✨ TLDR

Built an AI copilot for the support team at a hypergrowth startup (valued at $10B+). The agent takes a support ticket, searches docs and past resolutions, and drafts a response.

Went from 0% automation to ~80% resolution rate.


🎯 The Problem

Imagine being the support team for a company in hypergrowth mode. User base exploding. Thousands of questions pouring in. Every day.

Before I came in:

  • Everything was manual. No automation, no tooling.
  • Same questions kept appearing. Already answered in docs or past tickets.
  • Support Agents spent more time researching than helping. Context was scattered.

💡 The Approach

Build a copilot, not an autopilot.

The goal wasn’t to auto-respond to users. That’s risky for a company of this scale. Instead, I built a copilot that:

  1. Takes a support ticket ID
  2. Parses the user’s question and context
  3. Searches product documentation for relevant answers
  4. Searches past resolved tickets to find similar issues
  5. Drafts a response for the agent to review and send

The agent does the research. The human makes the call.


⚡ Shipping Fast

I got something working fast, put it in front of the team, and improved based on real usage.

This is how internal tools should work — ship, learn, iterate. The first version was rough. The latest version was hitting 80% resolution.


📊 Results

MetricBeforeAfter
Resolution automation0%~80%
Time-to-first-responseManual researchNear-instant draft
Agent workloadResearching everythingFocus on complex issues

The support team went from drowning in tickets to having breathing room for the hard problems.


🧠 What I Learned

Copilot > Autopilot for support. Copilots keep humans in the loop while still saving 80% of the work.

Search quality is everything. The agent is only as good as its retrieval. If it can’t find the right docs or past tickets, the draft is useless. I spent most of my iteration time on setting up the right tools for the agent.

Internal tools are underrated. Companies invest millions in customer-facing products but neglect internal tooling. A few days of work can unlock massive leverage for internal teams.