Welcome to Decoding the Future, a podcast where we explore the latest trends in the world of technology. I'm your host, Ker Yang, CTO of Fujitsu Asia. And in this episode, we will be discussing how organizations can transform themselves with Agentic AI. In our previous episode, we spoke in-depth about the data to AI Journey and Agentic AI was a key emerging trend. Today, I'm very happy to have Adarsh from Microsoft back with us to dive further on this topic. And we've got a special guest, Kenny from Frost and Sullivan. Adarsh, Kenny, would you like to do a quick introduction of yourself? Sure. Maybe I'll start first. Hello, everyone. My name is Kenny. I'm taking care of the Asia Pacific ICT practice and analyst and consulting firm Frost and Sullivan. Good to be here. Thanks. Thanks, Ker Yang. And thanks, Fujitsu, for having me here again. I'm Adarsh. I'm part of the Microsoft Global Black belt team for AI. I work with our customers on the most cutting edge AI initiatives that they have within their organizations. I’m happy to be here Thank you. Thank you, Kenny and Adarsh. So maybe, Kenny, start with you first, agentic AI is rapidly emerging as a top technology trend. Many analysts firm predict widespread adoption within the next few years, and there's also significant executive interest in it. So maybe let's start with the basics, Kenny. As an analyst, what's the current industry perception of AI? Sure. So I think one of the things that we realize. So my role today is really to sort of give everybody a context of where the industry is from outside Fujitsu, outside Microsoft to actually talk about this from an industry perspective. I think one of the things that we've seen around CIOs and CIO teams is they’re all interested in agentic AI. But one of the bigger issues, taking a step back first around AI in general, is that most organizations are not simply looking at the basic functionality of AI. I think one of the things copilot, ChatGPT and others; to improve your emails, make your presentations look better, do a bit of ideation, etc.. I think generally speaking, management teams want to do more. It's not the basic functionality that matters to them. I think that's, of course, a bonus. And, I have to applaud Microsoft for including Copilot into the basic Microsoft 365 subscription. So that everybody can use that. But I think one of the things that the teams really should think about is how do we get value back more? Some of them have very specific business problems. It could be for a very specific line of business. It could be very specific kind of value that they're trying to get. It could be new data product. It could be new efficiencies or a new problem-solving methodology. And that's where, back to agentic AI. It really requires a bit more advanced reasoning, right? Interoperability across different systems and really so that we can have better data and better sense of that data and decision making. So the disillusionment, I feel that there is some disillusionment in the sense of AI in AI investment, in AI transformation. But I think everybody is trying to experiment and understand this AI agents, because this is something relatively new from all the hyperscalers recently. So I think everybody's trying to get a figure on how to move forward. That's interesting. Adarsh, do you have anything to add? Yeah, I mean absolutely. I hear what you're thinking about some of the challenges that a lot of the enterprise customers that we're working with, are sort of realizing as you know, they're trying to adopt this very new technical landscape. But the way we see it is that there is going to be a fresh wave of productivity and efficiency gains that we're going to be seeing from this landscape of agentic AI. And primarily, the way we see it, is that these agents are going to have two capabilities. They're going to be able to plan, proactively plan on your behalf on how to achieve a task. And they're going to be having access to toolkits that are going to be allowing them to execute those tasks on your behalf. So that's what we see going forward. And that could be in the form of your employee agents, personal agents for employees to do the research, as well as customer service agents that can carry out end-to-end automation on behalf of the employee. And that's where we see that agents really bringing that additional value to the bottom line for the customers. I got a question for you, Adarsh, actually. So based on our Frost & Sullivan analysis, we've noted that AI has sort of progressed in overlapping waves in a way. First, it started with predictive, as sort of way of using AI. Then it moved to generative which was happening in the last year and a year plus. And then now we're sort of having moving into this agentic AI. Based on Microsoft and your experience, how do you define it? What do you tell customers when they ask you about agentic AI? Yeah. So that's a really good question. I think this wave of agentic AI that we're going to be seeing is really a continuation of a long list of innovations that we've been working on inside Microsoft. And, as such, the real game changer here with agents is their ability to create a plan and have access to a wide variety of data sources internally within your enterprise ecosystem to sort of like, have access to data that they can reason on, as well as external data through, access to APIs that they can call, for instance, to a search API to proactively retrieve information from the internet, or perhaps to a third-party provider that is providing, financial data or travel data or any other kind of information that they can consume. And the agents will have access to these different data sources, and they'll be able to then plan how they're going to use this, and then they will be able to automatically call those tools on your behalf. And that's a real game changer that we see here. So in terms of real world applications, we're seeing applications, for agentic AI, in primarily in call centers, of course, contact center automation is being a big game changer. They will be the first to gain value in a way, right? They will be one of the first to gain value. But we also see, in