Welcome, everyone. My name is Mansi. I'm a principal engineering manager here at Microsoft, where I lead the strategy and vision for AI powered applications within Microsoft Teams meetings.Today, I wanna talk to you about a shift that's reshaping the enterprise space specifically, and that's the rise of these AI agents into their growing role in decision making as well as insight generation. Some of you may think why, especially, we're having this conversation today. We have, five industry trends that I wanna talk about, the big trends, which is the advancements in LLMs and GenAI. Think of it like the foundation models that have become far more powerful and accessible, opening the door to these intelligent context aware agents. We have improved reasoning and planning capabilities that these agents aren't just, like, responding to, you know, some some prompt anymore. They're actually trying to reason.

They're actually trying to plan and even adapt to more complex scenarios. Third, I wanna talk about maturity, the frameworks, the toolings that's available, you know, whether it's with Google Vertex AI, whether it's with Microsoft semantic kernel, and how we're making this agentic development and orchestration. Specifically, I'm gonna talk a lot about orchestration today and how it's more easier than ever. Fourth, we have lots of, growing data volumes. If you are from the enterprise space, you know, you guys are swimming in data. So these agents are, like, helping with organizational digest, interpretation, acting essentially, for this data to be consumed faster and more intelligently. And last but not just the least, I wanna talk about the demand for this hyper automation and efficiency, which, every business is sort of chasing, you know, operational excellence and how these AI agents are the promise to a more faster decision, a fewer handoff states, and the smarter systems that can learn over time.

So that's landscape you are in, and it's the convergence of, like, need, capability, opportunity. So let's dive in into what the enterprise agents really are and why they are the key to the next phase of intelligent, enterprise transformation. So every day, enterprises, make thousands of decisions. What if I told you, you know, less than 50% of them had enough relevant data to actually make that decision? Maybe they had more, but maybe they didn't even know what data and insights they didn't have. And that's the gap these AI agents are closing by turning this raw data into an actionable decision, whether it's fast, whether it's, agentic. Speaking of agentic, if you have not seen Microsoft Build conference happening right now, Satya Nadella said in his keynote just two days ago, that we're entering the agent of Agentec we're entering the era of Agentec computing, and that's where these agents will collaborate with us.

They'll reason. They'll plan and act. And it's already here. The Sarah is already here. We're already doing multi agent orchestration. We're already doing, you know, all sort of, logic applications, that these agents can essentially choose their own next steps. This isn't just about, like, some dashboards, but it's more more so about making the decisions that are more aligned to your business goal. And I have a lot of context to share in the upcoming slides, but I want to touch really quickly on enterprise agents and how you guys might be defining it. So before, like, I dive into, like, very specific examples, I want to share what I think is the core characteristics. So we have autonomy. That's, like, our act with without minimal human input, perception, sensing the environment, sensing whether, you know, you have the user interaction, what the system states are, what's incoming data looking like.

We have a core characteristic like decision making that's going to be the central discussion for our talk today. So that relates to, like, applying logic, applying rules, learning patterns, and choosing the right action. Speaking of action, we have it really just doesn't mean, you know, your an your agent is analyzing, but it's actually doing something meaningful, like sending alerts, triggering processes, updating systems. I even wanna add a fifth characteristic, which is more around learning. So these agents are now capable of adapting and improving based on, like, feedback or some observed results. So I'm going to show you some, like, live video as well. Well, not like a recorded video from, from build conference that can really emphasize on the learning aspect as well.

And then to talk about the function categorization, this is where some people say, okay. In enterprise space, what kind of agents are we looking at? And these are some of the top ones that came to my mind. Like, you know, you have had data agents, you know, that are focusing on, like, ingesting and interpret it into protecting these large datasets. You have had decision agents. Think of them like, you know, insight to action agents. So they don't just, like, surface insights, but they recommend or even execute at the next specs the execute the next best action for you. We have had or we're starting to see more of these now, the process automation agents, for, you know, streamlining your workflows, handling approvals, handling scheduling, handling data sync.

The fourth one that I think of is customer facing agents. These ones, you know, they include, support for your boards, for your, virtual assistants, for example. Then you have employee facing agents. It this could be, you know, your onboarding boards. This could be your ID support, agents. This could be your, you know, any internal documentation for that matter, that can be made at GenTech. And then, the other category, last category that I could think of was security agents, which are already, you know, a a a very, it's sort of like the silent sentinel state where, you know, you're already detecting threats, monitoring for, anomalies, you know, triggering all the automated response.

