The August 2025 Windows 11 update shipped a quiet but transformative feature: a natural-language agent inside the Settings app that can understand commands like "make my mouse pointer bigger" and execute the change directly. The agent is powered by a brand-new, on-device model called Mu, and it marks Microsoft's first concrete deployment of the context-aware, multimodal AI vision that executives have been outlining for over a year. There is, however, a significant catch—the agent is available exclusively on Copilot+ PCs, machines equipped with a neural processing unit (NPU) capable of at least 40 trillion operations per second (TOPS). It is the clearest signal yet that Microsoft intends to use premium hardware as the gatekeeper for the most advanced Windows experiences.
Pavan Davuluri, the corporate vice president in charge of Windows and Devices, revealed the strategy in an internal interview published just one week after the Windows team posted its first public "vision" video. "The interaction experiences on Windows are going to change, the business models are going to change, and the experiences are going to change in the next coming couple of decades," Davuluri said. He described an operating system that becomes "authentic and multimodal, with voice and vision and pen and touch, just like we use a mouse and keyboard." The interview, along with Microsoft's engineering blog posts and the August 2025 cumulative update, paints a detailed picture of how the company plans to embed AI so deeply into Windows that the PC can see your screen, understand your intent, and act on your behalf—all while running models locally to preserve latency and privacy.
The Mu Model: A Tiny Engine for System-Wide AI
The centerpiece of the recent update is Mu, a small language model (SLM) designed from scratch for on-device deployment. Microsoft's engineering team revealed that Mu uses an encoder–decoder architecture, contains only a few hundred million parameters, and has been heavily optimized through quantization and hardware-specific tuning. On Copilot+ PCs, Mu responds at over 100 tokens per second running entirely on the NPU. That speed is critical for a feature that must feel instantaneous; users aren't going to wait for a cloud round-trip when they just want to tweak a setting.
Mu was trained specifically to map natural-language queries to Windows system function calls. It is the brain behind the new Settings agent, which can interpret ambiguous requests like "fix my trackpad sensitivity" or "turn off notifications during presentations" and perform the corresponding actions without the user needing to navigate through nested menus. Microsoft distilled Mu from larger Phi models, leveraging the company's broader investment in the Phi-4 family—models that handle heavier reasoning and multimodal tasks across text, code, and vision. While Phi variants can run on-device on powerful silicon or in the cloud, Mu is deliberately compact and task-specific, a clear example of Microsoft's hybrid AI architecture.
"Settings is a representation of over 25 years of Windows engineering work, and now we have an agent in settings powered by Mu that allows people to use natural language interaction to complete a task," Davuluri said. "You can just interact with Settings and then ask, hey, fix this mouse cursor size for me. And it just happens. It's really cool."
Not Just Settings: The Broader Multimodal Push
The Settings agent is only one piece of a larger context-aware suite that Microsoft is rolling out. Davuluri highlighted two other capabilities already in preview for Copilot+ PCs: an improved semantic search that augments the traditional lexical indexer, and Click to Do, a feature that analyzes screen content and suggests context-appropriate actions. "The beauty with the semantic indexer is it has the capability to understand content that is being searched in addition to keyword search," he explained. Search becomes capable of finding documents by meaning, not just by matching strings. Click to Do, meanwhile, can detect when a user is working in an app and offer to complete tasks, reducing friction across the OS.
These features are underpinned by a multimodal AI stack that can accept voice, touch, pen, and vision input simultaneously. In the future, Davuluri expects the computer to "look at your screen and be context aware" as a primary interaction modality. The vision includes scenarios where you speak to your PC while writing or inking, and the system tracks the whole context to infer intent. All of this depends on having capable on-device models like Mu and the Phi-4-mini and Phi-4-multimodal versions, plus the necessary NPU hardware to run them without crushing battery life.
The Hardware Gate: What Makes a Copilot+ PC
Microsoft defines a Copilot+ PC as any device with an NPU delivering 40+ TOPS, alongside a modern CPU and GPU. The NPU is mandatory for the full Copilot+ feature set; without it, the Settings agent, Click to Do, improved Recall, AI-assisted Paint object selection, and other advanced tools either degrade to cloud-dependent versions or disappear entirely. Initially, Copilot+ was tied to Qualcomm's Snapdragon X Series, but Microsoft has now expanded support to Intel Core Ultra 200V series and AMD Ryzen AI 300 series chips. However, even within those families, the 40 TOPS threshold creates a hardware tier: many existing laptops and desktops, including those that meet Windows 11's baseline CPU and TPM requirements, fall short.
