operations. Looking into for example, let's say you are running the data engineering jobs. Agents can check Did this agent did this job run successfully? If there was an error, it will be able to see the logs. Do some quick summary and provide you all the summary with respect to the logs. If you are running some customers analysis, it can go through all the customers within the list. Run some background checks on all of them or companies that can analyze some companies. And then give you some kind of output. There's a lot of administrative operations that you can take care with the agents and this improves the overall operational efficiency and the speed within the organization. It reduces the human errors in the multi-step processes. So, usually earlier, see again, 1 and 1/2 to 2 years back, even 1 year back, 2025 was called as a era of AI agents, where most of the effort was on building single agents. 2026 is being more called as a era of multi-agents. Where we are not just building one agent, but multiple agents are working together to solve certain problem with the multiple steps. So, now with multiple agents, a lot of human work is going to be taken care here and this is going to reduce a lot of human errors. Because the systems See, the systems will also make a error. It's not like they don't make an error. The LLMs are also like 95 to 100% accurate. There's still a 5% error. But, the thing is similarly humans also have this similar type of error. But, if some work humans are taking 1 day, agents can perform it in 1 hour. Okay, this is again some high-level picture I'm saying. Sometimes this 1-day work can even be done by 5 minutes and sometimes it may require some effort. It depends on the complexity of the work and the level of the work. And enhances the overall cross-department collaboration, which is possible with the systems. I will be talking to one of the startups today. They were building some agent tech system which was more around integrating lot of departments together into AI agent. And you can just ask it any question, I will be able to get the overall organization level any data that you ask it for. It enables a faster parallel processing of data, improving the overall performance and scalability of the systems. So, the systems are becoming more and more faster. Earlier this were slow, costly. Now the cost is reducing, overall usage is increasing, parallel processing is increasing, accuracy is increasing on a continuous basis. See, the last 3 months if I have to tell you, so much has happened with respect to open and closed itself. Okay, these are the two top models, even the Gemini. These three are the top-tier models right now. All of this over the 3-4 months have done crazy changes. Okay, the whole uh automation and the accuracy with the systems has gone rapidly up. Same thing again over the 6-month, 1-year, these things are going to keep on improving and improving with respect

to the overall accuracy. If we talk about a real-world use case with respect to the agent, I've given already some use cases, but let's take some example of some real-world implementations with respect to the AI agent. So, one of our real-world use cases would be Mercedes-Benz financial services where they're dealing with the CRM. The CRM is your customer relationship management systems where all the customer related details are managed, where sales agent will be able to keep track, okay, this is a new customer. Did we call this customer? Did we email this customer? What are the challenges with this customer? Is he using the service? Is he raising a ticket? Lot of the things will come down inside your CRM. Now, one of the challenge that happens with this sales people have to spend a good amount of time going through the whole history of a customer. What products they have bought? What are they Did they recently Did discuss any kind of a tickets? Was there any challenges? What is their overall spending with us? Lot of this information when you go through UI, it could be spread across multiple tools. Okay? All of this information and the sales person just takes a good amount of time to go through all of it, create his notes, bring out all the important information, and then make a discussion with the customer. But, the agentic based systems just changes the whole way the sales agent are going to work with this data. Okay? So, Mercedes-Benz displayed a multi-agent ecosystem integrated with the CRM. With specialized agents, collaborated to analyze the customer sentiment again from the tickets, emails, and all of it. You try to understand what kind of sentiment the customer have. Are they happy? Was there any bad tone? All of that the sentiments will come. Access the overall history of a customer. And resolve the complex billing disputes or service request in a automated way. Okay? So, agents can now read through all of this data and then some small issues, the agent can solve by themselves. Okay? Not the small repetitive issues. And most of the organizations have like some 70 to 80% of repetitive issues. 