Microsoft Design: building an AI-forward design system for Copilot

Microsoft’s head of design Jon Friedman describes how the company approached building the Copilot Design System — not by grafting AI onto existing interfaces, but by working from a more fundamental question: what does an AI assistant need from a design system to feel like an extension of how people actually think?

The system rests on four components. The Dynamic Action Button (DAB) functions as a contextually adaptive entry point, appearing where it makes sense in a given workflow rather than occupying a fixed corner of every screen. Chat handles the main exchange of reasoning and output between user and AI. On-Canvas is a lightweight surface for direct interaction with content already in front of the user. Suggested User Actions surfaces timely prompts based on what the user is currently doing, without demanding attention they haven’t offered.

The piece that holds these together is what Friedman calls the “throw and catch” pattern: Copilot can move fluidly between the four surfaces, handing work from one to another without requiring the user to re-establish context. The design intent is that Copilot behaves as “an extension of your thinking” rather than a separate application you have to remember to consult.

The central argument is that intelligence without continuity creates interruption rather than assistance. A chat interface that doesn’t remember where you were, or a suggestion that appears after you’ve already moved on, undermines trust faster than any slow response time. Continuity across surfaces is what distinguishes a competent AI from one that merely resembles one.

Friedman also notes the governing principle behind the team’s pace: they built “at the speed of life, not AI hype.” That phrase captures something worth thinking about carefully. Design systems need to outlast feature announcements, and the structural decisions made here — which surfaces to use, how context passes between them — will shape billions of product interactions for years after any individual model update.

Useful for teams designing AI features into existing products who need a framework for thinking about context persistence and how to choose between competing interface surfaces.