# Microsoft AI Tour, Mumbai | Keynote with Satya Nadella

*Microsoft India* · https://www.youtube.com/watch?v=JTJPMJW9JZU

## Summary

Satya Nadella delivers the keynote at the Microsoft AI Tour in Mumbai, celebrating AI innovation in India and walking through Microsoft's AI tech stack, from Copilot and agents to data (

## Key points

- Nadella says building AI solutions inverts the traditional process: start with the outcome and evaluations, then the spec.
- Copilot is framed as the browser for the agentic web, with built-in agents like researcher, analyst, and an Excel agent.
- Microsoft's data layer includes Work IQ, Fabric IQ, and Foundry IQ, which together enable context engineering for AI applications.
- Microsoft Foundry offers 11,000 models, including the open-weight Giga Time model that simulates a costly immunotherapy test from a tumor image.
- GitHub Agent HQ unifies coding agents across github.com, VS Code, and CLI, with mission control for all agent sessions.
- Nadella demoed a personal deep research app using multi-agent decision frameworks like LLM council, DXO, ensemble, and debate to pick the best ever Indian test cricket team.
- Agent 365 provides identity, security, and data loss prevention governance for agents.
- Microsoft announced a $17.5 billion investment in India, its biggest in Asia, plus a new 100% renewable Hyderabad Azure region.
- Maha Crime OS, sponsored by the Chief Minister of Maharashtra, reduced investigation turnaround times by 80%.
- The NCS portal, powered by Azure and Azure OpenAI, serves around 150 million workers with AI-based job matching and skill tools.

## Notes

### Opening Themes
- Satya Nadella opens the Microsoft AI Tour in Mumbai by celebrating AI solutions built in India, mentioning Crime OS sponsored by the Chief Minister.
- He stresses that the frontier of what is possible is changing, requiring a mindset of unlearning and learning, new skills, and a new tool chain.
- Building AI solutions inverts traditional development: start with the outcome, create robust evaluations (evals), then the spec — not left to right.

### Experience Layer: Copilot and Agents
- Copilot is described as the browser for the agentic web, with built-in agents: researcher (with computer use), analyst, and an Excel agent.
- Zoe demos Copilot on mobile, agent mode in Word updating a monthly business review, and agent mode in Excel building a budget dashboard with native formulas, charts, and conditional formatting, using GPT 5.2 released the previous day.

### Data and the IQ Layer
- Context engineering is key: Work IQ (tacit knowledge under M365), Fabric IQ (semantic models and operational data), and Foundry IQ (next-generation retrieval that plans and executes reasoning over data retrieval).
- Models are becoming a commodity; success depends on the experience layer and context engineering.

### Building Agents
- Agent builder, app builder, Copilot Studio, and Foundry (with multi-agent orchestration) make building agents as simple as documents or spreadsheets.
- 11,000 models available; Giga Time, an open-weight model built with University of Washington and Providence Hospitals, simulates a costly immunotherapy test from a tumor pathology image.
- GitHub as Agent HQ: coding models grounded in your repo (Code IQ); India expected to be number one in GitHub participation by 2030.

### Karan's Agent HQ Demo
- Delegating tasks to Copilot coding agent from github.com, VS Code, CLI, or mobile; mission control view of agent sessions; custom agents; Copilot code review; security fixes via Microsoft Defender for Cloud integration; plan agent in VS Code; bringing fine-tuned models from Foundry into Copilot.

### Satya's Deep Research App
- Built on his own Azure subscription using Windows 365, GitHub, and code spaces; decision frameworks include LLM council (after Andrej Karpathy), DXO (lead researcher, critical reviewer, domain expert roles), ensemble, and debate modes.
- He uses it to pick the best ever Indian test cricket team, showing chain of debate annotations.
- Agent 365 blocked his deployment due to DLP rules, demonstrating governance, compliance, and security for agents.

### Customer Examples
- Adani Cement: invoice to payment reduced from 20+ days to 4 hours; Aditya Birla: conversational interface; LTI Mindtree: wall-to-wall AI-first transformation; Mankind: AI for medical reps; Yes Bank: 40+ use cases.
- Maha Crime OS, sponsored by the Chief Minister of Maharashtra, cut investigation turnaround times by 80% and helped deliver justice to a crime victim in Nagpur.

### Infrastructure and Investment
- Token factory equation: performance per token is about tokens per dollar per watt.
- Azure regions in India: Pune (Central), Mumbai (West), Chennai (South), a DC with Jio, and a 100% renewable Hyderabad region next year.
- Sovereignty options: public cloud, sovereign controls, confidential computing, private cloud, and locally operated regions; Copilot data processed locally within India's data boundary.
- Announced a $17.5 billion investment in India, Microsoft's biggest in Asia, building on $3 billion announced earlier in the year.

### Closing Video: e-Shram and NCS
- e-Shram national unorganized workers database provides social security benefits; NCS portal offers AI-based job matching, an AI resume builder, and skill recommendations; around 150 million workers are on the NCS portal, powered by Azure and Azure OpenAI.