20% is where some actual human effort would be required where they have to think through and make some changes. So, 80% of this work, I would not say 80, but still 60% of the work, these agents will be able to take care in a automated way so that humans can now focus on the more complex aspects. Overall, the Mercedes-Benz basically says that they had a 25% reduction in the customer complaints and 15% increase in the cross-selling due to tailored real-time engagement. We as agents and humans can go through like for example one salesperson can go through 10, 20 customers or even 30 customers in a day. But AI with the scale can go through a number. It's up to us like how much we want to scale. So, the scaling again becomes a very big impact and that gives a real-time access. Now, with human being there, humans will be able to prioritize some important ones to the top and the non-important ones they may ignore. But AI can be very quick. The moment the ticket or the query is coming, within a quick few seconds or a few minutes, it will be able to come up with the response and immediately share the result. So, that's the power with respect to the AI agent factor use case. So, this is the one use case, but like this many companies are running behind different use cases. Now, if we talk about the different challenges and risks which are there inherently with respect to the agentic AI systems. But agentic AI is very powerful. Okay, we just talked about all the good things which are there with respect to the system. But every system that comes in with the good good, there is also some kind of a bad or impactful or risk-based aspects which are there with any of the systems. Okay? So, with the very powerful execution capabilities, but it also introduces some significant governance, security, and accountability related challenges in the system. Okay, so first is loss of control over the autonomous decisions. Sometimes if I'm giving you complete control, see as I told you earlier, the systems are 95 to 900% like accurate, 99% okay, 95 to 99 but then still we cannot have this 1% fully, that 1% is always going to be there but still now 4 to 5 on an average as a error is going to be there.

But when it is running, you can't have a full control around it. Some things can go wrong. So we have to keep it in mind that okay, some of the decisions may go wrong because we will not have a full control. There are some ways to set up better human in the loop guardrails and all those other strategies which we can use to manage the systems to some level. Misalignment with the business objectives. If you're not setting the systems in a right way, okay, if there are any mistakes, even a small mistakes, those can bring in some misalignment with the business objective. Okay, because again because of the small nature of issue issue, the traditional software engineering in that we mostly have is if the user is asking bank balance, I'll only search for the word bank balance. If bank balance is present, I will run the bank balance flow. If user is saying fund transfer, I'll run the fund transfer flow. But here it doesn't happen. Here the user may come up with a complex question that okay, I want to know my balance which is there in my bank account. Can you tell me that? Now here those two words may not come in one single go and come in a different way. So Agent KI can handle this kind of cases but with a small error. Security and access vulnerability. There are lot of security based aspects. Now when we say agent, okay, I'm giving it access to my database. The challenge is if you're not keeping all the right strategies when you're giving it access to the risky tools, it can do some unexpected things. For example, you guys may have already heard about some of the situations where recently, I think this last month, uh for cursor, one of our client got their whole database deleted. Okay? The agent just simply deleted the whole database. Uh this kind of things can have a huge impact. Where the whole organization may get impacted. So, in their case, what happened is even that uh full date like the history and backup also got deleted. So, that was a big impact. So, the thing is we have to build the system with the right control. Okay? If you're not building the system with the right control set, you're giving it some uncontrolled access, then these challenges can come. But, what are these challenges and where they can come? I just gave you one example, but there are many situations where this can have an impact. There are also hacking mechanisms now which are very specific to the agentic systems. People have got some hacking mechanisms where they can ask LLM, they know if I use business words, the systems will do things which it is not supposed to do. Okay? I can ask it, "Give me a bank balance for user of this this. " And it may do that. Okay? I can use some very specific words that is being proven that will help this LLM to ignore its existing instructions and allow unwanted requests. So, all of these things are there. So, understanding how and where to use this in a right way plays a very important role. Data privacy and regulatory compliance of this. Now, many regulatory compliant companies won't go with the AI. They stay behind the old technologies because they have to wait because there are huge penalties if something goes wrong in the regulated environment. So, they have to be very cautious what they are using, how they are using, what kind of an impact it has. So, they have to be very secure. They have to go with every single step. So, again, as I told you in the morning, in the start, not AI is not applicable in every single place. There are some use cases where