The Future of Windows: AI-Native Computing with Pavan Davuluri

0:07Hey, what's up everybody? We are at Microsoft Ignite 2025 and today I'm joined by Pavan Davuluri, who heads up the Windows and Devices team at Microsoft. Thank you for having me. Yeah. Excited to be here today. Welcome to the show. I want to start by asking you specifically, so under your leadership, Microsoft has continued to redefine what the Windows ecosystem means in an AI-driven world. Sure. How has your vision changed as AI has become basically like a huge part of that? It's a great question. I mean, you know, I personally think of this as an amazing learning journey in itself for us, both as a team and organization in terms of how we build our product, but also how we serve the breadth of Windows customers and users, both in commercial environments, certainly in consumer gaming and education and many, many other places.

0:56For us, I feel like this is a journey that started a couple of years ago, in a way. In fact, May of 2024, we started with this idea of Copilot Plus devices about the time we were also building the Copilot. And really in that window of time, one of the fundamental things we had to think about was what was the role of Windows in this environment and how do we build Windows in a way that can contribute in this space, but was also sort of modernized to be able to be effective in the world of AI. And so we spent a couple of years thinking about what does that mean from a system stack from the silicon platform, you know, that Windows devices are run on?

1:32What does it mean for the operating system itself in terms of being aware of AI workloads? We had to think about what the developer platform needed to be over time. And certainly we spent a good amount of time on what experiences do we want to bring to life on Windows, both as a Windows team, second from a Microsoft perspective, certainly with the co-pilot work that's happening across the company. But really for us, the magic is Windows being a platform to have others build on top of Windows, both in the consumer commercial space. So it has been kind of a multi-step journey for us to think about the layers of how to bring value to customers at the end of the day.

2:05Right, right. And Windows has always been the backbone. Like if we think of it, like it's the backbone of productivity, specifically in the workplace, right? But to your point, lots of other consumer use cases. I mean, I grew up with a Windows PC in my house. It's cool to hear it. Yeah, so we were Windows family. So obviously I've used it for everything. How do you see it evolving into a more intelligent platform given those different layers that you have going there? I think we should probably start with a piece that you opened with. I think a lot of customers have an expectation for resilience and security as a foundation for trust for Windows.

2:42And when we started with Windows 11, in fact, one of our core principles was making sure Windows 11 mapped the secure future initiative work Microsoft as a company was doing to make sure secure by design, by default. And we built a lot of our production systems around making sure we could deliver on that base foundation for Windows being a great, resilient, secure platform. And that's, for us, an ongoing journey. And so we continue to stay invested on it. In fact, today at Ignite, we have a ton of work in that space that we're going to share broadly. So that's, I think, that's sort of the layer cake at the bottom.

3:10People expect the trust and capability of Windows as a platform. The exciting thing that we have when we think about productivity and unlocking productivity, one of our big learnings has been And most customers are familiar with the Windows UX today. They understand how to get their jobs done. It's the place where they're productive all day long. They use it for a variety of instances. And so a lot of our thinking has been on how do we have AI come into your workflow as you have it today? How can Copilot be of an assistance to you in the world that you're already operating in? The simplest example I can think of is the interaction layers in Windows.

3:43And so we've had mouse and keyboard. They've been stable for computing for many decades. and then we've added capabilities like pen and touch, for example, in the most recent decades. And now with AI, in terms of unlocking productivity, interacting with a new modality with voice and vision for Copilot, we think is particularly powerful and interesting for a set of users who want to be productive and multitasking and be able to find AI, you know, have it serve itself kind of when they're in the workflow. And so one of the big features that we talked about in October, and we have now available for commercial customers as well and Microsoft 365 customers, is the ability for Microsoft 365 Copilot to have voice support in Windows, voice capability in Windows.

4:22The other big one is the taskbar. Integration into the taskbar, it is a familiar place for Windows users. It's a place where just hundreds of millions of people are using the taskbar and its features on a daily basis. So ask Copilot in the taskbar is one of these places where we've gotten a strong signal for customers wanting to interact with the Copilot, just similar to how they do file search and photos and apps and settings today. it's a natural intuitive place for them to go. And now they have the ability to add natural language interaction. And we can direct that to a co-pilot interaction in addition to all the jobs that you do on Windows today.