## Transcript

[00:02] [music] [applause] >> Good morning. It's It's uh wonderful uh to be back in India and here in Mumbai and have a chance uh this morning to talk about uh AI and mostly to celebrate uh all the AI and the AI solutions uh that are being built uh by people in this country and this state and everywhere. And it's just exciting to be able to come back and see uh the progress even year on year uh that um I you know see uh you know in action. To us, it's always about empowering people and organizations. So, when we talk about even new technology um you know platforms, it comes down to how is it being really used to empower an institution, an organization. Um I'm going to talk about in fact Crime OS that the Chief Minister was here, you

[01:03] know, really sponsored and made happen uh in this state so that even uh every citizen of this uh state can benefit. And that's the type of innovation we want to see um in any place. And but the key thing for us is to be grounded in the fact that the frontier of what is possible is changing, right? I mean, that's the exciting part for any of us is when we talk about technology, it allows us to reimagine what we can now do uh whether it's in our state, whether it is in our organization, whether it's in our lives, right? So, the ability to move the frontier. So, we always start there, right? With the customer experience. Now, you can think about and reimagine it. Whether it's the employee experience, whether

[01:49] it's the operational efficiency, uh whether it's the rate of innovation. If you're a drug company, can you really get through clinical trials faster? Can we get more new drugs out there? Uh that to me is ultimately moving of the frontier that we are all chasing. And to do that though, you kind of have to embrace a new approach even, right? You What worked previously in some sense, you have to do this hard job of unlearning and learning what's possible, right? Sometimes we sort of pattern match too much uh from what worked in the previous era uh may not necessarily work. In fact, there's a new frontier. So, when I think about that, it starts with that mindset, freeing ourselves uh that there's a new

[02:36] approach. For example, in when it comes to AI, right? If you look at even a building of an AI solution, you don't start with a spec. You actually start with the test, right? I mean, when people they have fancy new words talk people talk about evaluations, right? They you create create these rubrics to do evals on a particular outcome, then you create a learning system, then you hill climb on that learning system. Guess what? That's kind of literally inverting the process of building a an an information system. You start from the outcome, you create a robust eval, and then you spec versus going left to right. But the key thing is that means you got to pick up new skills, all of us. Uh and and in order to pick up the new skills,

[03:25] you kind of need a new tool chain that you are using every day. Uh and and to be able to then have all of your data that is being brought uh to harness this new tool chain, this new skill set with the new mindset. That's really uh the way we want to approach it. So, to facilitate this, right? This can't just be an abstract framework. This has got to be an everyday practice for all of us. Um you know, it's an information worker today or a software developer. Um all of us are going to be using essentially the same approach. To realize that, we've built a tech stack. And I want to just literally walk up this stack or walk down this uh tech stack. So, those starting with the experience layer where we have taken this idea that, you

[04:15] know, let's bring this AI copilot right where you work, right? You may be you may be in Outlook, you may be in Word, you may be in Excel, you may be in Teams, you may whatever you are doing your work, how do we bring the copilot there? And think of copilot though as not just a request response, right? That's what we've been doing now nearly for 2 years, right? You you know, we know now that hey, we can ask it any question, it'll come back with a better answer than anything that we have previously experienced uh in let's call it an information retrieval search type experience. But now, we have agents. In fact, think of copilot as your browser for the agentic web. In fact, right inside a copilot, you

[05:04] have very powerful agents built in. Like one such uh agent is your researcher. And researcher with computer use is a real breakthrough. It's literally like having 24/7 someone very sophisticated who is your research assistant who can go take any topic and then because they have a computer, right? So, they just are not like think of it as not a single shot I go and bring back information after thinking. They can the researcher can use a computer to do research. Right? So, it's not only you have a computer, your researcher has a computer. So, that's the way to conceptualize. Same thing with analyst. So, I can load a whole bunch of Excel spreadsheets, you give it to an analyst, and it's kind of like giving it to your data science

[05:52] department, right? Or the data scientist. So, it's again having a 24/7 data analyst who can then go give you insights. And lastly, it's like one of the things that I'm very very excited about is what happened in software development is going to happen with Excel and Excel agent, right? So, I in fact had a chance to participate recently in the Excel world championship. By the way, ESPN even broadcast Excel championships. We have an honest-to-god uh great way to sort of celebrate the Excel prowess. And so, I was able to participate in it you the digital challenge using this Excel agent. It is so cool to be able to see and I worked on Excel all my life essentially. And to be able to sort of see how the agent mode understands

[06:42] deeply the entire semantics of a spreadsheet, right? So, it's not just a one shot I created a great model. But it's the ability to iterate on a pretty sophisticated model continuously using this agent mode just like how we do in software engineering. So, it's just exciting to see this. And also, it's really to me at the end of the day, all these tools are like the new Word, Excel, PowerPoint, Office, right? That's kind of they're as simple as that. Uh and they're in the hands of everyone here. And I see it being used like that, right? So, no one is sort of you know praying at the altar of these tools. They're just using them. Uh and and to be to see the proliferation, the diffusion is probably faster than

[07:27] anything I've seen, right? I've come to back to India, I've come back to Mumbai throughout every generational shift in technology starting in the early '90s. And but when I see now the rate of pace of change, there's no difference quite frankly. I could be in any city anywhere in the world uh and to be able to see uh all of this uh applications just being broadly diffused. To show you some of this in action, let me invite up on stage my colleague Zoe. Zoe, take it away.

