AI agent that runs on intervals: reviews documentation for rules, drafts replies to incoming emails, and acts as a CRM buddy. Tracks leads and follow-ups without manual input.
The problem
Most "AI assistant" demos are a chat window. That's the wrong shape for ops work. Nobody wants to sit in a chat all day asking the bot to do its job. The real pain was follow-ups slipping: emails that needed a reply sitting unread, leads going cold because nobody logged them, the same context getting re-explained every time. The work is recurring and rule-bound, so it shouldn't need a human to kick it off.
What I built
- It runs on a cron, not a chat. The agent wakes up, reads the context docs, does its work, and goes back to sleep.
- Drafts land in Gmail with a label for human sign-off. Nothing sends unapproved, which was non-negotiable for a CRM touching real customers.
- Leads and follow-ups get tracked in a Notion board the client already uses, so there's no new tool to learn.
- It's deployed live on a private Google Workspace account.
The whole design is built around the human keeping the final say while the agent handles the reading, the drafting, and the remembering.
How it works
A scheduled cron job triggers the agent on intervals. It pulls rules and context from documentation, then uses OpenAI to draft replies to incoming email. Gmail and the rest of the work happen through the Google APIs, with drafts written to a labelled folder rather than sent. Lead and follow-up state lives in Notion, which the client was already running their pipeline on.
Outcome
Live and running on a private Google Workspace account, surfacing follow-ups and putting drafts in front of a person instead of letting threads rot. No public usage numbers yet. It's a working deployment, not a demo.