4:54Totally done that. Yeah, where you basically pull up the bar, you ask a question, and then you have it direct you to where you need to go next. Yes, exactly. Pull up a search or pull up a program. Bingo, correct. And so now you can, exactly, you start with files, you start with settings, you start with apps. And then now we have the ability to add new capabilities like an interaction, a natural language query that gets redirected to the M365 Copilot, for instance. Yeah, yeah. How do you think of that in terms of how much you want to be done on device versus in the cloud? Like, how much of that is happening on device?

5:24It's a great question. And I think it is a world that is evolving rapidly and somewhat concurrently. In fact, part of me feels like that's sort of the magic for us is when we think about this in the long term in a Microsoft stack in this space, I do think computing is going to become hybrid. And so there's going to be a ton of compute that happens in the cloud. frontier models are there. Our large copilot experiences are served through that. And there's robust reasons for why that is a great value proposition in itself for customers. And then there's computing on the edge. And computing on the edge will evolve as well in my mind.

5:55In fact, one of the big step function changes we made in Windows was to be able to add neural processors into the Windows architecture. That was important. They're essentially highly accelerated compute for AI. And we want to be able to do that in an energy efficient manner so it doesn't consume your battery life, for instance. And so I think those investments of building a more robust platform on Windows on the edge will complement those use cases of how AI is built in the cloud. Certainly, we're seeing use cases where customers would like to be able to use AI on the edge. They have a lot to do with data locality.

6:25And so when you have infrastructure and data on the edge, it makes sense for us to run inference on the edge in those instances. There's privacy and performance and latency requirements that drive it. Our solution for that is a thing that we call the Microsoft Foundry on Windows. And Microsoft Foundry in Windows gives you the ability to bring your own models and run them on the edge. It allows you to run open source models on the edge. And in Windows, we also carry inbox models, surfaces APIs, and you can simply use those APIs powered by the Windows models in case you don't want, you know, you're not in a place to be able to bring your own models or use open source models.

6:58So let's say I'm chatting with Copilot and I'm asking a question like, hey, where's my, you know, PDF for this report that I'm working on? So is that me talking to Copilot over the API, or is that me talking to Copilot on the edge, or does it depend on the... Great question. Actually, that's a scenario that we're exploring with the Copilot team right now in terms of how we can connect dots between the Copilot experience in the cloud and what your file system could look like in Windows. Primarily, that is a cloud experience. One of the new innovations we have that we're actually sharing at Ignite is this notion of agent connectors in Windows.

7:29So agent connectors in Windows are going to be backed by an on-device registry from the Windows team. And in that specific example, in fact, we're going to have a File Explorer connector on Windows. Awesome. So yeah, that was a great question. So your model in the cloud can ask Windows for access to your file system. And in the case of Copilot Plus devices, you can also ask for semantic access to the system. And so it's not just looking for names and files and folders. the model can actually ask a query to the agent connector locally on Windows for, find me a file with a red barn door and retrieve a semantic result.

8:04Yeah, or like the use case I think of is my mom has hundreds of thousands of photos on her computer. And if she's like, give me all of the pumpkin patch pictures from every year, maybe it might not get all of them, but it could get a spread of that perhaps. Absolutely. I think photos for consumers, Microsoft Office documents for commercial customers are all great examples of why the file system in Windows is something that people are familiar with, that they trust, they want to be able to unlock in their experiences. And so that was the primary motivation for starting with the File Explorer connector, in fact.

8:40What principles guide your team when you're trying to decide what happens natively in the cloud? I guess you kind of talked about this a little bit, but from a design perspective, how are you thinking about it? We fundamentally start with what makes most sense for the end user scenario for the customer value proposition in itself. We're very much anchored in what can we do to deliver, from a Microsoft perspective, the best end-to-end solution. And we typically work backwards from there. The reason why I find that's a helpful way to look at it is because I find the answer will change over time.