[07:59] >> [applause] >> Thanks, Satya. I'm Zoe, a product marketer on the Microsoft 365 Copilot team. And like all of you, I start my day commuting in and preparing for the day ahead. For that, I'm going to use Copilot on my mobile. Hey Copilot, what's on my agenda for today? You've got a few things lined up today. First, there's an AI Tour India meeting that started earlier this morning. Then at 11:00 a.m., there's a Contoso sales call. Later in the afternoon, you've got a team sync at 2:00 p.m. Thanks, Copilot. And has my manager sent me any emails? Yes, you've got a few emails from Nicole. Uh just yesterday, she reminded the team about a Friday mixer happening in the common area. She also mentioned

[08:54] that finalizing the monthly business review should be a high priority. She also Thanks, Copilot. Copilot doesn't just have my work context, it knows who I am, what I work on, and who I work with. It's why I use Copilot. It's a true AI assistant. Now, I wanted to take a look at my manager's ask around that MBR updates for my V-team. And you'll see I'm actually using GPT 5.2, which we released yesterday and it's immediately available to all Microsoft 365 customers. Do you Yeah, it's exciting. I know. >> [applause] >> What I love about GPT 5.2 is it's allowing Copilot to be even better at finding the most relevant resources. So, my responses are very contextually relevant and I'm able to take quick

[09:47] action. And it makes a very visual output. Now, only Copilot is able to combine this constant model innovation with your work data to provide the most relevant response. Now, I want to go and update my existing monthly business review document, but I want to change it with my November data. For that, I'm going to use agent mode in Word. Agent mode allows the agent to edit, to draft, to format the document, all with natural language, but it still keeps me in complete control. And I can see it's starting to look across all of that work data, right? My emails, my meetings, my team chats, my files. Now, that normally takes me hours to do. I have to find the right information, I have to sift through it,

[10:39] but now you'll see the agent is actually actively applying it on the left-hand side. And you'll see that it's actually leaving parts alone that haven't changed month over month over month, like our mission, but rather using this to update everything and keep me on par with exactly the November updates. Now, last, I want to make sure I understand our team budget for the next fiscal year. I'm going to again go to Excel, and I'm going to turn on agent mode and ask it to get me a little bit more aware of what's been spent, what's uncommitted, what's at risk, and to bid build a visual dashboard so it's easy for me to understand my data. Agent mode in Excel natively speaks Excel just like Satya has said. So, it's

[11:25] not going to create a random number that I don't know where it came from. It will actually build formulas right in my spreadsheet, right where my budget already exists. And it will even create native objects like charts and tables, things that I can continue to manipulate. It also walks through everything it's doing on the right-hand side. So, that way, if say my manager Nicole comes to me and asks, "How did I come to my conclusion?" Well, Copilot actually helps me understand what it did. That way, I can continue to work with it. I can also ask agent mode to do simple things like conditional formatting, which you'll see, highlighting things in red that maybe are at risk. All of these things to continue to work with Copilot

[12:09] and agent mode in order to get a better understanding. And you can see it's created this dashboard live. I can If I click in, I can see these rich formulas it's actually built to give me an answer. The conditional formatting to highlight what's at risk in red. And a native graph that actually I'm able to manipulate even further, should I want. Now, Copilot, it hasn't just made me better and faster at my job, it's helped me make more impact. And I can't imagine working without it. Back to you, Satya. >> [applause] >> Thank you so much, Zoe. Whenever I see that Excel demo, I like to think about it that, you know, when I had hair, I loved Excel, and when I don't have hair, I love Excel. Uh, I'll tell you, it's

[12:58] one of those generational things that just gets better and better in spite of my follicle challenges. Um, you know, to me, uh, if you sort of take that experience layer, one of the things that's super important in the age of AI is to think about your data as one of the most strategic assets. But you got to bring that data contextually to the AI. In fact, people talk about it as context engineering. Uh, and I think it's a good term for it because in fact, a lot of what needs to be done, both at this engineering level and even, I would say, business decision-making level, is to think about not the silos of the previous generation, but the contextual way you can feed intelligence with data. So, to that end, for example, underneath

[13:53] Zoe's demo, is perhaps the most important database in any company that uses Microsoft 365 today, right? If you think about it, the tacit knowledge inside any corporation, any organization, is the knowledge of people, their relationships with other people, their work artifacts, whether they're spreadsheets, documents, their conversations in email, meetings in Teams. To be able to relate all that, because that's where the tacit knowledge of an organization is. Uh, and it's in that database. And so, fundamentally, what we have transformed that is into work IQ, right? So, when we say work IQ, it's the stateful part underneath M365 that now can be brought to bear in the context of any AI solution you want to

[14:46] build. Not just Copilot, any agent you build out there, you can now use work IQ. Think about that. The second thing we've done is, in fact, one of the other ubiquitous tools all over is Excel and Power BI. So, that means you've taken real great pain over decades building these semantic models to do analysis. And now, with Fabric and Fabric IQ, you can take all of that data, along with all the operational data, and make it available to your AI. And then the last one sort of really makes it all come together. So, what we've done with Foundry IQ is said what people describe as retrieval augmented generation, right? So, one of the biggest things that we have done in the last couple of years as we build these AI solutions is

[15:37] to do retrieval of data. And we've now built the next generation retrieval system, essentially. It does more than retrieval. It in fact plans and executes and reasons that plan all that plan on how to retrieve relevant data from work IQ, Fabric IQ, and any other blob storage, unstructured storage, or what have you, and bring it in the context of your AI application. So, that's what this IQ layer is. In fact, it's probably one of the more salient, yet important pieces on how anyone who builds any AI solution going forward will be more successful, right? In fact, direct correlation between the evaluations or the sort of performance of your models will be how good is your data being brought. Just because in some

[16:30] sense, models are becoming, quite frankly, a commodity, right? So, there is lots and lots of models out there which are very, very capable. The question is, how do you marshal that capability? I think it'll come down to having a great experience layer where people can discover the agents you're building, and having a fantastic sort of capability around context engineering. So, the next layer for us uh, is really about now that we have an experience layer, we have the data set up, let's build agents, right? And the beauty of building agents, quite frankly, is like building spreadsheets or writing a document. It's not more mystical than that. In fact, agent builder that's built right into Copilot

