AI Automation

Human in the Loop Automation

Task decomposition

An inbox is polled for new unread email (checked every minute).
The incoming email is used to build a reply-drafting prompt: reply naturally, keep it brief, ask to schedule a call for personal matters, and decline events/invitations/promotions politely.
A draft reply is generated.
The draft is sent to the user for approval, along with the original email it’s replying to, and the workflow pauses until a decision is made.
If approved, the reply is sent immediately as a real reply to the original email thread.
If rejected, the user is asked for free-text feedback on what to change, and the workflow pauses again until they respond.
The feedback is folded into a new prompt, which explicitly shows the AI its previous failed attempt and the user’s specific complaint, and the reply is redrafted.
The redrafted reply goes back through the same approval step, and the cycle repeats until approved.

Objective of the workflow

This is the human-in-the-loop counterpart to the judge-threshold version, instead of an AI judge deciding when a reply is good enough, a human approves or rejects each draft directly, with rejection feedback fed straight back into the next attempt.
It shows a different control pattern for the same underlying task: trading full autonomy for a review checkpoint on every outbound message. Built for CIS 515 (AI and Data Analytics Strategy).

Architecture