A year ago, Microsoft pitched Copilot+ PCs as a new class of computers where on-device NPUs (neural processing units) would run powerful AI models locally, delivering private, instant experiences that would change how you use Windows. Today, after months of hands-on testing and slow rollouts, the reality is more mundane: a handful of useful on-device capabilities, a cloud-dependent feature set, and a branding mess that leaves many buyers underwhelmed. The hardware is excellent—class-leading battery life and snappy performance are real—but the “AI revolution” Microsoft promised is partial at best.

The Promise: Instant, Private, Local AI

At the heart of Microsoft’s Copilot+ PC vision was the idea of offloading AI tasks to a dedicated NPU. Small language models (SLMs) and vision models would handle things like natural-language search, content summarization, and task automation right on your device. No cloud round-trips, no privacy concerns about sensitive data leaving your machine. The marketing leaned heavily into this on-device narrative, and OEMs rushed to ship ARM-based Surface devices and x86 Copilot+ variants, all emphasizing efficiency and AI capability.

The Reality: A Mixed Bag of Delays and Cloud Dependency

The actual shipped feature set tells a different story. While some features do run locally and feel genuinely useful, many of the most hyped capabilities either arrived late, were scaled back, or still require a Microsoft account and cloud connectivity.

  • The Settings agent powered by Mu is the standout on-device success. This feature, available on Windows 11 version 24H2 with KB5062660 and only on Copilot+ PCs, uses a compact language model called Mu to understand natural-language queries like “My mouse pointer is too small” and directly suggest the relevant setting. The model runs entirely on the NPU, so responses are instantaneous and private. It’s the clearest example of how local AI can eliminate friction without sacrificing privacy.
  • Recall, the feature that captures desktop snapshots to let you search your past activity, was initially delayed for months after security researchers discovered the unencrypted database. Microsoft added encryption, mandated Windows Hello, and made it opt-in, but the reputational damage was done. Several apps and browsers, including Brave and AdGuard, moved to block or restrict Recall’s functionality. Even with mitigations, a screen-recording feature demands ironclad trust that takes years to build.
  • Click to Do overlays contextual actions on selected content (e.g., summarize, search, email). Microsoft says the analysis happens locally on Copilot+ devices, but many of its richer suggestions require cloud services—undercutting the on-device-only promise.
  • Generative features in Photos and Designer are cloud-first: image generation and editing rely on server-side models like DALL·E, requiring a Microsoft account and, in many cases, a Copilot Pro subscription for premium capabilities.

Deep Dive: The Features That Matter (and Why They Underdeliver)

Windows Recall: Powerful, Polarizing

Conceptually, Recall is compelling: a local timeline index that lets you find anything you’ve seen. In practice, the initial implementation stored screenshots in plain text, leading to widespread alarm. Microsoft’s later security overhaul—encrypted snapshots, biometric verification, filtering—was technically sound, but trust is harder to restore than code. For many users, Recall remains disabled, and for security-conscious enterprises, it’s a nonstarter.

Click to Do: Clever, but Limited

Click to Do is a productivity tool that performs local analysis of your selected content and offers actions. While it’s fast and keeps sensitive snippets on-device during analysis, the most powerful actions (like generating a detailed summary or drafting an email) often trigger a cloud call. This hybrid model makes sense given the limitations of on-device models, but it blurs the line between local processing and cloud dependence—exactly the line that Copilot+ marketing tried to draw.

Settings Agent and Mu: The Real On-Device Hero

The agent in Windows Settings is the closest thing to the Copilot+ vision done right. The Mu model, detailed in Microsoft’s documentation, is a quantized SLM tuned specifically for Settings metadata. It maps natural language to concrete Windows APIs, offering a one-click fix. Because it runs entirely on the NPU, there’s zero latency and zero privacy risk. The feature is available in English, French, German, Hindi, Italian, Japanese, Korean, Portuguese, Spanish, and Chinese (Simplified), and it works in all countries except China.

But Mu’s scope is narrow: it can’t help you with general questions, summarize documents, or perform complex reasoning. It’s a brilliant micro-optimization, not a general AI assistant. The gap between this focused agent and the broader Copilot experience highlights the whole Copilot+ problem: on-device NPUs excel at tightly defined, low-latency tasks, but the AI everyone expects still lives in the cloud.

Photos, Designer, Image Creator: Cloud-First Creativity

Microsoft’s integration of generative AI into Photos and the Designer tool sounds impressive, but the heavy lifting happens on Azure. Image generation uses DALL·E models; edits require cloud processing. A Copilot+ PC can certainly run these features, but so can any Windows 11 machine with an internet connection. That makes them a poor differentiator for the Copilot+ hardware.

NPU vs. GPU vs. Cloud: Who Does What Best?

Understanding the technical landscape explains why Copilot+ PCs haven’t revolutionized AI workloads.

