Copilot and Agents

Engage brings together company announcements, questions, communities of practice, CEO town halls, and help desk requests. It is both a professional platform and a social space, which created two problems for us to solve.

People often hesitated to write for a broad, formal audience, even when they had something useful to share. At the same time, knowledge was siloed in communities, which meant certain users could miss helpful information simply because they were not part of the community where it was shared. This revealed two opportunities for Copilot: help people transform their thoughts into material they felt confident sharing widely, and assist them in surfacing knowledge from communities. In these early days, there were few established patterns for AI in products. Basic questions remained on the table: Should the interface be a chat? How should answers signal their limitations? What role should citations play?

So we started by making our assumptions explicit, documenting what we believed and where each assumption came from:

Then, drawing on that research and the emerging direction from Microsoft’s Responsible AI initiative, I led the team through a pair of workshops to cement the non-negotiables:

The first problem was hesitation. People had ideas and experiences worth sharing but were reluctant to translate them into posts to share across the entire company. An assistant that simply writes those posts would remove the friction, but also the person. We wanted to prevent a feed full of generated posts that nobody would trust or read.

To address this I designed Copilot in Engage to write alongside the person, not for them. It opens beside the composer, drafts a post, cites its sources, and then stops. Nothing reaches the post until the person presses Add to post. A plain note flags that AI-generated content may be incorrect, and thumbs up and down provide feedback. These may look like obvious guardrails now, but in early 2023 they were choices we had to argue for. The follow-up prompts reinforce the principle. Rather than offering to write another draft, they return the writer to their own experience: How can I add a personal example? The goal is to help someone get past the blank composer while keeping their perspective and judgment at the center.

The second problem, siloed knowledge, called for a different approach. Catch Up summarizes what someone missed in their communities and links back to the original threads. It gives people a way into conversations they might otherwise never find.

I led design for Copilot through its general availability release in Viva Engage, including team dogfooding sessions, integration alignment with Teams, and a holistic inclusive design pass. We achieved a 59% conversion measurement of Copilot chats resulting in a platform-based action, like posting, commenting, or summarizing, with 23% of this number engaging meaningfully for the first time. In this way, we had managed to convince the unconvinced.

The community agent shipped after I left Microsoft, building upon the foundational framework I designed for the feature. Although I can’t speak to the specific results of the product, I can outline the success criteria I helped form, which included recurring-question load removed from named community managers (i.e. the number of high-confidence answers posted autonomously), review latency paired with answer acceptance, and a permission and label leakage gate.