[17:11] is like saying, "Hey, I doc new," right? Or spreadsheet new. Uh, and so we start there. In fact, now we have an app builder. So, in fact, Bill's big thing always, for the from ever since I joined Microsoft in the early '90s, was, "Why the heck do I have a difference between a document, a website, and an application?" Like, in you know, the his concept was there should just be one thing, and it should I should be able to transform. And guess what? We finally have it. Uh, which is app builder can you can go in and just specify the app you want to build, and it'll build it. It's kind of like a document, it's kind of like a website, it's kind of like an app, right? So, that's the power. Then Copilot Studio is another way to think

[17:52] of it as build out of your agent. So, again, as a natural you're using a natural language prompt, you can go build an agent, you can ground it in some data, uh, and you there you have it. You have an agent. Um, and so we're building out Foundry, in fact, without any cliffs to be able to then say, "Okay, we now have agents. Can I even orchestrate multiple agents and build fully agentic systems?" Microsoft Foundry allows you to do that as well, because the multi-agent frameworks are there. All of the tools, uh, that you need in order to be able to build multi-agentic systems are there. We also now, of course, have all these models. We have 11,000 models. You say, "Why Why do I need 11,000 models?"

[18:37] In fact, even for me this week, it is a great reminder of why having this diversity of models is going to really make this world a better place. In fact, we launched, uh, and we announced, uh, in Foundry, uh, as an open weight, uh, model, uh, a model called Giga Time. It There was an article published in Cell, and quite frankly, it gave me goosebumps when I sort of read about it and learned about it, uh, because it's work Microsoft did with both the University of Washington as well as Providence Hospitals. Um, and what it does is, if you think about the challenge of someone with cancer, um, today, when you sort of try to understand whether immunotherapy will work, you kind of have to model the

[19:26] immune system, essentially. And so, there's a test for that. The problem with that test is, uh, it's a complex test, it takes a long time, and it costs a lot of money. And so, what this particular model does is, it does a simulation of the test. So, you can literally take a simple pathology of the tumor, right? So, literally like a microscope image of a a tumor. And then use essentially GenAI to fill in. So, it will do the simulation of what is the basically spatial proteomics of the immune system and say, you now can have a test. So, I was thinking about it, right? With that model, that means some entrepreneur in Mumbai is going to pick it up and say, you know what? I'm going to build an entire sort of simulation lab

[20:27] in silico and make it available to every city in Maharashtra and in this country, right? So, every hospital can start doing what now today is done in the few tier one hospitals and cost a lot of money. I mean, that's in some sense, when you're playing the game, sometimes you forget why you're playing it. This is why we are playing it. We are playing it so that we can actually use this innovation to make a real difference uh in the world. So, it's exciting to see uh how this is all coming together. Now, we're bringing, as I said, all this together in this tool chain. This tool chain is going to become what you will use. Uh in fact, GitHub is where you would start interfacing with all of this app stack,

[21:11] right? So, if you think of the new app server as Foundry, uh the tools are Visual Studio and VS Code and GitHub. And and it Talking about GitHub, it's really exciting to see the growth of GitHub in India. It's I mean, it's tremendous. By 2030, India is expected to be the number one country in terms of GitHub participation. [applause] You know, it's it's tremendous to see not only just the participation, but the but the ambition level of the projects uh coming out of India are just continuing to improve and, you know, get better. And so, it's fantastic to see that. Now, we're building on GitHub, by the way, as Agent HQ. So, what I mean by that is you go to GitHub and you will have access to all the coding models

[21:58] that you can use in the context of your repo. Uh and so, that's the idea. It's not about sort of having one model, but you have your code repo, just like Work IQ, think of this as Code IQ, right? So, that's kind of really it. Like, your code is your most important resource. Yeah, you want that. And then you want many many models over the years to sort of keep working and creating more of that repo uh and the repo features. And so, that's what And by the way, the form factors, right? If people say, oh, wow, will I only use autonomous agents? Will I use only an IDE with agent mode? Will I use a command line? Guess what? I use all three all day long, right? I mean, that's it. Like, the form factors

[22:39] uh of how I interface with AI are going to remain essentially always uh some IDE, some uh CLI interface, uh and a a new chat interface, if you will. And that I think is the way you'll interact with it. But to show you all of this in its full glory, I wanted to invite up on stage my colleague Karan. Karan, take it away. >> [applause] >> Thanks, Satya. A few weeks ago at GitHub Universe, we announced Agent HQ, which is an agentic platform that unites every agent, making them native to the GitHub flow you're already familiar with. So, let me show you a demo of what that looks like. All right. So, I have a demo app here that's a storefront for some of the nostalgic toys, which might have been common in many of our households.

[23:30] However, as you can see, the app doesn't really have any search or filter options and I want to add that functionality. So, I have my code repo here on GitHub. With the new agents panel, I can now kick off an agentic task from anywhere on GitHub by just describing the task and delegating it to the Copilot coding agent. So, Copilot starts working on it in the background. I can assign a task to Copilot from anywhere I work on github.com, in VS Code, in the CLI, or even on my mobile. All right. While working across multiple repositories and tasks, I have delegated many of these tasks to Copilot and I want to know what's going on with it. So, now I can see all of my agent sessions in a single mission control

[24:21] view, both ongoing as well as completed ones over here. Not just that, I can also start off a task from right here while using custom agents that I have created or have been built by partners. So, these custom agents are domain experts, which extend the Copilot coding agent across all of your tools and workflows. All right. So, let's dive into the uh session that we just kicked off and see what Copilot is working on. So, I can see the progress of Copilot here on the left and also the draft PR on the right. So, I'm a developer, so which means I am still in control. So, I can even steer a ongoing session midway to have Copilot work in a slightly different direction. So, this will take a few minutes, so