software development. So where traditional software development has focused a lot on developers spending a lot of time building not just code, but the pipelines around the code, building the testing and the automation that's required for an enterprise grade software package to be created. And so where we see agents coming in is they can, bring in some of the sort of productivity benefits even to this ecosystem by automating things like ticket creation, watching the board, cleaning up the board as well as, of course, code generation, which is a core capability of a lot of these are in.. How did this really differ from how the chatbots are now, right? Because generative AI has already enabled many of the chatbot capabilities to the automation that you talked about. I think it's an extension of the chatbot. So chatbot is a very, as we say, it's a human centric, sort of like, UI portal. The way we see it, is chatbots are going to still possibly, probably be there as a way for the humans to interact with these agent, with the underlying agents. But ultimately, the current generation of chatbots, many of them are Q&A chatbots, where they have the intelligence to be able to read through a lot of information and possibly answer questions. But at the same time, they are not always designed to take action. So possibly they'll return an answer. For example, if you're working in a contact center, they can tell you, okay, what could be the possible root cause of your particular issue? If you call in with a specific complaint to a contact center, but they're not going to be able to generate, for example, a ticket to a human, reviewer to come and possibly take a look at your action. So what an agent could do is not only sort of answer the question, that the user might have, but also then at the back end generate a ticket to the ticketing system to schedule maybe a follow up from a human being. Also, enter the details into a CRM system so that, you know, this record of the customer interaction becomes a part of the 360 view for the customer. That then enhances your future interactions with the customer. So that's where we really see the value of an agentic interface to AI. Thank you. I think that's a really interesting discussion, understanding the definition of agentic AI. And of course, we look at some of the real world examples for quoted by you, Adarsh. And of course the call center example has a use case that is really compelling. So based on this, what are the three key characteristic that you can name of to let the listener have an understanding of agentic AI value differentiator. So the three key differentiators, as I would see it, is that we need agents to plan on your behalf. So where we see agents are not going to be, so where we see limited interaction, human just give it an objective, and the agents will be able to plan on your behalf. The other big differentiator that's going to basically empower these agents is how do you make sure you give them access to the right data sets that they can reason on and the toolkits that the agents have? And finally, we need to look at how we build the ecosystem of platform that make building and maintaining these agents as seamless as possible. That's going to be one of the key differentiators as well, because, with every technology shift that we have, the platform that you build this on, ultimately gives you the scalability that you need to be able to sort of like get that value addition. So that's what we see. Of course, we can look at the kind of, different kind of tools that they might have access to. And that's, I think, another area that we are trying to differentiate ourselves on, such as on Azure AI Foundry, we have the largest collection of models available that you can plug and play seamlessly into your agentic interface. So you're not limited to a single model provider. We have access to search APIs that can be deeply integrated within your own enterprise data ecosystem through something like an AI search or through external data that you might want to get from a Bing search. So both those tools can be integrated very seamlessly into your agents building on Azure. So that's the kind of like value prop that we see, for agents. That's interesting. Then, maybe to both of you, how do you find differing intelligence level, affects the performance of this agentic AI and how do you see it impacting the autonomy and the way that the actions will come about in terms of the workflow? Maybe you want to start from a technical, more technical perspective? Absolutely. So from a purely from a technical perspective, right? So when we talk about intelligence, we have a lot of benchmarks to measure model performance. So, we see two different paradigms of models coming up now in the market. The first versions were the traditional chatbots, where, it's a Q&A kind of format where you give the model precise instructions and then it returns, an answer for you that may or may not work for you. The latest innovation that's been happening, is catching a lot of wave is the reasoning models. With the reasoning models, where we see is that we give the model a task rather than precise instructions. And the model behind the scene follows something we call ‘chain-of-thought’. And then it basically self-plans the instructions required to execute that task. And then it comes up with that. So of course, in terms of, like the different models that are available, do we need a reasoning model for every kind of use case? Possibly not, right? So what we see is that we also see a need for a mixture of models. The reasoning models are reasonably expensive and they're compute-heavy. And they can be also latency-heavy. But we can mix those reasoning capabilities along with a smaller, faster language model that can do quick Q&A and together they can come together to, sort of like build the entire agentic interface for the customer. That's interesting. Any thoughts from you, Kenny? Yeah, I will come across from a business angle. I think from a business angle, I love your explanation around the different models and the reasoning capabilities today, but I think for a general organization, that would be too much, right? My