- Microsoft is combining on-device and cloud-based AI models to handle different tasks.
- Copilot will be able to find, organize and work with files and perform actions across Windows.
- Microsoft says local AI agents will operate inside permission-controlled sandboxes designed to limit what they can access.
Microsoft is pushing artificial intelligence deeper into Windows with a new approach called “Hybrid Intelligence”, giving Copilot greater ability to work with files, apps and settings on a PC.
The company says the system will combine local AI models running on a Windows device with cloud-based models, allowing Windows to decide where a task should be handled.
That could mean simpler or privacy-sensitive requests are processed locally, while more demanding tasks are sent to cloud models.
The change moves Copilot beyond a traditional chatbot that responds to questions. Microsoft is positioning it as an agent that can find information, organize files and take actions on a user's behalf.
What can Microsoft's Copilot now do on Windows?
Microsoft demonstrated Copilot working with files and everyday administrative tasks.
In one example, Copilot could find and organize a user's tax documents, then draft an email to an accountant with the relevant files attached. The broader idea is that users will be able to ask Copilot to complete multi-step tasks rather than manually opening folders, searching for documents and moving information between applications.
Microsoft also showed AI-enhanced Windows search that can help users find settings using natural language. Instead of navigating through menus, users could ask Windows to locate a setting such as dark mode.
The operating system is also getting a conversational interface for tasks such as sending a text message.
For businesses, the changes could be particularly significant. Employees regularly spend time searching for documents, extracting information from files, preparing emails and moving data between applications. An agent capable of handling those steps could reduce some of that administrative work.
But the technology also raises a more important question: how much control should an AI system have over a computer?
Microsoft says Copilot's agentic features will not have unrestricted access to an entire file system. Instead, the system will use integrations with Windows and Microsoft applications to carry out specific tasks.
That distinction matters because an AI that can act is potentially much more powerful than one that simply generates text.
What is Microsoft's ‘Hybrid Intelligence’?
Microsoft's approach is designed to split AI workloads between the PC and the cloud.
The company's HydraFusion model router can determine whether a task should use a local model or a cloud-based model. Local processing could be useful for relatively simple or privacy-sensitive tasks, while more complex requests could be handled by larger cloud models.
Microsoft said cloud-based options can include models from OpenAI and Anthropic, depending on the task and the software involved.
The company is also expanding Windows ML, its machine-learning inference framework based on ONNX Runtime, to support Llama.cpp, an open-source inference engine used for running AI models locally.
Microsoft says local agents will run inside Microsoft Execution Containers (MXCs), which provide sandboxed environments and fine-grained permissions that can be applied while an agent is operating.
Pavan Davuluri, Microsoft's executive vice president for Windows and Devices, described the approach as making Windows “the most secure platform for agents.”
Keeping some AI workloads on the device could also reduce the amount of information that needs to be sent to Microsoft's servers. It may lower cloud computing demands for simpler tasks, although running AI models locally can place additional demands on a PC's processor, battery and other hardware.
Microsoft is initially rolling out the hybrid intelligence capabilities to GitHub Copilot, its coding assistant, with the Copilot app expected to receive the functionality over the following months.
Is giving AI access to your files safe?
That will depend heavily on how Microsoft implements permissions and how users configure them.
The company says its sandboxing and permission controls are designed to restrict what agents can access and do. Microsoft also says Copilot will not simply search through personal files without the user asking it to perform a task that requires that information.
But the shift represents a significant change in the relationship between users and desktop AI.
A chatbot that generates an answer has limited ability to affect the outside world. An agent that can search files, attach documents to emails, change settings or perform actions across applications has a much larger potential impact if something goes wrong.
That makes permission controls, audit trails and clear user approval particularly important, especially for businesses handling financial, customer or confidential information.
For enterprise users, the technology could eventually make document-heavy work such as accounting, tax preparation and administration considerably faster. But companies will need clear rules about what Copilot can access, which actions require approval and how those actions are recorded.
Microsoft's bet is that the future of Windows will not simply involve AI answering questions. It will involve AI doing things on the user's behalf, with the PC and the cloud sharing the workload.
The success of that approach will ultimately depend on whether Microsoft can make those actions useful without making users feel they have given an AI too much control over their computers.