it can do very good, and some it cannot. So, even in the regulated environments, where can it have a larger impact? Creating to that is more important. Error amplification at scale. Now, let's say there are five pipelines. Then, agent one, agent two, agent three, agent four, agent five. I told you that agent one, every agent now has a 95% accuracy. Okay? If this one has a 95% accuracy, and this one also multiplying with 95 95, the output of the pipeline is not 95. It is less than that. Okay? Just multiplying these two, the accuracy will drop down to 90 two. Something close to that. Okay? I'm not giving exact math, but close to 92. Again, you multiply this 92 with the next 50, it'll again reduce. Again, you multiply it with the next module, it'll again reduce. So, if we are not testing the systems in a good way, the overall error factor can go high from all the cumulative errors which could come. Ethical bias and fairness. Now, again

the systems are more biased toward certain aspects because we are using human data, human-created data to feed and train the systems. Human data, by nature, is biased. For example, more of a male-driven data, politically biased data, religiously biased data, or the biased data. So, all of those also become part of the LLM. So, those sometimes can have a larger impact. So, we have to understand where it can go wrong. Integration complexity with the legacy systems. Legacy systems again may not be very easy to integrate. Uh so, that becomes a one more challenge. Over-reliance and the workforce disruption. Then is some kind of a workforce disruption that is going to be there. Some jobs are going to be eliminated. I just came, like last to last week I went to a museum. Again, the museums are there to showcase the history. Okay? And there I was seeing that there are a lot of jobs that we had, like the hand cart or the bullock cart or even let's say creating the utensils and all kind of thing which were all being happening manually by the hand created things. So, those all got automated in the industrial revolution. Same thing, this kind of jobs that we had now, also they will be eliminated and people have to adapt, upskill with the newer age technologies. And that will be there. Monitoring and accountability gap. Now, many of the time the question comes is who's going to take a responsibility if the AI is making a mistake? Will it be the Open AI, Claude, or company who's implementing? So, who will be the responsibility if there is a mistake? Okay? So, monitoring that, taking the accountability, all those things are there. Okay? So, not every company out of over the online, you may see a lot of people talking about AI, but very few actually get to implement and see the things in the production. Okay? So, getting the systems from the POC stage to the production requires a lot of different engineering pieces. Learning about AI is one part. but then again, thinking of how do you integrate it into the whole problem statement plays a again a different perspective. So, some of the tools which are widely used there is a n number of tools. There is a few. There is a line chain, graph, there is a crew AI, there is a hugging face, there is a the recent I forgot the name from the Google one. But there is one framework from Google. There is a one framework from open AI is a security key. There is one framework from uh the AWS itself called as AWS forms. So, everyone has got the framework. Over the future they all will be combined together into only few uh frameworks because that's what happens. When the things are hot, many people will come, but over the time it'll all get consolidated and then only few people will come in that. Okay? So, that's the some of the framework. Okay? So, great. Now I think I'll give it to Shilpa. Thanks a lot everyone. So, Shilpa I'll just take few seconds to take the questions. Yes. Absolutely. Perfect. So, looking to see great like please go ahead with the questions. I'll just go through the ones which are there right now in the chat. So, you don't give it right permission, but prompt injection. Yes, yes. See, again, giving the right permission Anubhav is sometimes we know what can go wrong. We have to give it right permissions in some situations. Like I have to create a new user, lead, or I have to do some insertion of records. So, I have to give it a right permission in some situations, but what kind of impact does it have? How do we handle that? So, those all become important situations. Okay? So, it will not work that we will not give it right permission, but sometimes that could be part of a problem statement. Okay, prompt injection, yes, that's one of a very risky thing which is there within the AI ecosystem. Google AI studio, yeah, so it's not framework. Google AI studio is an interface through which Google exposes its own LLM model. Okay, so Google AI studio, Google just comes up with notebook LLM and all these are like products that Google comes up with. Is this the end of session? Yeah, but in terms of the content it will be we are going to build models or agents going forward as IT engineers. See, Ashwini, like the model building would be required to some level, but how much of hand building that is the question. Okay, for example, one of a very

