Microsoft’s new strategy for Windows 11 does not indicate a retreat from artificial intelligence, but rather a change in how it appears within the operating system. Instead of adding Copilot buttons and separate functions to every part of Windows, the company wants the system to become a platform capable of running AI models and agents directly on users’ devices, while retaining cloud services for tasks that require greater capabilities.
This direction emerged in statements attributed to Satya Nadella, Microsoft’s CEO, and Mark Linton, the company’s executive vice president for Windows and devices, during the IFA 2026 event. Microsoft used the expression “unmetered intelligence” to describe its vision of using AI without permanently relying on cloud inference calculated according to usage.
From a Copilot Button to an Intelligence Layer Within the System
According to the source material, refocusing on Windows’ stability and quality does not mean that Microsoft will abandon AI. The proposed change is to reduce reliance on presenting it as a button or a visible product for users, and instead make it part of the infrastructure that allows applications and agents to carry out their tasks within the computer.
Examples cited by the article include removing Copilot from Microsoft Photos and Snipping Tool, and renaming the Copilot integration in Notepad to “Writing tools.” These changes alone do not prove a complete technical shift, but they illustrate a trend in naming and presenting functions to users, away from presenting every feature as an independent Copilot.
What Is Changing in Practice?
Microsoft believes that high-performance computers can run more AI models on the CPU, GPU, or NPU instead of sending every request to data centers. According to the statements cited, this includes some everyday tasks and smaller interactions, while the cloud remains an important part of the ecosystem, and local execution is not being presented as a complete replacement for it.
The intended economic impact is connected to the cost of tokens, or tokens. If the model runs on the user’s device, the request itself will not necessarily need to consume tokens in a cloud service or require a subscription tied to each inference operation. However, describing usage as “unmetered” does not mean that Copilot will become free without limits; it refers to a difference in where execution takes place and how inference is calculated. Local models also remain tied to the device’s capabilities and the resources available on it.
Windows as a Platform for Agents
Microsoft links this vision to the shift from chatbots to agents capable of carrying out longer and more autonomous tasks. According to the article, the company discussed computers capable of running models containing tens of billions, and even hundreds of billions, of parameters, as well as the goal of making Windows the “best platform for AI.”
One stated priority is improving the developer experience through a simpler environment for building applications and agents. Microsoft also referred to the developer environment and ongoing work to develop tools that allow software to be created from within Windows, without the article providing complete technical details about these tools or their release date.
Control and Security Remain Open Questions
The most important practical limitation in this direction is that a local agent should not operate freely within the user’s device or an organization’s network. The article quoted Linton on companies’ need for security and manageability while preserving existing network policies and rules.
For this reason, Microsoft is discussing the ability to audit agents’ behavior, review the logs they produce, and control more precisely the operations they are permitted to perform. According to the source, the company is also working on containers and subsystems dedicated to AI, with the aim of preventing agents from affecting personal files or carrying out operations beyond control.
certi.news analysis: The actual change here is not the disappearance of Copilot, but the shift in discussion from a single assistant interface to an operating architecture that combines local hardware, the cloud, and agents. The success of this model will depend on Windows devices’ ability to run useful models efficiently, as well as Microsoft’s ability to provide clear auditing and management tools. Until the company announces additional details later in the year, the limits of the models that will run locally, hardware requirements, and the mechanism for managing “tokens” locally remain unresolved questions in the available material.