view is from a business angle that organizations really need to just start, experimenting because I think a lot of, this technical capabilities are developing, will always continue. I think, Microsoft, along with the rest of the industry, is really progressing on AI very, very quickly. And, general enterprise with general business requirements would not have that kind of requirement for very special model that to be created of their own. They will be probably user of an existing model then to actually create something that's custom for them. So I think from a business angle, I think the organizations really should get started. It really shouldn't be waiting for a finalization in a way, or stabilization of the development or innovation is heading. The innovation will always continue. So I think we had just to do, put the flat in the sand and just get started in that way. Thank you, Kenny. I think, it's rather interesting that we understand from Adarsh first on the cost of implementation. Then with Kenny, with your explanation, that sort of state the business justification for it. So I think today, it's really very compelling that agentic AI has a business case. So maybe for us, from Fujitsu point of view, this is where we find that, in our real, some of the case studies that we developed with our customer for agentic AI, one of the latest one that we do is how we can integrate, intelligent workflow, towards the few support services. I think it is an extension of what Adarsh, you previously mentioned from the call center. So for us is, we are looking at using integrating vision AI together with this, agentic AI solution where we’re looking at detecting certain operation constraint or operation scenario, for example, in the warehouse where, you have many of these forklifts, moving about and that actually, encompass some safety concerns, whether the staff or the workers are near the forklift that will maybe, have issues to safety. And this is where we also can apply it to another scenario where you are in the airport, where you are on the runway, where you look at how planes, maybe is departing and whether all the safety personnel are actually evacuated out of the path of the aircraft. So this actually we can use vision AI today to detect, and it will trigger an auto alerting system into the ticketing system, or maybe to the emergency alert system to notify the control room that, hey, there may be a safety breach. If the staff is actually coming too near to the aircraft or to the forklift. So these are some of the easy way where you can actually quickly trigger the alert and let the workflow take it through and actually comes from the safety perspective to prevent any mishap from happening. So maybe Adarsh, Kenny, would you like to share more additional thoughts into this kind of workflow area? No, I mean I want to, I actually I wanted to get your view. But on my point, I think from CIO, CEO perspective, many organizations don't know the case studies, frankly. So, the case study that you mentioned, I think is very important for technology vendors, technology service providers, to really demonstrate that kind of case study to the end user, because agentic AI is so new, even if you go to the websites of the hyperscalers, very often it's a hypothetical scenario. It's not an actual case study with actual measurement and actual value or actual results. So I think that's going to be important part. And that's my feedback from the people that I’ve been discussing with so far. Thank you, Adarsh? Absolutely. I think the kind of use case that you described is an excellent example of what we think of as multimodal reasoning, where you need to reason on top of, the images that you're seeing, from possibly your camera feed. But also you need to possibly reason on top of, some instruction manual that may be part of the safety briefing provided, right? And agents may be having access to these multiple data sources and reason on top of them together. That's the kind of like differentiating the sort of value that we can bring. And definitely, I agree with your point that, this is still a very new area of, sort of innovation that we're going through. And so there's going to be some level of, sort of trepidation for CIOs to adopt this technology. But at the same time, this has been built on top of a legacy of like, lots of innovation that's been coming out from the different sort of research agencies. And I think the kind of benefits that we've already seen with the first phase of generative AI application proves to us that the next phase is going to be equally productive. So we are going to see that value addition coming from, the adoption of generative AI. And you can see this, for example, in the cost reductions, right? So the scaling laws for generative AI models have been faster than even Moore's Law in terms of its innovation. So we are seeing a cost reduction, almost a 50% reduction in a lot of these model providers in six months timeline. And this is being spurred on by innovation from different countries and different model providers. So we definitely going to see that continuing. And so that's something that we are looking forward to as well. Seeing that how we can build the business ROI that's required to make some of these agentic use cases actually feasible. So maybe, we have talked quite a fair bit to introduce agentic AI. So what do you think are the key concerns that organizations can have to implementing and governance need? I'll go from the technical angle and Kenny, you can add from the business side of things. From the technical angle, I think the risks of agentic AI adoption are largely similar to the risk of generative AI adoption, right? The risks around responsible use of AI. So we make sure that the use cases that we're using it for, they are governed under the responsible AI principles that we inside Microsoft have and our customers also, generally are open to adopt those principles. And this can be regarding how we make sure that the interactions with the generative AI models are safe and consistent for our users, but also the other major concern that a lot of CDOs and CIOs have is the risk of data