# AI Agent: Personal Inbox Triager

> An agent that reads new email, drafts replies for low-stakes threads, files reference material, and escalates anything that needs a human in under an hour.

**Platform:** Agent  
**Category:** ai-agents  
**Difficulty:** Advanced  
**Tags:** agents, email, productivity  
**Updated:** 2025-05-08  
**Source:** https://promptfoo.tech/workflows/ai-agent-personal-inbox-triager

## Overview
Inbox zero via agent isn't magic — it's a small number of well-scoped tools and a strict policy about what the agent can send without approval. This design pattern uses read + classify + draft + file, with sending gated behind explicit human approval.

## How it works
1. Trigger: every 15 minutes during work hours.
2. The agent lists new mail with a Gmail tool, classifies each into: reply-needed, FYI, reference, spam.
3. For reply-needed with low stakes (scheduling, thanks, confirmations), it drafts a reply into Drafts — never sends.
4. FYI is archived with a label. Reference is filed to Notion with a summary. Spam is reported.
5. A morning digest lists what was done and any drafts waiting for your review.

## Benefits
- Recovers 30–60 minutes a day of shallow inbox work.
- Never sends without approval, so risk is bounded.
- Filed reference material becomes searchable in your knowledge base.

## Use cases
- Founders drowning in intros and scheduling threads.
- Consultants juggling many low-context client emails.
- Anyone whose inbox is 80% noise and 20% signal.

## Step-by-step
### Step 1: Define the policy
Write a one-page policy: what the agent can draft, what it must escalate, and forbidden actions (no sending, no calendar changes, no unsubscribes).

### Step 2: Pick your stack
OpenAI/Anthropic + a Gmail tool via MCP or Zapier + a Notion tool. Keep the toolset small.

### Step 3: Build the classifier prompt
A single system prompt with your policy, example classifications, and a JSON output schema.

### Step 4: Wire the actions
Each classification maps to one tool call. Drafts go to Gmail drafts; FYI applies a label; reference writes to Notion.

### Step 5: Add the morning digest
A scheduled job posts a digest to Slack/email with counts, drafts to review, and any escalations.

### Step 6: Run it in shadow mode first
For a week, log what it would have done without acting. Read the log daily and tune the policy before enabling actions.

## FAQs
**What model should I use?**

A mid-tier model (GPT-4o-mini, Claude 3.5 Haiku) is fine for classification. Reserve larger models for drafting the trickier replies.

**How do I keep it from making things up?**

Structured outputs plus refuse-when-unsure prompts. Never let the agent invent facts about the sender.

**Is this safe to point at a shared inbox?**

Only after weeks of shadow-mode runs. Shared inboxes have higher stakes and more edge cases.

---
Licensed under CC-BY-4.0. Attribution: PromptFoo.tech