  • NPUs are optimized for sustained, low-power inference on small models. They can run SLMs like Phi Silica or Mu at a fraction of the energy cost of a CPU or GPU, making them ideal for always-on, privacy-sensitive agents.
  • GPUs remain the workhorses for large generative models and high-throughput inference. NVIDIA’s Tensor cores and AMD’s compute units deliver the memory bandwidth and parallel processing that NPUs can’t match. Tools like Nvidia Broadcast (real-time background blur, noise removal) require an RTX GPU and won’t work on an NPU alone.
  • Cloud remains where the biggest models run. Adobe’s Firefly, GitHub Copilot’s advanced models, and most image generators are impractical to run locally on any current laptop hardware.

The takeaway: NPUs add value, but they don’t replace GPUs or the cloud for heavy AI. For a laptop NPU to be a selling point, there must be a compelling set of NPU-exclusive features that meaningfully improve the experience. Right now, that list is short.

The Ecosystem Gap: Third-Party Apps Still Aren’t on Board

A hardware accelerator only matters if software uses it. And the Copilot+ ecosystem has a chicken-and-egg problem.

  • Blackmagic Design demoed DaVinci Resolve running on Snapdragon X Elite NPUs, but most creative professionals still rely on GPU-accelerated workflows or cloud processing.
  • Topaz Labs explicitly stated in community forums that its AI tools don’t use NPUs and instead leverage GPUs and CPUs.
  • Microsoft’s own developer tooling for NPU targeting, while improving, hasn’t yet spurred a wave of third-party adoption. The Windows Copilot Runtime and APIs exist, but broad developer engagement lags.

Until apps like Adobe Premiere Pro, Blender, OBS Studio, or Zoom can offload specific AI tasks to the NPU—and until those tasks are more than gimmicks—the NPU will remain a checkmark on a spec sheet, not a daily differentiator.

The Branding Mess: Copilot, Copilot+, Copilot Pro

Microsoft’s naming strategy compounds the problem. “Copilot” now refers to the cloud assistant, “Copilot+” is the hardware tier, and “Copilot Pro” is a $20/month subscription for premium features. Confusion is inevitable. A buyer might reasonably expect that buying a Copilot+ PC grants full, unlimited access to Copilot’s best capabilities. Instead, they get a machine that can run some AI locally, but still needs a subscription to unlock the full cloud assistant and productivity features.

This disconnect has real-world consequences: customers feel nickel-and-dimed. A smarter move would be bundling a year of Copilot Pro with high-end Copilot+ purchases, similar to how phone vendors often include cloud storage or premium services. It would build goodwill while the on-device ecosystem matures.

What Copilot+ PCs Actually Get Right

Despite the AI shortcomings, the Copilot+ hardware is genuinely impressive.

  • Battery life and efficiency: Snapdragon X Elite and the latest Intel/AMD chips with NPUs deliver class-leading battery endurance and quiet operation. Reviews consistently praise 15+ hours of real-world use, a tangible upgrade for mobile workers.
  • Privacy-first engineering, where applied: When Microsoft commits to local execution—as with the Settings agent and the secure enclaves for Recall—the design is sound. A fully local AI assistant that never phones home is the right architecture; it just needs more features and a trust rebuild.
  • Meaningful on-device UX wins: The Settings agent shows how a small model can eliminate friction in everyday tasks. These micro-interactions, scaled across the OS, could genuinely change how we interact with Windows.

Risks and Unanswered Questions

  • Security surface area: Features that capture or analyze screen content expand the attack surface. Even with Microsoft’s mitigations, Recall’s existence will make some users uncomfortable. Any new sensor-heavy AI feature must be designed with extreme caution.
  • Developer adoption: Without a critical mass of NPU-accelerated apps, the hardware goes underused. Microsoft needs better SDKs, sample apps, and perhaps even financial incentives to nudge developers.
  • Economic model misalignment: Charging separately for Copilot Pro while marketing Copilot+ as the AI PC is a tough sell. The value proposition only works if on-device features become so compelling that users can reduce their reliance on cloud subscriptions.

What Microsoft Should Do Next

  1. Recenter messaging: Sell Copilot+ PCs on battery life, efficiency, and specific on-device wins—not on a vague AI transformation. Use the Settings agent as a demo, not a promise.
  2. Accelerate third-party NPU support: Ship polished NPU SDKs, clear documentation, and performance benchmarks. Offer a certification or badge program for apps that truly run on-device.
  3. Bundle Copilot Pro temporarily: Include a one-year subscription with high-end Copilot+ purchases to demonstrate the full Copilot experience and build user habits.
  4. Build trust visibly: Publish independent security audits, clearly show when on-device indexing is active, and make all AI features opt-in by default.
  5. Broaden hardware compatibility: Ensure that NPU-driven features don’t arbitrarily exclude x86 or older GPUs. If a feature requires an NPU, offer a GPU-accelerated fallback where possible to keep the ecosystem inclusive.

The Bottom Line

Copilot+ PCs are not a failure—they are excellent laptops that happen to have an underutilized AI accelerator. The Settings agent proves that on-device intelligence can work elegantly; the hardware proves that efficiency gains are real. But the “AI magic” Microsoft marketed is still largely in the cloud, subscription-gated, or delayed. For now, buy a Copilot+ PC for its battery life and build quality, not for an on-device AI revolution. That revolution may still come, but it will require better execution, wider developer support, and a lot more trust.