[25:14] I'll let Copilot continue working on that in the background, but I want to show you one of the uh agent sessions I kicked off earlier for the same feature. So, I can see it's completed and then there's a summary as well. And oh, nice screenshots as well, thanks to the Playwright uh MCP. I can also take a look at uh the files it has changed over here and also ask it for any change, if I need, right over here. All right. Amidst all of this, if I want to have uh you know, a chat with Copilot to ask it something or so, I can jump into Copilot chat right here as well. So, Copilot provides me a choice of various different models from different providers, including GPT-5.2, which just launched last night and

[26:03] already available on GitHub Copilot. See how fast it is. And yeah. All right. Not just that, I can also bring in my own models from Microsoft Foundry. So, I have a few models that I've deployed on Microsoft Foundry, including a couple of them that I had fine-tuned myself and I have brought in all of these models to Copilot by connecting them to my deployments in Foundry. So, now I can actually use all of these models directly within Copilot, be it on GitHub or VS Code. Now, I was just doing something and I fine-tuned one of the model uh just for fun to see if I can converse almost like a developer, you know, in Mumbai. So, I just asked it something simple saying, what is Kubernetes? Your pods, nodes, kuch samajh mein nahi aata

[27:00] hai. And this is coming and it says, simple language mein samjho. This is what containers are. This is what Kubernetes is. And it's so awesome, right? Because this Yeah. This is coming from my own model that's deployed on Foundry, which I can use in Copilot, right? That's pretty awesome. All right. So, speaking of deployments and everything, um I had also created a PR to implement a new product detail page uh you know, sometime earlier and I can see that Copilot has already reviewed my PR and also left a few comments on it. So, let's take a look at some of the suggestions it has left, like this one here, which speaks about uh DB connection pool exhaustion or, you know, this one uh right here, which speaks

[27:52] about type safety. So, this is the new agentic Copilot code review, which combines LLM detections, tool calling, and also deterministic scans that help you with smarter reviews. Not just that, I can actually ask Copilot to help me implement all of these suggestions in a single PR, which will again kick off an agentic task for me. All right. You all know how important security is, right? So, thankfully, I'm also a security admin for this account on GitHub. So, with the new integration between Microsoft Defender for Cloud and GitHub Advanced Security, I can now specifically filter for runtime risk issues in production that have been identified by Defender. All right, right here. So, that's really

[28:47] nice and a campaign to address all of these alerts was created and oops, I see some of them in my repo as well, but not to worry because Copilot can also help me fix these security issues. So, I can just select all of these and then say, assign to Copilot. And then Copilot will go off and then create a PR to address all of these security issues. All right. You know, so now we have multiple different agents working on different tasks in the background, but I want to write some code myself as well. So, here is uh the agent session for the search and filter function that we were looking at earlier and um I want to make some changes to it. So, I can continue iterating with this by opening this directly in VS Code, which will open up

[29:42] the session for me. So, I can either check out the branch or even apply the changes directly in my current branch and I can continue working. So, I will open up the familiar chat interface here in VS code and I'll switch to my agent mode probably use one of the models I deployed from foundry and I will ask it to make one small change, which is I also want an auto complete feature for the search bar as well. So, this can continue working in the background and you would have seen that just like on github.com, I also have a mission control view in VS code as well, where I can see all of my local agents, my cloud agents, my background agents, which is running as well. All right. So, I want to implement a new feature.

[30:39] The toys are great, but I want to make a bundle of it and almost gift them as a play box. I know many of you would love that as well, right? But, I need to plan how to go about implementing it. So, with the new plan agent in VS code, I can ask Copilot to create an implementation plan and also keep iterating with it until I am confident and ready with it. So, I had created one of you know, this plan earlier as well, which I'll just open up right over here. So, you can see it gives me a detailed plan for all the steps that needs to be accomplished. So, whenever I'm ready, I can kind of continue coding this or I can hand it off to an agent as well to work on it in the background. So, I can hand it off to the CLI agent and I will

[31:33] say include the changes. That's fine and this will start working on this agent in the background. All right. So, some of this will take a bit of time to complete. So, meanwhile, if when all of this task would have been completed, let's see how our app would look like. So, it would look like something like that with all of the filters that are there and also my search and auto complete feature and I can go start playing my brick game. All right. >> [applause] >> So, in the past few minutes, we went from so many different ideas, planned them and also coded them using Copilot coding agents in the cloud on github.com, in VS code, in background using CLI agents, had that code reviewed with the Copilot

[32:26] code review agent and also implemented the security fixes using Copilot auto fix agent. All the while having ability to look at all of my agent sessions, steer them in the single mission control view on GitHub or even in VS code. And that is Agent HQ. Thank you. Back to you, Satya. >> [applause] >> Thank you so much, Karan. And so, you know, when you look at all of these tools that Karan showed, you know, what does it do? It inspires you. So, recently we had Thanksgiving in the United States. So, we had some time off. And so, what does one do when you have time off? You build. And so, I built an app. And so, this is by the way, this is my regular laptop you are seeing. So, this is um in fact, my Azure subscription.