And this is where I I can see a lot of enterprises really investing a lot in the near future. So each of these agents, while they have a very different role, they share a very common goal, that is to shrink the gap between the insight and intelligence action. In fact, if you're coming to this talk and you have a traditional AI background, I wanna say, you can actually instead of having the this function categorization, you can draw parallels to goal based agents, utility based agents, you know, reactive agents for that matter. But that's sort of the foundation. I wanna show a video now from build conference, and this is this will, like, sort of set the stage. You know, we we talked about dashboards decision. So the next logical question that comes to mind is, like, how are these enterprises actually making this real? So this video is actually a very powerful example that we just announced at Microsoft Build, when I say b, Microsoft.

And it's something that marks a new era on how we think these AI agents will work in the enterprise space.

Every patient is unique. And with increasing condition classifications and treatment options,

precision medicine requires sophisticated investigation of vast Can you guys hear the audio just to be sure? K. Perfect. Data and knowledge. Azure AI Foundry helps developers jump start the process of creating a

health care agent to address this challenge. Health care agent orchestrator enables specialized agents to collaborate and quickly complete tasks that would take hours, like analyzing medical images, synthesizing current literature, sourcing clinical trials, and building a chronological patient timeline using clinically grounded knowledge and health care adapted safeguards that support accurate and reliable results.

Developers can fine tune models on their organization data, test performance in a guided playground experience, build or customize agents, and extend agents with Copilot Studio and model context protocol to connect with existing internal or external agents, tools, and services.

Agents can be deployed in Microsoft three sixty five applications like Teams, where users can interact with AI agents in their existing workflows. Throughout, the AI models adhere to the principles of trustworthy AI. By combining orchestration of multiple multimodal AI agents with simplified customization, extensibility, integration, and deployment into existing workflows, health care agent orchestrator empowers developers to work with clinicians and accelerate innovations that can impact their work.

Every Alright. So I think that video is very powerful. What I was really trying to show was multi agent orchestration. And there were a lot of key terms that that video used like MCP, Azure AI Foundry. I want to touch on all of that. I'm gonna see how much I can do in the next ten minutes, so people here can actually take back something meaningful. To start with, the very basic topic that that we do touched on was multi agent orchestration, and that's, the the sort of the groundbreaking transformative, you know, you can the thing that you can take back about Copilot Studio today. And that's a major shift we're talking about. It's not just like a single assistant that's responding to commands. We're not we're now talking about, you know, agents that collaborate across enterprise domains. You have HR, you have IT, you have marketing.

And these are all, like, coordinated, think of them like goal driven units. So each agent is responsible for part of the workflow, and they dynamically delegate these responsibilities based on, like, a business logic or maybe a task context. Let's take practical example. Employee onboarding. One agent handles, let's say, identity verification. It does system provisioning. The other agent curates maybe a personalized training path. Third agent maybe integrates policy document and perhaps even compliant checks. And the magic here is that they're all working in tandem with zero human handoff. And this isn't just like automation. It's about distributed decision making at run time. Think of this like Kubernetes but are an enterprise process for for enterprise processes, and even powered by cognitive agents.

And that sort of marks a clear evolution that we're moving away from, like, a single agent, single purpose copilot to multi agent systems, and that can scale, that can make your organization that might be working in silos to actually work together. A very natural thought here would be, hey. We've talked about, okay, these teams, these functions, and how do how these, automated coordinated workflows need to work. But what if you want these agents to reflect your specific domain, your tone, your terminology, your logic? This is where Copilot tuning comes in. Copilot tuning, it just doesn't give you, like, you know, enterprise fine grain control over their AI agents, but also, you know, it allows you to inject propriety domain specific knowledge, your tone, your terminology, your task specific nuances directly into the model behavior.

So this isn't just about, like, prompt engineering that you might have seen in, like, very basic flows, but this is about agent level specialization. So organizations can combine the semantic configurations between business, you know, grounded training examples to agents that can actually reason, like, internal exports that you might have. One good example here could be, let's say, you are from a law firm and you want to tune your Copilot that that can, like, draft your contracts, maybe perhaps based on, like, just this like, some jurisdiction specific clauses. You might have a firm approved legal language, and even, like, site precedents, for example, like, all without retraining your foundation model. And that's the result of Copilot tuning. That's how these AI agents don't just respond. They represent your brand. And that's where you can leverage something as powerful as Copilot tuning. And they may reflect your logic, your internal, you know, best practices, but, this is where you have the generic Copilot moving towards more enterprise calibrated agents that can bring you precision, that can bring you compliance, that can bring you trust, all at scale.