9:13Your compute construct will change. The environment, which are the devices being used will change. Model capabilities. Exactly. Model capabilities will change. How these connectors, for example, will evolve on the edge as they will in the cloud. So we try to think about what is the end user instance and the value proposition in the use case, and we kind of work backwards from there, I would say. Okay, so a really good example of how the cloud copilots work with local AI capabilities are on Copilot Plus devices. On Copilot Plus devices, we have a capability in Windows called Click2Do. And Click2Do essentially gives you screen understanding under user control.

9:46And so if you're in an instance where you have a PDF document and you're looking at tables, files, formats that are really complicated, and you as a user want to be able to manipulate them, the local device capability gives you the ability to understand the screen a local model runs that takes that screen content and turns it into a tabular format for Excel. Then you have the power of M365 co-pilot. In Excel, you can fire up agent mode, be able to manipulate that thing, evolve it, graph it, understand it, run analytics on it. In my mind, it's a beautiful example of an end-to-end workflow where you start with essentially a static document.

10:20It is essentially a picture or a PDF document that is typically hard to manipulate and synthesize and draw insights from. And in a matter of two steps with using local AI and then an agent mode in Excel, you really get to an empowered state of being able to do much richer things that same data set. That's interesting. So one of the use cases that boggles my mind because it doesn't feel fixed yet is there's so much when you're writing with AI and you're using AI content, there's a lot of middle work where you're changing the formatting and this or that. I would love to be able to click to do and be like, just change all of the format of this, like, you know, remove all of the like the bullet points or add bullet points or just tell it to do it.

10:59So much of that is like manual now. And I feel like like I'm letting the AI do a lot of the fun part of writing. And then I'm in there doing the manual. That's a good note for us. We will take that back to our team. This is this is not how this is supposed to work. It's supposed to be the opposite of this. True. In fact, in at Ignite, we're actually talking about writing assistance on Windows as well. We have the M365 writing systems capabilities on these devices on Windows. but I love your note on, can we also take some of the drudgery out with file and formatting? I love it. Yeah, let's solve this.

11:29We can do it. We have the technology. The other use case I really like that I think if you're taking notes, I really would love for AI to be able to draw on my screen when I'm looking for like a button and be like, it's here, click here. You know what I'm talking about? Like in the same way that if I was to take a screenshot and send it to my friend and I would like draw on it and I'd be like, oh, you need to click. Like I'm trying to tell him how to navigate a really complicated UI or something, click this button here. I would love for AI to be able to take a screenshot of my screen, show me, here's what you need to do, click this, and then in this menu, click this.

12:04We might have something for you. We have a couple of capabilities that are in Copilot Labs and Windows Insiders right now. We call that feature Copilot Highlights, where you can choose to share your screen with Copilot, and Copilot uses an understanding of your interaction with Copilot and can highlight content on the screen to help you find hidden menus, find new features. So I'm going to send you a link after this and you should tell us what you think. I love it. It's a start, but we're getting great signal art, quite frankly, for the exact reason you mentioned. Yeah, I mean, I think that's where the edge use case beyond File Explorer really makes sense.

12:38Yes, exactly. Like now it's like, we're going to help you use a computer better. Correct, yes. Everyone's grandparents, right? They've needed the kids to be tech support. Now you can build tech support into the computer. Assistance for troubleshooting, assistance for in-workflow tasks and activities, and assistance for how to use your own computer are among the big signals we see in terms of queries in Copilot. And we as a Windows team are very much down the path of how can we take advantage of AI to simplify these workflows, make it more accessible for users, reduce friction. For the exact reason you described, we have a very broad user base in Windows.

13:12There's a set of customers who can really get supercharged with these capabilities when it comes naturally in their workflow. Yeah, I mean, that's one thing that I think is, you know, how can you improve accessibility with a lot of this? Like, voice UI, I feel like, is a really, really easy way, you know, not that it's easy, but it's a simple way to help with accessibility. How else are you thinking about accessibility with your design? Accessibility is a central tenant in how we think about Windows design and how we think about our features as we ship and build them for the breadth of our user base.

