Microsoft pulls copilot from Windows apps and redesigns it as an 'OS for work'

Microsoft pulled Copilot out of several Windows apps earlier this year, a quiet admission that bolting a chatbot onto every menu bar doesn't automatically make it useful. The company is now betting on a different approach: one Copilot, built around three capabilities called Home, Code, and Autopilot, pitched as an "OS for work" where chatting, building apps, and running agents all happen in the same place.

Jacob Andreou, Microsoft's EVP of Copilot, told The Rundown that the earlier Windows app rollout looked good on traffic dashboards but failed to drive real usage. That gap between surface-level metrics and actual adoption now shapes how he ships AI products. He expects the same pattern to repeat this fall when Microsoft pushes more personal agents to consumers.

The model-switching strategy Microsoft won't advertise

Andreou said something most vendors avoid stating plainly: no single model wins everywhere. Copilot's job is to pick the right model per task rather than bet on one lab. If OpenAI's models handle Excel-style logic better and Anthropic's handle visual and code consequences better, the smart architecture is modular - routing logic that swaps models by task instead of locking into one API.

For developers and founders, this signals that model choice shouldn't be a one-time decision baked into a product. For researchers, it confirms that benchmark leaderboards are losing value as a proxy for real-world fit. "Best overall" means nothing once you get specific about the job. Microsoft has the resources to build that abstraction layer at scale through Copilot. Smaller teams building AI products will need to decide whether to build similar routing themselves or wait for platforms like this to do it for them.

How much autonomy before a human steps in

The conversation covered how much independence an agent should get before a human needs to intervene, why a smarter model alone doesn't solve the product problem, and what AI-native work will look like once the current wave of features settles. Andreou tied the product direction to a broader concern about who captures the value from AI advances.

"If we don't build products that translate raw intelligence into real value for real people, intelligence will centralize in the hands of a couple of companies, and value will centralize along those same lines," Andreou said. "That's what we're trying to solve with Copilot."

Whether Microsoft ever names which tasks go to which model - or keeps that routing logic under the hood - remains an open question worth watching. The company's AI for Product Managers Courses reflect a similar modular approach to building AI products, though the Copilot team is operating at a scale few organizations can match.

Why this matters for product development and IT teams

Microsoft's shift confirms that the market is moving past the "pick one model" era. Product teams should architect for model flexibility from the start rather than treating an API choice as permanent infrastructure. For IT leaders, the implication is clear: evaluate AI platforms on their routing and orchestration capabilities, not just which foundational model sits underneath. The abstraction layer is becoming the product.