well-known person called as Andre Karpathy recently said that he runs the model training over the night all through the AI agent, but he has to set it up. Okay, still up that how do I train this model? What will happen in this model? What kind of parameter tuning would be required? So, those all things still have to come from the human. Okay, so agent will itself not do the complete work. We have to still define. So, there the knowledge of how these things works plays a very important role. Anubhav, what are your views on world models? Can you explain the subsequent sub-quadratic approach for models? See, I think world models are still very new. That is still not completely productionized kind of a system that is being the exploration and that is still going on. So, I would not much comment on it. The person the Yann LeCun is the one who's working on this kind of a world models right now, but I think it has some good promise. Maybe some good outcomes will come. We cannot say that as of yet. Okay? So, memory is one of the very hot topics in the whole LLM. So, I would say the overall memory optimization, there are a lot of things. There is a hardness engineering is one of the recent thing. There is a context engineering. So, all these different things are there to manage the whole memory aspect. Okay? We'll fine-tune the existing model for our use case. See, when we talk about the program, so within the program, the fine-tuning as approach would be there where we will be showcasing that because that's going to be one of the hot trend over the future. Currently, we are mostly going for paid models, but these are like very large models. So, SLMs are one of the hot things which is going to come and get stable over the future. And with SLMs, we have to do more of a fine-tuning, domain-based fine-tuning. So, over the next few years, more of that will come. What are the easy way to learn and hosting a local LLM? For experiments, I would say go with the Ollama. And for production kind of thing, go with vLLM. Okay? Thanks, Anubhav. Uh I've read a big percentage of agentic models fail after the promising POC. Will the hype settle down and put it in the future? See, I think more of Aditya, and I would say it is uh see, everyone is running behind without understanding what is AI agent. Okay? Most of the people I talk with, they even don't understand how the systems work. They just run behind it. And the systems are not like traditional systems. The more of a reason for the failure is people not able to understand the overall utility as where these things can be used and where these things cannot be used. That is I would say the poor reason for the failures because they're not picking the right problem statements. If you pick up the right problem statements, then it will perform in a much better way. So, hype is allowing people with the ignorance to just jump into the problem statements and when things are not working because they don't know how the system is working and what they're built for, they just say it's a failure. Okay? So, I would say it's not a complete reality. For production, local development, it's a complex thing, I don't know. I would say you will never do a production deployment on a local server, but V L M is a server that you can go ahead if you want to do the local development with the open source L M. Okay? Yes, V L M is a good one. Great. Great, guys. Thanks a lot for all the great questions. It's a great time to spend with you all. I hope you have got some good learning out of the session. Okay? So, I'll leave it to you now. Yes. Yes, absolutely. Even I did Rishikesh as well as usual. And uh thank you so much for the insights and uh to all of you all, we'll just extend the session for another 5 minutes because uh we're going to talk about the, you know, the gap that we need to bridge right now. Although we discussed about the possibilities of the learning uh the uh ethical uh agentic AI. And Rishikesh, if it is possible for you to still keep sharing your screen, that would be great. I have a little logistical issue in terms of sharing. If that is okay, that would be great. Thank you so much. Okay, I hope I'm audible to you guys. Uh a thumbs up, a yes on the chat box would be great. Um Okay. So, uh talking about the bridge uh in uh or rather the gap that we want to bridge right now is basically as very rightly, I think uh even uh you know, Rishikesh, you mentioned that is to learn the depth of the uh machine learning. Not just understanding as an end user how each of the tools, each of the models are going to just work. Because when it fails, you need to actually get into the nitty-gritties to

get it you know solved. And that's where the core learning journey begins, right? So, I think being an alumni of the Great Lakes as well as the UT Austin program yourself, Rushikesh, I think you shared that experience earlier with us. To all of you all, there are you know structured learning processes and at this point in time, the questions that you all have asked, be it Ashwini, be it Anubhav, all of these being able to provide sustainable solution work and drive AI or navigate your career into the field of AI and machine learning or you know agentic AI right now would require a comprehensive learning. And that's what the organizations also look at, right? So, being being a part of Great Learning, I think Great Learning is one of the leading EdTech company which has been there in the EdTech in the higher education system