13:39In fact, we're also thinking about places where accessibility can be enriched with AI capabilities, both in terms of voice interaction pieces. One of the beautiful things with AI, like you mentioned, is being able to understand screen content. And so that example of sharing for navigating complex file systems or file menus in an app is one. But just being able to understand screen content for a user who could benefit from accessibility guidance where they're not navigating through complex UI that was not really built for that use case. is another great example where we're trying to have Narrator and Windows, for example, take advantage of being able to semantically understand screen content.

14:16In fact, local closed captioning in Windows today on Copilot Plus devices and live translation capabilities in Copilot Plus devices are using local AI models to be able to do these real-time on the fly. That is so cool. How do you ensure consistency and performance across the wide range of devices that you're overseeing here with us? That's a great question, especially in the world of AI. And I think in my mind, there's a couple of different ways to think about that space. One, there's a set of capabilities that we think customers will expect out of the Windows user experience on a frequent recurring enough basis to our conversation that we just want them integrated into the workflows and just kind of be a part of Windows as you think about them.

14:59There, we have a real responsibility for making sure performance, battery life, latency are all well within all other parameters of the operating system. So it feels like AI is a part of the device, part of your Windows experience, not a thing that got added on later, for instance. And so for there, a great example for us are these Copilot Plus devices where we've had the ability to work closely with our silicon partners for making sure they have the capability in the silicon platform. where we as a Windows platform are able to deliver certain guarantees to the end user because we can expect those base capabilities in the platform and in the use cases that show up in those devices and themselves.

15:36And so neural processors, memory bandwidth on the silicon themselves, expectations around system configurations for the hardware itself are all great ways for us for making sure any one of these devices has a certain guarantee of experience. The other end of the spectrum is where customers know and expect Windows devices to have a diversity of capabilities based on what the device can do for you. GPUs and gaming are a great example. You as a user know for AAA titles, you have to have a certain level of GPU card performance, a certain level of memory, GDR specs, and so on. And so we have a lot of developers, enthusiasts in the AI space in Windows for sure, and a certain class of customers as well who are okay with the ability to be able to take advantage of AI on the devices that they have right now and understand that you can get better AI performance as a platform gets better.

16:23And so we have customers who use GPUs today. And the beauty with the framework that we have built with Windows ML is Windows ML lets us take advantage of NPUs for these deep integrated Windows experiences where it makes sense. It fully allows us to take advantage of GPUs for either customers, app developers, AI enthusiasts, or developers themselves to be able to take advantage of GPUs for running local AI, knowing that the performance will vary based on the platform. Is there ever stuff that you develop that you can't ship yet because it can't fit all of the different devices? Like you have stuff on the frontier where you're like, man, we just need all of the surfaces to get this next generation and then we're good to go with this?

17:00It's a great question. I think the Shurder says yes. And that's the beauty of it at the end of the day. We feel like we have a pipeline of ideas, capabilities, new features. And we are pushing the animal for how can we find creative ways of delivering this value proposition and maintaining the value propositions that customers know and love today and expect out of Windows. That, I think, system engineering and making sure they're deeply integrated is the magic for us and kind of our role. And yes, so we have lots of work that's happening to make these models, have them be more efficient, have them be more performant, kind of update these models with newer models, like you mentioned, even in the edge space, kind of like we do in the frontier space.

17:38And I'm actually really bullish about the space in the collective because the models themselves are getting much better and much more capable. The runtimes and frameworks that we're using to execute these models are also getting better and more capable. The hardware platforms are getting more capable as we go forward. And then for us, it is opening up the surface area for how user experiences can be built and how the Windows developer platform can surface more value. So we're kind of excited because I think we're at the start of this runway. And I think about this in the years to come. We're just opening up the surface area of what's possible, I think.

18:08I have a slightly out there question for you. Do you think that we at some point are going to need to reinvent the computer for this AIH in some meaningful way? I know we're doing it while the plane is flying, but is there going to be some, looking at it from a system level, all of these things are improving each other at the same time. Is there a point where we need to take it in a different direction that really gets us to escape philosophy? It's a great question, something we as a Windows team ponder often. And, you know, I forget who exactly said this and how, but I think, you know, there's this thing around the future is already here.

