The organization
A program team answers a high-volume support inbox. Most messages are variations on questions they’ve answered before — but every reply represents the program, so it has to be accurate and sound right.
The problem
The answers already live in the team’s FAQ, yet each reply still meant finding the right entry, adapting it, and writing it in the program’s voice. AI can draft from the FAQ while staff review every reply.
What we built
This is a concept, shared for the reusable pattern. It connects the team’s FAQ (in Google Sheets, editable without code) to its Gmail inbox, with an AI step between: it spots likely support requests, finds the FAQ entries that answer them, and leaves a draft reply on the thread for a person to send.
There are two ways to build it. The team scoped a custom Python service. Today there’s a lighter option: an agentic assistant like Cowork can do the FAQ-matching and drafting with no custom code, leaving mainly the inbox-watching to automate.
How it works
As designed:
- Load the FAQ and tone examples from Google Sheets.
- Scan recent Gmail threads for likely support requests; retrieve the best-matching FAQ entries.
- An AI step drafts a reply in the program’s voice, grounded in those entries.
- The draft lands on the original thread for a staff member to review.
- Every interaction is logged, surfacing where the FAQ needs work.
What it produces
- Draft replies on the original threads, ready for review
- The FAQ citations and confidence behind each draft
- Logs of every interaction
Where humans stay in the loop
The service drafts; it never sends. Keep the confidence score visible, but require staff review for every draft until the system has been tested against real outcomes. The FAQ’s blind spots are where drafts are weakest and judgment matters most.
Why it works
- The FAQ grounds every draft. Grounding drafts in approved FAQ answers makes mistakes easier to trace, such as selecting the wrong entry.
- Review improves the FAQ. Reviewing drafts also reveals gaps in the FAQ.
What to copy
- Build the knowledge base first. The FAQ is the real asset; the AI uses it to draft replies.
- Choose the review level deliberately. You can keep the send decision with staff for sensitive or consequential replies, while lower-risk categories can move to lighter review after the system has been tested. The event email draft builder uses the stricter draft-only pattern.
- Log citations and confidence. Reviewers move faster seeing why the system answered as it did, and the logs double as a to-do list for the FAQ.
Since this design touches a live inbox, read Data security and privacy basics before building anything similar.
