Microsoft 365 + CoCounsel

Problem

Legal work is scattered. A general counsel learns about a new regulation, then spends days bouncing between email, a document repository, chat threads with outside counsel, and legal research tools — copying context from one place to the next by hand. Every hop costs time, and every manual transfer is a chance for something to get lost. Our bet: stop asking lawyers to come to the AI. Bring the AI to the apps where their day already happens.

Constraints

I was designing inside someone else's product. Copilot extensions dictate where and how a third party can appear in Microsoft 365, so I worked within Microsoft's Fluent patterns while keeping CoCounsel's identity legible — a user should always know whether an answer came from their own files, from Thomson Reuters content, or both. The second constraint was trust. Attorneys don't act on text they can't check, so every response had to carry visible sourcing, and nothing could execute without a human approving it. And because the first public appearance was a live keynote, the flow had to hold up in front of thousands of people with zero room for error.

Process

As the only designer, I ran the full arc myself: discovery calls with general counsels and outside firms, mapping their current process end to end, low-fidelity flows to align Thomson Reuters and Microsoft stakeholders, then high-fidelity Figma prototypes built on the Fluent kit. I tested with practicing attorneys under NDA, cut anything they hesitated on, and worked shoulder to shoulder with engineers on both companies' teams through keynote rehearsals.

The turning-point decision

The demo follows Sam, in a busy general counsel's office, on the morning an email arrives from outside firm Marshall & Partners: Canada has banned the import of goods made with child labor, and the company's supply chain is exposed. What used to kick off days of manual coordination across legal, risk, and compliance now runs as a single assisted flow — and the screens below walk through it exactly as it ran on stage.

Final solution

Source-first answers: Every claim links back to its origin — a clause in the company's contract, a policy line, Thomson Reuters content. Citations sit at the same visual weight as the answer, because for this audience "where did this come from" matters as much as "what." Drafts, never actions: The system proposes; the person disposes. Across all thirteen screens there is not one moment where CoCounsel sends, signs, or publishes on its own. One grammar, five surfaces: Prompt, sourced answer, proposal, approve — the same loop in Outlook, Teams, Word, HighQ, and the browser. Users learn it once and it works everywhere, while each surface still feels native to its host app.

Measured outcome

The demo ran live during the Microsoft Build 2024 keynote — one of a handful of partner demos on that stage. The interaction patterns above now extend into CoCounsel Drafting, which entered presale shortly after with a market launch that July, letting customers draft and refine contracts and playbooks without leaving Word.

What I learned

Designing AI for lawyers is designing for skepticism, and that's a feature. The more legible I made the sourcing and the approval steps, the more willing our test users were to hand the tedious parts over — which is exactly what the product needed them to do.