18:44It's just not uniformly distributed. And so we definitely see that with a lot of our customers today. We have a continuum of customers, some of whom are in their current use cases, know and love Windows for those use cases and absolutely celebrate them. And then we are learning as we watch a set of customers who are now using Windows in ways that we ourselves are coming to appreciate. and we're in a fast, closed loop with them to appreciate how they're building and using Windows and places we can evolve the platform. The agentic evolution is one good example there where the uses of agents in an enterprise instance, for example, was a thing that, in my mind, is happening faster than some expected.

19:19And it is a great thing for us because it's an opportunity for us to build platform primitives, user experience capabilities, integration capabilities, Windows to serve them. And we get great signal for them, to your point, on where does Windows go in the future in itself? And I think for us, it is our responsibility to kind of do both pieces. We have to walk and chew gum. And so we have to be able to make sure Windows is great for our customers today as they know and love it. And then it's our opportunity to find a way to bring these new capabilities in ways that make sense to them across the diversity of our user base.

19:47So there's this concept of AI fatigue, right? Yeah. where everybody, there's a certain subset of people probably in the Windows community, just like the larger society that is tired of AI, maybe doesn't want AI in their Microsoft Word or their Windows. How do you balance the powerful automation and the transparency that, you know, what you're explaining where you're using AI and, you know, maybe even making an option to turn it off when people don't want it? Yeah, you nailed it, actually. I think for us, the reality is we have a breadth of customers, which by definition means different customers are in different places in terms of how they think about AI and where they are in the adoption of the product and the technology in itself.

20:30I think it's our responsibility in Windows for making that be accessible to customers at the rate at which they find it valuable and useful for themselves. And so we do have a set of principles that we use in terms of how we product make and how we actually roll out these capabilities. The first was exactly what you just mentioned, which is we think deeply about making these core features opt-in for customers. We have mechanisms and listening systems in our products and with our customers for making sure, one, we have good signal from them in terms of where they are on the journey in itself. For sure, make sure we start with a set of opt-in capabilities.

20:59So it is a thing that customers can use, try, learn, experiment, both give us feedback as well as learn where it adds value in their workflow. The second big construct for us is making sure these experiences are secure. And so we ourselves feel like we have a responsibility for making sure the Windows platform is secure in itself. And then we're creating an environment where others can build on top of the platform with primitives that are also inherently secure. A lot of the work we're doing, in fact, in the agent space, for example, is to introduce this notion of an agent workspace. And it's to be able to give people or give developers and customers an environment that they can monitor, observe, where there's parameters and guardrails for security, for instance, and corporate governance for policy.

21:42And so we're bringing new concepts to the table that allows security, compliance, and governance to be base constructs in how the platform parameters are being created. In my mind, that makes them a step function more secure, more robust than even how apps are built today, for instance, in some places. So that's the second big element for us. The third one is this notion of user consent. And to your point on privacy, even with the use of agent connectors, for example, we are very much grounded with agent connectors and use of the Windows on-device registry is a capability that we expose through user consent.

22:14And so users are aware of what is happening. They're in control. They can choose to continue or not continue with a certain task. And it's the same thing for Windows 11 today, for client computing and for Windows 365 in the cloud. In Windows 365 in the cloud, you can imagine the entire cloud PC is a session for an agent to use for computer use automation, for building tooling for another task, for infrastructure. And we want to be able to make sure that Windows 365 instance is a thing that, you know, enterprise customers, IT and policy and governance can completely audit, have complete visibility around, there's true, you know, logs for it.

22:48And so we think about these concepts, you know, fundamentally how we build a product, but also how we roll them out to our customers as well. Yeah, I mean, that's going to be especially big as you're working with MCP, right? Like now that you've got MCP local on the device, you know, that's, you don't want anyone to be able to come in and, you know. That's actually one of the important parts exactly with the MCP is the Windows on-device registry needs to be a place that end users can trust as a set of connectors that are in the registry that they see as being trustworthy and reliable and can use.