[33:23] And this is in fact my app that's deployed right now, I think in in Canada Central. This is my repo. This is my GitHub repo. So, this is my code. In fact, the way my standard sort of way I work is I have a Windows 365 setup. So, it travels with me everywhere. And then, of course, I use GitHub. And then I use code spaces. So, so this is I think it'll show up with the code space here. So, it's kind of like turtles all the way down, right? So, I have my code space running on GitHub, running on Windows 365. So, that's kind of my virtual environment that I travel with. And of course, like as Karan was showing, every morning I get up and my favorite place to go is here, right? Which is I go to my repo and I go

[34:08] to my you know, Copilot and start assigning it tasks. Like every day there's some multiple agents I kick off. Like this morning I said, "Please upgrade this entire app to GPT-52." Right? So, and it's actually done and when I'm done with my keynote and I'm on a flight back, I'll kind of get there on my flight Wi-Fi and be able to actually look at the PR and push it. If you will. In fact, I'll test it once with my code spaces and then I'll commit. But, that's in fact, it shows you the power of how you can literally start really taking simple tasks to big updates and start powering them powering through them. So, what did I build? Um So, what I built is essentially a new deep research tool. You know, one of my life's ambition is

[35:01] to figure out how to get a job in the Copilot team. And so, I'm preparing for it. I'm sort of desperately trying how to get competent enough to get hired. And so, therefore, I said, "Okay, let me build a new more sophisticated deep research application." And so, the what the thing that I did is you know, I said, "Look, there are all these models available. So, why don't I create decision frameworks that work on top of the model?" And one of those decision frameworks is Andrej Karpathy, who was a famous AI researcher, came up with this LLM council, he called it. Right? So, the idea is you will have like like a selection committee, you will have multiple members of the committee, right? So, in this case, I can have GPT-51 or 52

[35:48] today, Claude, Gemini, Lama, you know, Grok, whatever, right? Pick your favorite selection committee. Then you elect even a chairman. Like you can say, "Who is the chairman? Is it Claude or Gemini or you know, GPT or what have you?" Pick your chair and then go ahead and issue a query and ask it to do stuff. And then another one I did was DXO. This is a pretty cool decision framework. This is comes out of health care. In fact, we first implemented In fact, we've implemented it in health care, where you have just different roles. So, for example, one role is a lead researcher, who does Let think of it as exhaustive breadth first research. There's a critical reviewer. And the And by the way, you can assign different

[36:36] models to these different roles. So, I have a lead researcher that happens to be Claude, a critical reviewer that happens to be a GPT. And what does a critical reviewer do? They That particular role finds method problems, right? Say, recency bias or period bias, whatever are the issues which are cognitive biases. Then you can have domain expert, data analyst, so different roles. And In fact, when we applied DXO in health care, it turned out that having multiple roles work together in a multi-agent system outperformed any single model. Right? It's intuitive, right? Why not have all the smart people working with you on a particular decision versus one. And so, that was DXO. Then I implemented even another one called ensemble.

[37:21] And the idea with ensemble is you sort of fire off parallel queries, anonymize the responses, so there is no bias even in selection and then synthesize, right? In fact, last night I said, "Well, let's even implement a regular old debate." So, this is a new feature I added and thank God it's showed up here. >> [laughter] >> And basically, think of it as chain of debate. In fact, I I think of all of these as instead of thinking of them as chain of thought, thinking of all of this as chain of debate. And I said, "Let's actually have a full-on debate." So, the bottom line here is you can have different formats of you know, even debates. You can have pros and cons, SWOT. You can have risk versus impact, counterfactual

[38:07] versus devil's advocate, right? So, you can even pick your form framework. You can have critiques. So, each sort of each person makes an argument, everybody else critiques. You can have multiple rounds of it. So, I was just kind of I had a lot of time, I guess, in my hotel room yesterday. But, so I went on and added features. Like So, I said, "Okay, let's have" Oh, by the way, I even added because I was spending so many tokens, I said, "Man, I might as well actually add cost control to my thing." And so, I said, "Do I need fast lower cost or deep higher cost?" I'm picking balanced here. I can pick the different research analyst, system architect, risk officer, three types of roles, and then have

[38:44] multiple rounds. And so, you get the point. So, of course, you build all this, what does one good old South Asian do? the best ever Indian test cricket team. I think we need one. And so, if I go to my history, you'll see it. Like I've been at it. I've been working at it. In fact, I have a fantastic MLB lineup as well. But, since we are here in Mumbai, I'll show you some of the stuff, right? So, I'll show you the council one. I'll show you what it came back with. In fact, let me kind of increase the font here. So, this is the chairman's synthesis. It shows me, you know, obviously Sunil opens with Sehwag, Rahul, Sachin, Virat's A man, it picked So, this is the synthesis and it shows VVS at six and seven and then it goes

[39:31] on. Here's the interesting thing, right? The chairman's sort of report says that there was unanimous essentially consensus, right? Gavaskar, Sehwag, Dravid, Tendulkar, Kohli, Kapil Dev, Ashwin, Bumrah. But, the big debate, of course, was, you know, do you need an additional batsman or not? And thank God they selected, at least being a Hyderabadi, I'm really proud that they made VVS Laxman in there. Um and here's the interesting thing. So, he waited 5-1 and Claude's inclusion. I mean, think about this. These two models bid for my man Laxman, right? Because uh why? Of his crisis management. And sort of made the point about why. Uh and then the other debate was who's the captain? Um in fact, I think Oh, by the way, the Kumble

[40:16] versus Zaheer was also a very good one, right? Which is uh it sort of chose Anil versus uh Zaheer. Zaheer is going to be 12th man. Um you know, depends, right? I I guess that's where I think there's more work to be done. I'll pick Zaheer some days and sometimes um uh depending on where where you're playing. Captaincy debate, you know, it picked uh Kohli. So, this the key thing is I see the annotations uh of every the chain of debate and it gives me insight. Um in fact, if I go back um to show you one other case, let me go back to my history again. Uh let's see. Let me go to the uh the DXO view. I think this is interesting because the DXO view is slightly different, right? Uh it says here is what the lead researcher did. Um

