Microsoft · 2023–Present

Prototyping the systems behind AI product design

I brought complex behavior-authoring experience from Nuance Mix.dialog into Copilot Studio, then moved into Microsoft Core AI to help a 30+ person design organization explore, critique, and build AI products faster.

Copilot Studio

Core AI

Azure Portals

Prototyping

Role progression

The scope grew from a product problem to a design-system problem for AI work itself

The promotion and transfer matter because they show a continuous expansion of responsibility—not three unrelated Microsoft projects.

I joined the team designing conversational AI authoring experiences, bringing prior experience with deeply nested logic and behavior authoring from Nuance Mix.dialog.

“The transferable insight was not a screen. It was a model of where conditional complexity becomes hard to comprehend.”

Chapter 01 · Copilot Studio

Turning “feature parity” into a decision about which complexity mattered

Copilot Studio needed stronger support for nested conditions. My Mix.dialog background gave the team a working reference for how enterprise authors build, read, and debug complex conversational logic.

Problem

Not a checklist exercise

Parity affects how authors scan branches, understand precedence, recover from errors, and trust what a system will do at runtime.

Approach

Working from a Mix reference point

I worked toward the functional depth users had in Nuance Mix, while fitting Microsoft’s product model and interaction patterns.

Insight

Model, not a screen

The result was a working model of where conditional complexity breaks comprehension, debugging, and trust.

Three-day design sprint

Aligning stakeholders around the “big rocks”

I planned and ran a three-day workshop to convert a broad parity goal into a shared view of the highest-value problems to pursue.

Sprint structure

Day by day

The three days moved from shared understanding, to explored options, to prioritized commitments.

Day 01

Build the shared model

Align on the Mix reference point, Copilot Studio’s current model, user needs, technical boundaries, and where “parity” concealed different assumptions.

Day 02

Explore the opportunity space

Generate and compare approaches across authoring, comprehension, validation, and debugging instead of prematurely converging on one interface.

Day 03

Choose the big rocks

Prioritize the areas that could create the largest step change and leave stakeholders with a clearer product sequence for deeper exploration.

Chapter 02 · Microsoft Core AI

A prototyping layer for a 30+ person design team

Production repositories were too costly and constrained for many early design questions, and static design tools could show a state but not the behavior of an agentic system. My transfer to Core AI made that gap the center of my role: build smaller, safer environments where designers, PMs, and engineers could make behavior real early enough to learn from it.

Explore

Agentic playgrounds

Agentic playgrounds and focused GUI tools for early-stage exploration.

Make it real

Mini products

Mini products with representative behavior and data that make a concept testable.

Evaluate

Live feedback

Live annotation, feedback, and reusable agent skills that turn critique into next steps.

Flow

Idea → behavior → feedback → iteration

The throughline connecting exploration, building, and evaluation into one practice.

The prototype portfolio

Different tools for different kinds of uncertainty

The output was not one monolithic platform. It was a family of focused tools that reduced the cost of answering a design question—including early feature explorations for Azure SRE Agent, presented here as an application of the method rather than finished product work.

Official work · private

Agentic CodePen

A GUI prototyping environment where designers use agents to create and modify working interface experiments in a tight visual loop.

Official work · private

Live annotation

Tools for marking up a running interface and converting feedback into clearer, actionable context for an agent or implementation partner.

Official work · private

Product mini apps

Focused versions of Microsoft products that preserve the behavior needed for design evaluation while staying fast to change.

Official work · private

Agent skills

Robust instructions and context packages that help Copilot and other agents apply correct product conventions with less drift.

Azure portals

Making agent-generated prototypes feel native across Azure portal experiences

Some of the mini-app and agent-context work extended into Azure portal experiences. My contribution was primarily the shared prototyping infrastructure and team enablement—not product ownership for an individual Azure service.

Product context

Which Azure experience is being changed?

Authority map

Which source answers tokens, components, and behavior?

Agent decision

Which Fluent 2 variant and resource should be used?

Why product context mattered. “Use Fluent 2” was not precise enough—an agent also needed to know which product surface, repository, component source, and local conventions were authoritative for the task.

Leadership as infrastructure

The tools mattered, but adoption required a weekly practice

I paired implementation with teaching: weekly learning sessions, hands-on guidance, and patterns the design team could reuse without waiting for a specialist. That built shared capability with LLMs and agents, made interactions executable earlier so teams could react to real behavior, and gave agents more reliable context—reducing plausible but incorrect implementation choices.

Let’s build the systems behind better AI products.

Erik Drouhard · Senior UX Designer and UX Engineer

Get in touch