for the past 12 plus years, right? Yes, Ashwini, there is a placement support for this particular program. I'll come to questions and answers immediately after a quick discussion with you all. So, we have had you know seen 15,000 plus career transition. These are more recent, of course, and we have 4,500 plus hiring partners at this point in time just to answer that question of yours. So, we serve in 170 plus countries and we have courses which is to which is for every segment of you know folks, people coming in from non-tech, technology, people having different segments of experience and so on and so forth. And all of these programs come with you know a high completion rate of 98% with all the top universities that you can think of because ultimately it's about the outcome that we are looking at, you know, that is the career growth. So, be it with University of Texas at Austin, be it associations with MIT, with Northwestern, with John Hopkins, with Walsh, and so on and so forth. Now, having said that, Rushikesh, if you can help me with the next uh slide, if you can move on to the next slide. And I can talk about the PGP program in AI and machine learning, which is uh you know, which is what Rushikesh is an alumni for as well. So, uh the 12-month program, uh while it is it has a structured learning journey, it actually um takes every one of you through the entire uh journey, starting from uh traditional um machine learning to be able to go ahead and work on the modern AI part of the process, which is be it generative AI, and you will be able to work and build models from ground up, right? So, the 12-month program has a flexible structure, 600 hours plus of learning. It's a dual certification PGP level program, which is 80% practical and a 20% theory kind of a journey that we work with. Uh and that gives you that hands-on uh you know, uh understanding to be able to build scalable you know, solutions for the organization as well, which is the expectation. So, there are programs, of course, which is there is a way that you can fast-track this 12-month learning in terms of learning, which is more uh effective as well. So, for that, you can definitely uh you know, connect with us uh individually later. And this 12-month program, it is going to be a personalized learning journey. So, I know a lot of you uh who already have an understanding of AI or have a technical background would think that, you know, uh how is it going to be relevant to my or how will I be placed here? So, of course, this program is going to be a personalized learning journey, wherein you'll be put in a group of 25 learners coming in from a similar portfolio background domain to ensure that the learning is significant and uh you know, it is personalized to the individual as well, and there is a personalized attention from the day one as well. So, that's one of the major things, and as I think even I know Rushikesh very well highlighted to all of you all uh that's the problem-solving approach. There has to be a structure, and being a US-driven program, this is already we're already ahead of time in terms of the technology integration, and uh re- being relevant to the market, right? So, those things are already a part of this immersion immersive experiential uh experimental uh you know um hands-on learning as well. All right. So, uh having said that, uh if we can move on uh to the next slide, uh Rushikesh, that would be great. Uh and this will talk about uh you know this learning journey while it's a combination of sessions taken by um you know industry mentors, top mentors who've been working in AI like the example that we showcased today that was uh Rushikesh. You see the kind of diversified experience he's coming with in the field of AI being able to take those sessions in the weekend so that you learn hands-on. And uh when it comes to the faculties, these are again top-notch faculties from uh you know University of Texas at Austin. We have it from uh you know again Great Lakes Great Learning to be able to go ahead, and all of these faculties, if you really pay attention, they have been uh

educated in the field of machine learning, uh you know uh statistics and analytics, and that is what they will be training you on. These are not random faculties just, you know, coming with any random experience, and that's where the difference lies. And those are the top faculties working with different organization you different, you know, universities and different colleges, and uh they will be the ones who have curated the curriculum as well. Moving on uh to our next topic, uh Rushikesh, if you can help me here. Yes, I think a part of this we have already uh Rushikesh did, uh you know, talk about the um core foundation as well as the deep learning journey that we are talking about, wherein uh the 12-month curriculum is going to be comprehensive enough to start from the Python and then get into SQL, then you know, applied statistics, then it is going to take through the EDA and the data processing model, and then although it says uh self-paced, by the way, for the multi model, uh generative AI recommendation system, object detection, uh you know, segmentations, all of these things are just uh one of many things that you'll be having as a prep work to ensure that we start from scratch, and we go step by step. Now, of course, this is going to be personalized. There