governance and data over sharing. So how do we build any toolkits required to make sure that the data that is provided to these agents are authorized and the correct kind of data that they need to make the decisions. So we are not over sharing or under sharing the data. And so we can bring in different sort of technical toolkits required to make sure we have end-to-end data lifecycle map even within these autonomous agents. So that's something that we're looking at from a technical perspective. I agree with you completely. In fact, I think, I echo the what you mentioned about data, right. So the use of data, the loss of data, potential loss of data, in the use of tools, especially nowadays, there's so many different tools and many organizations’ staff are really using it on their own personal basis, their own personal account, etc., so that loss or potential for data loss is very real. So I echo you directly on that. But I think the second one from a business angle, one constraint or one concern, I feel as a consultant, we should really, organizations should really be proactive because I feel that most organizations or many organizations maybe, not the top 10% more proactive organizations, the rest of the 90%, are just waiting and seeing still up to today. So my view is that the risk is actually a risk of inaction. You're not taking action on AI. Cause we seen so many different phases of innovation, right. We've seen IoT, we've seen so many different technologies come and go over the years. And I think AI is very clearly here to stay. The one question that have always been asked to us is, Is this hype, right? Is this going to fade away? And, the state of innovation, the amount of investment, the amount interest in this is undeniable in the AI space. I believe this is here to stay and I think is really for organizations to take that step and don't wait. Just get started. I think that that really has to begin. Thank you, Kenny. I totally agree with what you have said. I think do not wait for the technology. Of course, the technology of tomorrow will always be better than today. But if you don't start today, you will never get started at all. And you will always be back behind. Yeah, yeah, that's very good. So coming back, maybe I'll just like to touch on the bit on, we’ve talked quite a fair bit on the workflow, the AI, the responsible use from Adarsh. I think the next things that probably we need to focus a bit on is, security. This is where, Fujitsu, actually, we have been working to make sure that we address the security essence of this generative AI, especially on the large language model. So we have worked quite a fair bit to develop solution that address or mitigate hallucination, AI hallucination. This one part of it. But another part of it is, we are also looking into attacks that target, phishing out the sensitive data of this GenAI or this large language model. As we see more and more front-end deployment of all these generative AI, this is where we need to be cautious. How we can actually secure the data that GenAI have, to make sure that, people accessing it, are not, are really the one that has the right access. So this is where we deploy, and develop a kind of, AI security agent that actually address the vulnerabilities that GenAI have. Okay. So coming today, we have talked about so many use cases, so many in depth. I'm wondering looking ahead, maybe this is a question for Kenny, what are your predictions for the broader adoption of agentic AI? And what advice can you give to organizations considering investing in this technology? Well I think just like we mentioned, get started. There's just one thing, but most of the time we would see that, maybe I just give you an example because in Frost and Sullivan, we've also had our own seminars, we have our own groups, coming together as well, and the number one area of interest is AI Every time we ask for feedback, it’s AI all over the place. And the number one question that is being asked is what should we do next? So it's very clear that organizations are still in that phase of “How to? What do I do?” They still don't really know. So I think that one thing that we really as a, as a community in technology, is really showcase value and showcase ROI. The number one issue with AI transformation really is how do we get value or how do we gain value for that organization? Every organization is different. Some are very tech forward, some are in the middle and some are very, very backward. Maybe even digital transformation is not done yet. But every organization is to figure out on their own, where they're going to get value from AI, generative AI and now agentic AI. What's the ROI? How do I make money using AI? Right? I think that's going to be the big question. I think in terms of area of focus, in terms of what we see moving forward, we think there is going to be enhancing customer experience in some form or some manner. The chatbot you mentioned was a great example of that, the field service example, you mentioned is a great example of that as well. How do we enhance customer experience, enhance customer safety? I think, enhancing user experience as well. That means the people in your team using systems as well. And then of course, better organization, better automation or better integration. I think that's generally something that we have to think about. But taking a step back, I think AI transformation is fantastic, but I want to really pull back to the core of it. To really begins with the business ROI. So the business teams, even though we keep thinking of, AI as a technical endeavor, really, the starting point is the C-suite, they have to see the business value. They have to understand what's the business ROI, and then they will open their cheque book, because, unless they don't see the business value and business ROI, they won't open their cheque book. And it will always be a small PoC, a small case study kind of project. So I think the important thing for all the guys on call and all the guys listening to this, is really how do we sort of help the business teams understand the value behind AI. And agentic AI, in this particular case. Hope that