A very other natural this, discussion from when you move away from there would be, okay. Hey. We talked about making agents, you know, domain aware. Now how do I tune them to think and act more like my internal expert in my organization? And that's where Azure AI Foundry comes into picture. So Azure AI Foundry, you know, it was positioned as the if you if you watch Microsoft build 2025, you know, it it has been Azure AI Foundry has been positioned as mission control for modern AI development. You know, it it now supports, I believe, over 1,900 foundation and open models, and that includes, GROG three. It includes OpenAI, From the very beginning, there's Meta, ZLLM, Lama, and, there's Mistral. And this turns the foundry, the Azure AI foundry into truly multimodal, multi agent platform. So think of it as a at a place where, like, all these developers, they are building, they're fine tuning, they're orchestrating, and they're deploying their agents all within one, like, observability first framework, like, how I like to call it.

The foundry can give you, you know, telemetry on your performance such as, like, latency and throughput. It can give you telemetry on your quality, whether your model is hallucinating. Do you wanna detect that? Do you have any response relevance? It can also give you telemetry on your cost. When you talk about cost for models, you're looking at token level billing insights. You can have, you know, all the insights on your safety as well, whether it's toxicity, API exposure, any jailbreak detection for that matter. So you're really either I if if whether you're spinning up, like, you know, a customer support agent that can maybe integrate with Salesforce or you're building your own in house, you know, analytics assistance, you can really leverage Azure AI Foundry for, not just like a model catalog today, but also like the MLOps layer for all sort of agentic enterprise space.

So we've talked about, you know, the way I like to call it now with with the tools to build, fine tune, and monitor these agents at scale, you know, across multiple models, like I mentioned, GROG three, LLAMA, and all, with, like, built in safety, with cost controls. This one critical challenge that, your enterprise systems could be facing that how do these agents are act like, they actually get plugged into enterprise systems. Like, how do they interact with my files, my calendars, my customer records in a secure, structured, and scalable way? This is where MCP comes into picture. So, MCP actually is like it it's really a game changer for, like, all sort of interoperability, and which is why you can see it being referenced as the USB c of the agent world. So think of it like an open source standard designed to help AI agents to understand and interact these software systems more naturally, not just by scraping screens or, like, you know, reverse engineering some APIs, but through structured context rich access. It's a universal interface, that all companies and enterprises are adapting to, and it's sort of expected to allow agents to plug into these enterprise softwares securely and more seamlessly, and, you know, truly make contextual, decisions.

And this sort of marks a major milestone in in our open in our agentic web era. A very natural question here would be, okay. So we have MCP that's supposed to act as, you know, this major integration problem that we have with all of these, you know, different agents that need to orchestrate together. Very let let's take it one step further, which would be what if these agents could not only see what's happening, but decide what to do next in real time based on goals. And this is where agent loop unlocks. Agent loop, again, it's it's a breakthrough feature in Azure Logic apps that embeds, AI native decisioning, you know, into directly into enterprise workflows. So historically, think of Logic Apps, you know, following a more rigid, predefined set of steps like a flowchart where, like, if this happens, do that.

It's useful, but it's not wrapped up. So agent loop changes that game. It allows an AI agent to actually pause the workflow, evaluate the context, and decide what should happen next. This is all based on, like, real time inputs. It's based on business goals. It's based on logic. Let's say you're processing, like, a customer escalation. So instead of always routing to, like, a tier two support, this agent can, like, you know, analyze sentiment, check historical, you know, resolution data, ticket resolution data. It can even, like, you know, consider priorities. You know? Some tickets might have a higher priority. It can dynamically choose an optimal resolution part. And this isn't just like a flow automation. It's a goal seeking orchestration.

And that's how you should leverage, like, these, like, agent loop to actually now replan, reevaluate, and adjust according to, you know, some mid execution level. I know I gave a lot of examples more relevant and more centered to, like, you know, the work that's happening at Microsoft. But needless to state, the entire industry is working towards it. I want you to take one example and drive you through it. So I touched on MCP, agent loop, and a lot of different things that, like, Azure AI found, Recopilot Tuning. I want you guys to take that back and, start small, at at your pace, at your organization level. So my key takeaway here, you know, we started with a very simple question.

You know, what if your enterprise didn't just surface insights but actually acted upon them intelligently, autonomously? The answer lies in AI agents. These aren't just your tools, but these are your collaborators. From multi agent orchestrations to MCP, we're entering a decision we're entering an era where decisions are made for faster, smarter, and aligned business goals. The three key takeaways here are, you know, you have you're shifting from those dashboards to goal driven decision systems. You're shifting from combining, you know, orchestration to furthermore into tuning and context for deep intelligence. And last but not the least, it's it's not about reporting, but it's about action.

So smart you know, the one thing I can recommend anyone to take away from here is, like, start small, but start now. Embed an agent. Let it evolve with your enterprise, because these proactive transformations, they they happen one decision at a time. And I wish you all good luck in in bringing these agents to your enterprise if that's what you're looking for in this talk.