will be folks uh who are already hands-on with Python. As I said before, we do not for you, you will not be repeating those Python basic languages uh basic learnings anymore. So, that's going to be customized there to ensure that what you're learning, uh you know, is related to you, and that's where it is going to be taken further. Uh moving on in terms of the depth of the curriculum, if I talk about, I think the domain exposure will start from supervised unsupervised learning, uh introduction to neural networks, deep learning, computer vision, NLP, all hands-on. Currently, the program has actually 38 plus tools that you learn here. This is going to be the older PPT. So, 38 plus tools that we have, and the curriculum gets updated every 3 to 6 months, even to an extent where the curriculum has Anthropic Claude learning of 5 hours as well. So, tools-wise, you learn all the tools, uh be hugging face, be it, you know, crew AI that uh even uh Rishikesh highlighted, all of those will be hands-on in the program as well. And then, uh you know, you'll be part of individual projects through and through in the curriculum from the beginning till the end, hence it's a hands-on driven program. And there's going to be a capstone, which is the major project that you'll be working on. Uh every weekend will be case study-based learning, where you solve real-time business case studies, where you solve for the process and so on and so forth. So, you will learn the proof of concepts. Peter, there will be peer-to-peer networking. There will be AI at work. All hands-on that you'll be learning here. Uh, moving to the next uh, aspect of our discussion, uh, I believe this does to an extent answer Ashwini's question, which is the career transitions. Well, uh, I think this is one of a kind program which provides or comes with a dedicated career support, right? So, we have been I mean, it's not the first or the second year. We've been doing this since 2017 onwards in the field of AI itself. So, that in itself talks about the kind of legacy that we have built and uh, the kind of opportunities that it, you know, it gives. So, because all of these career transitions are only possible um, because of the credibility and the skill set that you learn, without which companies are also not interested because the basic uh, gone are those days just a basic knowledge or just understanding the tools where, you know, is enough. No, it's not enough and that's where you have to have uh, end-to-end understanding with those skills it to be able to crack an interview as well. So, that's where, you know, our expertise into career transitions comes into the picture and there's just a few examples that I talked about or that you can see on your screen as well. Be it uh, people moving in from a non-technical role like in a system manager to uh, you know, uh, complete technical role. Be it people coming in from a software, uh, you know, background technical background and then moving into data scientist AI uh, roles as well. And um, moving on uh, to our uh, you know, nes- next aspect uh, a few more career transitions for all of you to, you know, basically look at. Uh, now, any one of you or every single one of you, I know you will be having questions. I know there will be a couple of, you know, expectations and for that uh, we are all open for you to uh connect with us. Uh we have consultants who speak to you, who help you understand what would be the right fit for your learning journey. Uh if you are already hands-on with AI and you want to learn advanced AI, there is a separate segmentation of learning that we have. Uh if in a case you're a person looking out for a career transition, there are programs that is going to be specific to a 12-months program dependent on what level of depth you're looking at. There are doctoral-level programs for folks who are really interested into, you know, research and development and so on and so forth. So, all of these things

keep moving on and we would, uh you know, want you all to go ahead [snorts] and uh raise your, uh you know, inquiries as well. And for now, I think I'm almost done in our terms of our conversation and I would like to take up any questions if you have. I have Ashwini. I think I I'll just answer this question now, uh you know, directly. So, Ashwini, I hope you're there in this particular meet and uh the placement is a part. I mean, I would say it's a dedicated career support, you know, that we provide in this particular program which starts typically from the 60% completion of the course itself. It doesn't just jump into the career opportunities. It does, uh we have a one-to-one career mentorship first because what we work is not just for somebody to be here and, you know, get outdated just because you're doing a short-term program, you're just relevant for 12 months. No. We build a career long, basically, a sustainable growth for you. And we call it exponential growth where you're growing both in terms of roles and responsibilities as well as in terms of your salary. For that, for anyone, any single one of you, if you really want opportunities, growth, one of the common denominators is how good you are in terms of problem-solving, how aligned of, you know, actually understanding the business requirement and then, you know, solving the problem for them or they really acting in terms of solving the problems