makes sense. What do you think? Yup, definitely, Kenny. Thank you. So, Adarsh. What are the key technological advancements needed to make agentic AI accessible and cost effective? What's your thoughts? Absolutely. I mean, first of all, just echoing on the statements, that any agentic AI use case or innovation that we bring in to our users and our customers has to be generating business value for them. And what we've seen one of the big sort of hurdles for adoption of generative AI applications, has been one is the cost factor. Hey, how much is this going to cost me? And the other one is going to be technology skill set factor. How do I adopt this as quick as possible? I don't have the right skill sets within my organizations to achieve some of this. How do I make sure that I can re-skill my team for that? So from that perspective, we really see three things that we, in Microsoft, are looking at to enable our customers to adopt this new framework as quick as possible. So from the cost perspective, as I mentioned, the scaling laws for model providers has been quite strong. We are going to continue seeing, possibly a cost reduction, and cost of training, we've seen. But the cost of inferencing, the cost of generating outputs from these models is going to continue possibly going down in the future, and that's going to continue improving the business case for some of these generative AI applications. So possibly at this point of time, you might say, you know what? I'm not sure this is going to be worth it, but this is going to be an investment that you need to make into the future that's going to be enabling you to achieve those kind of cost benefits in the future, as well. So that's on the scaling laws of model providers, that we're definitely going to see that. But also in terms of how do I bridge the adoption curve for my employees to be able to build on top of it? That's where we are looking at it inside Microsoft from a platform perspective. How do we build the platform that's required to support these agentic interfaces and make the adoption as seamless as possible? So within Microsoft, for example, we have our Azure Agent service. That's sort of a one click deployment for agent interface that brings in a lot of pre-built connectors to data sources that you might typically want to use. For example, it has pre-built connectors to your CRM systems, it has pre-built connecters to your Microsoft Fabric data lake., it has pre-built connecters to external data search using Bing search. So those kind of platform, two links that we can bring together, to support our customers as they start adopting these AI solutions. So their cost of innovation goes down quite a lot. And the other thing that we're doing in general is we're coming up with a lot of open source frameworks as well to support our customers as they build on top of our platform. So, within Microsoft, we have open sourced, two very popular, agentic frameworks. One is called, Semantic Kernel, which is a production grade open source library that is available in multiple languages that developers within your organization can start take up, and quickly build an agentic interface with like, just about 30 lines of code, and they're up and running. And then we also have a more experimental version of a multi-agent library called AutoGen, which is providing the cutting-edge capabilities in developing, not just a single agent interface, but multi-agents. How do we bring together multiple agents to talk to each other and provide the different intelligence capabilities that may be needed to solve a very complex problem. So, we inside Microsoft are innovating on an end-to-end sort of delivery kit for our customers. Thank you. Thank you, Adarsh. I think definitely we need to leverage on the skilled, where the benefits comes in from Microsoft to deploy all these generative AI or agentic AI through pre-built API. I think the development that Microsoft has done would definitely help, and also address the business ROI perspective that Kenny put forward. So coming back, I think, today we really, really need to drive agentic AI moving forward. And from Fujitsu point of view, I'm looking at the vision that tomorrow, hopefully, we can have autonomous decision from agentic AI that helps us to drive the business to move forward. So, other than that, maybe I would like to have a final hear from Kenny and Adarsh. Do you have any more to add on to let our listeners get further insights? Well, maybe I'll start I think, maybe just to sum up, start with basics. What's the business ROI and then do a digital transformation, then the AI transformation. I think the third part, is really work with vendors, work with service providers that have that case studies, that have that experience working at it. So those will be the three things I want to maybe leave for the audience. Yeah, absolutely. I completely agree with that. Get started. I think that's, my biggest thing. And I think you mentioned this many times, Kenny. Just get started, the cost of experimentation is much lower in this particular boom cycle that we are seeing in technology adoption compared to previous cycles. So I think it would be, like the cost of experimenting with these techniques is going to be much lesser than previous times. So just get started and work with, as you said, your service providers, work with the different sorts of, like experts in the area to identify, the ROI for different use cases. And how do we calculate that and how do we, make sure that we're following that. And as I said, the innovation is going to keep continuing. So keep up to date. There are lots of different sources online as well as, offline sources where we can sort of like, work with our customers to sort of build in the training required to sort of keep up to date with the innovation that's going on. And so there's definitely, three things that I'll be looking at. Thank you to both of you. It has been such a great discussion today. Thank you, Adarsh and Kenny, for sharing your insights with us and most important, for sharing the insights to our listeners for this podcast. With that, I would like to sign off for today. Thank you. Thank you.