Shaping Microsoft’s Multi-Agent Future: How an Incubation Project Influenced Product Strategy Across Teams

The Approach: From Research to Executive Presentation

Phase 1: Finding the Right Problem

We started broad, exploring multiple automation scenarios across industries. But breadth wasn’t the goal. Depth and believability were. I took the lead on identifying invoice processing for shipping logistics as our north star scenario. Why? Three reasons:

1. Real pain, real money: Our Fortune 500 customer was manually processing invoices from contracted carriers, losing $200M annually to overcharges (5% of $4B in shipping costs)
2. Complex enough to matter: The workflow required multiple specialized agents (duplicate detection, taxonomy normalization, contract validation, discrepancy flagging)
3. Validated by experts: Industry advisors confirmed this was a widespread, high-value problem. I personally reached out to supply chain advisors and customer engagement specialists, which led to direct access to the logistics manager who became our validation partner throughout the project. This wasn’t just research. It was building the credibility we’d need when presenting to executives.

The complete agent workflow, from invoice ingestion to discrepancy resolution. This diagram became our team’s north star and a key artifact in executive presentations.

Phase 2: Designing the “Teaching” Experience

The breakthrough insight: Users shouldn’t need to understand AI architecture to build multi-agent systems. They should just describe their process, and the system should do the intelligent work of decomposing it into specialized agents.

I took ownership of designing the conversational AI interface where users “teach” the system their workflow. This wasn’t a traditional chatbot. It was a guided, collaborative dialogue where:

• The AI asks clarifying questions (“What happens if a duplicate is found?”)
• Users provide natural language instructions, not code
• The system translates intent into agent logic transparently
• Users can test and refine as they go

Designing this required balancing simplicity with power. Too simple, and it wouldn’t handle edge cases. Too complex, and we’d lose our core promise: “Focus on your process, not on learning AI.”

I used AI (Copilot) extensively here, not just to explore design ideas, but to craft the actual conversational prompts and response patterns that would feel natural and confidence-building.

The creation flow, from initial setup to process mapping to agent configuration. Each step builds user confidence while gathering the intelligence needed to automate correctly.

Phase 3: Proving It Works

Concepts are nice. Working demonstrations are convincing.

I led the effort to bring this vision to life through high-fidelity prototypes and animated videos. This meant:

Organizing the chaos: With dozens of screens designed by four different designers, I created the sequencing and information architecture that would tell a coherent story.

Building the animations: I partnered with the lead interaction designer to animate each section, ensuring transitions felt intuitive and the AI’s “thinking” was visible without being overwhelming.

Crafting the narrative: Using Copilot, I developed the voiceover script that would guide viewers through the experience, not just explaining features, but telling Mona’s story of transformation from manual processor to strategic analyst.

Producing the final videos: I assembled everything into two presentation videos: one focused on creation (how users build agent workspaces), the other on execution (how the system handles invoices and surfaces insights).

The workspace in action: identifying discrepancies, managing exceptions, and surfacing strategic insights. This moved Mona from “invoice checker” to “negotiation strategist.”

Phase 4: Validation and Stakeholder Alignment

Throughout the project, I maintained a dual feedback loop:

External validation: Regular check-ins with our Fortune 500 customer and industry advisors ensured our scenario was accurate and our solution was viable. Their eventual endorsement gave our executive presentation credibility.

Internal alignment: Weekly team reviews, critiques, and task assignments kept us moving forward without losing sight of the bigger picture. When the team wanted to add more features, I made the call to wrap it. We had enough to prove the concept. More would just dilute the message.