[41:06] then it says here's what the critical reviewer did. But here's the cool thing. Here Look at the way it found the biases, the error bias. Right? So, the critical reviewer's job was to say, "Okay, when I'm selecting something, are you mixing up your stats?" Right? I mean, like I always say, like, you know, sometimes you could say, "Wow, the fitness levels of the modern teams are so great and wonderful and what have you?" But man, like to watch Gavaskar go and bat without a helmet or you know, even a I don't know um Hazare go and uh you know, uncovered wicket in play uh in England. Like, how do you really equate for that, right? So, that's the error bias. And so, to me, to be able to kind of have all these models debate

[41:52] that issue uh is what this is all about. So, anyway, so you get the point, right? Which is I You can use these decision frameworks as the new form of meta cognition, right? So, if you have all this abundance of models, the ability to literally do what I did, which is just build your own multi-agent system, right? That's the new commodity. I mean, the reality is, you know, this is lit- all of us are going to be doing this work just like how we do spreadsheets or documents. And if we have that power, then the question is what are we going to do with it? And one of the things I think is exercise better judgment, better decision. And in high-stakes situations, in health care, uh in financial services, in insurance, in supply chain.

[42:43] And that's the beauty of it, right? To be able to think about these as meta cognition frameworks for us to apply better judgment uh in a world where we have all of this. One of the other things I should mention is I built all this. I even made it in fact a co-pilot. So, this is all in my Azure private tenant. Um and I So, I then did a you know, basically an uh an open API uh to it. Uh and so then I made it into a co-pilot agent. And then I said, "Great, let me go deploy." And guess what? You know, I was caught. So, in other words, the Microsoft DLP data you know, loss prevention came in and said, "You can't deploy it uh because uh you this has to be deployed inside the Microsoft tenant. You just can't

[43:27] point it to an external agent uh because that's a massive exfil problem." And that is thanks to this other part, which I think is the runtime of agents we have built called Agent 365. And Agent 365 takes our identity management, our Defender, our things like Purview and data loss prevention, all of that and brings it to the agent world. Just like what we have done for end-user computing, if you will, we're now doing it for agents. Because if people are going to build powerful agents like the one I built, you want full compliance, you want full visibility, you want full governance of it. And that's what the Agent 365 does, right? Because you can't just run around and say, "I have thousands and hundreds

[44:08] of thousands of agents without that security and governance and visibility." And that's really uh what uh Agent 365 delivers. You know, the the beauty of all this is there is significant momentum, significant ambition level to translate this ultimately into people doing unbelievable things. So, this morning I had a chance to meet with, for example, the folks from Adani Cement uh who are really doing this invoice to payment uh and reduce something that took 20 plus days to 4 hours. Uh I had a chance to meet the team from Aditya Birla where they took their mobile app that's doing complex products and sort of said, "Okay, here's a conversational interface that just makes it simpler for them to be able to interact." I had a chance to

[44:54] meet with the folks at LTI Mindtree. They're doing a complete wall-to-wall transformation using things from co-pilot to co-pilot studio to a foundry to do every business process and relook at it from an AI-first perspective. Mankind, it's a beautiful use case of really putting that a power of AI and the expertise in the hands of the people who are use the med- medical reps who are visiting doctors in rural areas to be able to make sure that you can the medical rep is confident in answering the questions and then improving care outcomes. Yes Bank has done a fantastic job of really the invoice turnaround times, right? So, they're also doing a wall-to-wall transformation with 40 plus use cases. But really seeing the ROI in

[45:41] each one of these examples. Now, one of the chances I had again was to meet uh with the team that built this new agentic system called the Maha Crime OS. Uh so, this is was sponsored by uh the Chief Minister of Maharashtra who is here with us. >> [applause] >> And

[46:06] And it was really great to meet uh both the SP as well as uh the people in the team uh and in fact, the investigative team. Uh and the pride they had in in being able to use this system to deliver sort of justice to a citizen in Nagpur who was uh unfortunately a victim of a crime. But the ability for them to use this agentic system to speed up the time to justice was just fantastic to see. So, let's go ahead and play the video. >> [music] >> Today, policing is all about cyber crime. It is all about financial frauds. >> [music] >> And it is global. Nagpur rural receives about 30 cyber crime reports every single day. So, you can imagine the kind of influx of cyber crime reporting that that's taking

[47:04] place. >> [music]

[47:12] [music]

[47:17] [music]

[47:29] [music] >> Maha Crime OS serves as an investigation co-pilot. You're able to have an agent that is able to detect the patterns, able to make sense of it, and have it speak to the direction of the investigation itself. The Maharashtra government through Marvels has collaborated with both Microsoft as well as Cyber Eye in terms of adapting Crime OS for the needs of Maharashtra. Ethical [music] and responsible AI for public good is our motto. And Microsoft has provided us a very safe and advanced platform.

[48:13] >> [music]

[48:19] [music] >> Maha Crime OS AI is built using you know, multiple technologies. Underlying this is Azure and AI services through Microsoft Foundry. We're also using Defender for for cloud and Microsoft [music] Fabric. With Maha Crime OS AI, the turnaround times of investigations are >> [music] >> decreased by 80%. Almost 100% of investigations are getting digitally registered and investigators are acting upon. >> [music] >> So, that makes me really proud that we're able to adapt emerging technology to fulfill the needs of several millions of of people across >> [music] >> the state and make government function much more effectively and transparently for them. If AI is properly used in governance, we can make the lives of

[49:14] people easier.

