TLDR
Built an AI support assistant for a developer-tools company. It searches the places a support engineer would normally check, then prepares a draft for human review.
The workflow went from 5–10 minutes of research to around 30 seconds.
The Problem
Answering a product question rarely started with writing. The first step was investigation:
- Search for similar forum threads
- Check internal discussions and product documentation
- Look for an existing issue in the tracker
- Pull the useful context into a response
Doing this manually took 5–10 minutes for a typical question. The information already existed, but it was spread across different systems.
The Approach
I built an assistant that follows the same investigation path:
- Understands the user’s question
- Searches forum history for related discussions
- Searches internal conversations and documentation
- Checks the issue tracker for known bugs or planned fixes
- Produces a review-ready answer with the relevant context
It does not send anything automatically. A human reviews the answer, corrects it when needed, and decides what the user should receive.
Results
- Research-to-answer time dropped from 5–10 minutes to ~30 seconds
- Review stayed with the support engineer
- More time could go toward diagnosis and unusual product issues
The useful metric here is speed to a review-ready answer, rather than an automation percentage.
What I Learned
Drafting the final message was usually quick. Finding the right thread, document, or issue took most of the time.
A confident answer is useless if it was built from an outdated discussion or the wrong issue.
Support often needs judgment, especially when product behavior is changing. The assistant prepares the context. The engineer makes the call.
Related
- about: More about my work