for them. So, those proactive approaches, those learning journeys are the key to get those opportunities as well. So, one of the major things that we work with is the one-to-one career preparation first, which starts from the 60% completion and then that is when you will be getting the exclusive access to our own in-house career platform called as GL accelerate. And this Great Learning Accelerate platform is only for Great Learning candidates, by the way. All right, so these are again, we have a dedicated career success team that bifurcates job opportunities basis your profile, basis your years of experience, and they are sent to you. And you get to choose which company, which position you would want to apply for as well. And it's going to be a continuous journey from the 60% completion for the next 2 years. For a 1-year-long program, we provide you dedicated support for 2 years. And that is where you can see the difference in terms of, you know, where we come from. And all of this is not possible without having a credibility. Right? So, I hope this answers your question, uh Ashwini. Uh any other questions? Guys? Yes. So, cost of the program would, you know, more uh importantly, we would like to talk about the selection and the cost has certain parameters uh dependent on the program that will suit. So, uh Ashwini, for this, I would genuinely request you to please uh you know, reach out to us. Uh for everyone's uh you know, context, I'll just draft my email ID so that you guys can reach me out. Um and I can also send my number. 8050554070. Yes. Okay, this is my WhatsApp, though. So, guys, if in case you would want to reach out, please do. Because the reason why Ashwini, I'm not answering that question immediately is because it depends on the outcome that you're expecting. So, basis the outcome, there are different programs which is to ensure that we are able to reach the goal. Because all of my programs, all of these programs are outcome-driven courses. So, if it has to match, it has to, you know, also ensure that, you know, we you and us, we are on the same page. So, yeah. We do have I mean the master's program, the doctorate program, all of these things are going to be a part. This one-year program that I just explained has a uh has, you know, an opportunity for you to go ahead and extend it into a 12 24 months basically at master's program and a doctoral level program as well. So, that's the kind of comprehensive journey it is. Um so, so please reach out to me, Ashwini, for us to go ahead and help you out here. That's number one. Number two, Anubhav, I think you have asked me a question, is there any exam-based entry programs? Uh well, there is. Uh there are programs which we have from IIT Bombay where you really it's an 18-months PG diploma program which requires you to first qualify, appear for a test, and then only you can get into that particular program as well. Uh Sure. Right. So, so there are Although that depends, Anubhav, again, I mean the same parameter which I just told to Ashwini as well to every single one of you who is still sticking, you know, in this conversation. Uh whether you enroll for the program, that's still going to be something that we going to discuss about. The first thing that we need to understand is what is the end goal, right? So, and basis

that, we need to talk about what would be the right solution. Just not about any specific program per se because ultimately we have to reach that goal, and it's your career we are talking about. And that's something that we are very, very specific and we give that amount of time. So, please uh you know uh Anubhav, uh I have given my number in the details. If that's possible for you, please reach out to me. There is going to be an opportunity for us to have a discussion and then accordingly we can talk about which program can make sense. Uh any other questions? There are other folks? Anybody else have any other questions? Uh again, I mean Anubhav, uh just giving you the link does not change the fact that we need to still understand what's the end goal. Uh giving you the link is not a problem, right? And that will not solve any purpose as well. You might give an examination, you might qualify, you may not qualify. What's the goal that we are What's the problem that we are solving? Are you reaching Are you getting a career growth? The answer is still no, right? So, let's let's just, you know, first understand. You've asked very good questions in terms of uh you know, throughout the program. It seems that you have done a lot of learning. So, let's just try and go ahead and continue, yeah? I mean, connect with us for a further discussion. Anybody else? Any other questions? Jude, thank you so much. Yes, so I think we are almost towards the end of our discussion. Yes, uh you know, to all, thank you so much for uh you know, being with us today and it was really nice uh having you all on board. Please review everything, keep learning and definitely, you know, ensure that you're ahead in this wave and please go ahead and uh you know, send me a hi, ping me or just email me so that we can go ahead and connect and have an understanding of what can be the best suited options for you, yeah? So, with this, I'll just go ahead and end the session. Um Yes, sure. Yeah, okay. Okay, Shruti. Thank you so much, guys. Have a good evening ahead. Bye-bye. Thank you.