[49:25] >> [music]

[49:33] [music]

[49:40] >> It's [applause] It's a great example, quite frankly, of using the power of all this technology to help citizens of this country. And more importantly, it also sort of is a great example of how quickly this technology is diffusing, right? I mean, if you would have thought about any previous era of tech and the rate at which it would come to sort of impact in a context like this versus where we are. And that I think is the opportunity in front of us. So, it's fantastic to see. So, the last piece I want to talk about is powering all of this, of course, ultimately is the infrastructure, right? You have got to run essentially a token factory. And when it comes to token factory, one of the most important things is this one

[50:25] simple equation, yet very very important, which is performance per tokens is all about tokens per dollar or per rupee per watt. Right? You have to optimize. And And this, by the way, is going to be the most important thing, right? Whether any country, any community, any state that is able to sort of really have the infrastructure that delivers this efficient frontier of the token factory is just going to have more abundance of tokens that can then be deployed to drive the productivity, which directly is going to be correlated to the GDP growth in the economy. And so, to that end, we are all in, right? We are building out Azure as the world's computer. We are investing in building out Azure in India. We now have multiple

[51:19] regions. In fact, we right in in in Maharashtra, we have a Central India region in Pune. We have West India in Mumbai. We have South India in Chennai. We have a DC with in partnership with Jio. We are also very very excited about sort of our new South Central region that is going to come out in Hyderabad next year, which is going to be on the next big data center region. And we keep expanding. And to that end, I'm really By the way, the the region in Hyderabad is going to be 100% renewable, right? So, that's the other cool thing is we are really focusing on sustainability as well. >> [applause] >> And And to us, we are also building all of this with the making sure that there's real sovereignty. Because after all, at the

[52:11] end of the day, this is not just about bringing the capabilities, but it's also ensuring that there is real sovereignty controls. And so, for that, we have a full set of options, right? You can use public cloud. You can use public cloud with sovereign controls. That means you have you know, you're the key management locally. You can use confidential computing in concert with it. So, that means when the data is being processed, even in use, at rest, in flight, it's encrypted and the keys are with you. We also have the private cloud option. We also obviously have a region that's operated by local partners like Jio or NIC and others. And so, we have a full set of options. So, think of it as a portfolio that you can then

[52:55] use that portfolio for the different workloads you have and have the assurance of your own sovereignty and control over it. We are also doing, in fact, sovereign data control. So, Copilot and all of the things we talked about are all being locally processed. So, none of the data is going outside of the the data boundary of India. And so, we're very excited about sort of all of these options being available. Now, one other point I would also make, since we talked about sort of crime OS and and what have you, is one of the biggest challenges is for any online infrastructure is cyber and cyber risk and cyber security. And it's an intelligence game, right? And ultimately, whoever has the best intelligence is going to be able to

[53:44] defend you the best. And so, Microsoft sees trillions of signals every day. We see every nation-state and every cybercriminal everywhere in the world. And we are able to harness that to protect. In fact, in the demo current showed was pretty cool, right? Which is he basically was able to take your production sort of instance, have Defender check in real time what are all the vulnerabilities, and then have you even go back and fix those vulnerabilities. And that's not a static list. That is a dynamic list that is sort of being monitoring essentially what is the cyber activity and the cyber attack sort of graph. And so, to me, really thinking about sovereignty and cyber resilience are two sides of

[54:29] risk, right? So, when you think about even you managing your risk, you want to think about it. Because if you get disconnected, you have a sovereign everything, except it doesn't have the best intelligence, it's sort of you're going to be very very vulnerable. And so, therefore, it's a much more nuanced discussion that needs real thought. And then, at the end end of the day, it's not a one-time decision. It's a continuous monitoring of that portfolio of risk and opportunity. So, to that end, we are very very committed to this country. And therefore, one of the things that I was excited to have a chance is meet the Prime Minister earlier this week and even announce the biggest announcement. >> [applause]

[55:14] >> It's It's our biggest investment in Asia, 17 and a half billion, which builds on top of the 3 billion we announced earlier in the year, in fact. And so, we are obviously very excited to bring all of this infrastructure to empower ultimately all the citizens of this country, all the developers of this country. And so, that's where I want to close out with, which is to us, all this is about empowering every person and every organization in India to achieve more. To us, you know, whether it is you know, things like gigatime, which allow us to think about making what is personalized care or even precision oncology possible in every tier two city in this country because they can an entrepreneur here will take

[56:06] an advance like that and parlay it into health outcomes that matter. Or the Maha crime OS that's going to deliver the justice to someone who was a victim of some crime in Nagpur. To us, ultimately, that's why we do what we do. All of us working together to realize the opportunity ahead of us in this age of AI is an exciting time. I hope you all have high ambition for what you can do. We sure do for you all. Thank you all very very much. >> [applause]

[56:45] >> India has more than 400 million unorganized workers. They're earning on the day-to-day wages and not getting any [music] benefits. We need to shift the people from the informal to the formal sector. [music] That would help us to develop our economy. Through the initiative of e-Shram, the national unorganized workers database was developed and we are able to provide those workers [music] with social security benefits, also giving them access to the national career service portal. NCS is a one-stop solution for job seekers, employers, training providers, and counselors. We have AI-based matching between [music] the job seekers' skills and the employers. So, any informal worker can go to the NCS [music] portal and match

[57:26] with the right employer. NCS has an AI resume builder. And if a person is lacking certain skills, the [music] AI tool helps him to develop those skills by giving them recommendations. We were expecting millions of unorganized workers to register in this system. [music] So, we wanted to have scalability, reliability, and security. Microsoft Azure met all these qualifications. In addition [music] to that, Azure Open AI has become a game-changer for unorganized workers.

[58:00] >> [music]

[58:11] >> Around 150 million workers are now on the NCS portal. We are also [music] now trying to bring in more and more schemes to benefit these workers. AI is supporting us to provide opportunities for the individuals and youth of the country. [music]