AI-Powered Knowledge Base & Copilot Agent
The problem
When a new service entered the market, as single-threaded-leader, I chose to become point of contact for every question that came with it pre-launch / post-annoucement - from strategic queries to basic FAQs, landing from team members across the business. Why? Useful for gathering intelligence, useful for consistency in the face of ambiguity. Completely unsustainable as a long-term model.
What I built
Rather than deflect the volume, I leaned into it deliberately. Every question that came to me became a data point. I worked with cross fucntional teams to build out strategy and process around the service, and documented everything into a comprehensive knowledge base in SharePoint — structured, searchable, and built to stay live as new questions emerged.
But before I put an agent on top of it, I tested it properly.
The methodology
The build had three distinct phases:
Manual. I answered every query that came in, referencing and improving my knowledge base as I went.
Semi-agentic with human quality check. I used Copilot AI to reference the knowledge base directly - copying in real questions from the teams and reviewing how accurately and completely it synthesised the answers. This gave me a clear picture of where the knowledge base had gaps, where the structure wasn't working, and where my prompting needed refinement.
Automated. Only once I was confident in the quality and consistency of the responses did I build the Microsoft Copilot agent on top of the knowledge base and hand it over fully. The setup itself took about thirty seconds. The work that made those thirty seconds possible took considerably longer.
The result
GTM and Account Management teams get accurate, contextual answers instantly without routing through me. My active hours on the project dropped significantly. The knowledge base continues to update as new questions surface, which means the agent improves over time.
I went from being the bottleneck to being the architect of a system that doesn't need me in the room.
What it demonstrates
AI scales what already exists. The three-phase methodology — manual, semi-automated, automated — exists because automating an untested system just automates the problem. The agent inherited clarity that was built and validated before it ever went live.