The August 2025 update (part of the 24H2 cycle) delivered these AI experiences first to Snapdragon-powered Copilot+ devices in the Insider channels, with Intel and AMD systems following later. Non-Copilot+ machines received the usual security patches and minor UI polish but none of the on-device AI features. This staged rollout is deliberate; Microsoft's documentation signals that it intends to preserve the premium perception of Copilot+ hardware, and to gather telemetry and user feedback in controlled waves before broader deployment.
Strategic Chess Moves: Windows 10 EOL and Platform Lock-In
Why is Microsoft gating its most compelling innovations behind new hardware? The answer is part engineering, part business. Running inference at 100 tokens per second on battery without melting the device genuinely requires NPU accelerators; offloading the same work to a CPU or GPU would be slower and more power-hungry. There are also legitimate privacy arguments: keeping sensitive data like screen content and personal files local avoids cloud exposure. These are valid design choices.
But the timing aligns with a massive commercial opportunity. Windows 10 end of support arrives on October 14, 2025, after which security updates stop flowing. Enterprises and consumers who want to remain supported must migrate to Windows 11, and now Microsoft can dangle advanced AI features as an extra incentive to buy new Copilot+ PCs rather than simply upgrading an old machine. The company has not announced a "Windows 12," and Davuluri carefully avoided any mention of a new numbered OS, instead emphasizing continuous evolution inside Windows 11. This makes the hardware tier even more important as a differentiator.
Community Reaction: Excitement Tempered by Fragmentation Fears
The Windows enthusiast community has responded with a mix of enthusiasm and caution, as evident in forums and technical discussions. Early testers praise the Settings agent's speed and accuracy; the ability to speak or type a casual request and have the system comply instantly feels like a genuine leap forward. Developers are eyeing new Edge APIs that let web apps call local Phi-style models, opening a path to low-latency AI in browsers. Enterprise IT admins, however, are raising hard questions. Features like Recall, which periodically snapshots screen content to enable later semantic search, demand rigorous privacy reviews. Microsoft says on-device indexing stays local and offers opt-in controls, but the telemetry footprint remains ambiguous—what metadata, if any, is shared with Microsoft?
There is also frustration over the two-tier experience. "Locking the most compelling features to Copilot+ hardware risks fragmentation," one IT manager noted. "We have thousands of perfectly capable PCs that won't see these tools, and budget cycles don't align with Microsoft's release cadence." The branding confusion doesn't help: "Copilot" is now an AI companion on the taskbar, "Microsoft 365 Copilot" is an enterprise productivity suite, and "Copilot+ PC" is a hardware certification. Users struggle to understand which one covers what.
Security researchers have flagged the expanded attack surface. An agent that can execute system changes on a user's behalf must be tightly sandboxed; otherwise, a compromised model could become a powerful persistence mechanism. Microsoft has published sandboxing details for Mu and the Phi runtime, but independent audits are still pending.
Practical Steps for IT and Developers
For organizations, the path forward requires deliberate planning:
- Inventory and assess: Identify which existing devices meet Copilot+ hardware specs (40+ TOPS NPU). Most pre-2024 machines will not qualify, so hardware refresh timelines must account for AI feature needs.
- Policy and privacy: Before enabling Recall or any screen-aware feature, review data handling and compliance implications. Microsoft provides Group Policy and MDM controls to disable these features centrally, which many regulated industries will likely use by default.
- Pilot programs: Run Copilot+ devices in controlled pilot groups to validate feature behavior with line-of-business applications and to gauge help desk impact. Start with Settings agent and semantic search, which are less invasive than Recall.
- Developer exploration: The new Edge APIs and on-device model access allow web applications to tap local inference for tasks like real-time translation or content generation without round-tripping to Azure. Test these in Edge Canary/Dev channels using Phi-4-mini models, but design graceful fallbacks for non-Copilot+ hardware.
A World of Ambient Computing—With Strings Attached
Microsoft's vision is technically impressive. A Windows that sees your screen, hears your voice, and anticipates your workflow could genuinely reshape productivity. The Settings agent is a convincing proof point: it solves a real pain point (the ever-morphing settings UI) with a natural interface. Semantic search and Click to Do promise to cut through the application silos that have defined PC work for decades.
Yet the hardware gating forces a trade-off. Consumers and businesses that skip the Copilot+ investment will watch new capabilities pass them by, while those who buy in become part of a grand experiment in ambient computing. Microsoft says the models will get smaller and more efficient over time, potentially running on less exotic silicon, but for now the line is sharp. The company's advice to AI skeptics—"try it"—is reasonable, but only if you're holding the right device.
The next 18 months will test whether on-device AI is a premium gimmick or a foundational shift. Microsoft has placed its bet on local inference, multimodal input, and continuous, hardware-tethered updates. Success depends not only on engineering excellence but on transparent privacy practices, equitable access, and clear communication. For Windows users, the future has arrived—but it's sitting behind a Copilot+ badge.