All work
082025Internal use

Inbox Triage & Auto-Reply Drafting Copilot

An Outlook copilot that classifies, drafts, and learns — giving senior staff ~5 hours a week back.

Before

Senior team members spent 60–90 minutes a day classifying and replying to repetitive stakeholder emails.

After

An Outlook add-in classifies incoming mail by intent, drafts a tone-matched reply, and learns from accept/reject signals. It saved ~5 hours/week per user, reached a 78% draft-acceptance rate, and held 91% triage accuracy across 12 categories.

Challenge

Senior team members spent 60–90 minutes a day classifying and replying to repetitive stakeholder emails.

Approach

Outlook add-in that classifies incoming mail by intent, drafts a tone-matched reply, and learns from accept/reject signals.

How it was built

  1. 1

    Intent classification

    Phase 1

    Built a 12-category intent classifier for incoming mail — triage first, because routing the email correctly is half the time saved.

  2. 2

    Tone-matched drafting

    Phase 2

    Generated reply drafts that match the user's tone, surfaced in Outlook where they already work rather than a separate app.

  3. 3

    Learn from accept/reject

    Phase 3

    Fed accept/reject signals back in so drafts improved over the first two weeks — acceptance climbed to 78%.

Key architecture decisions

Draft, don't send

Why · Keeping a human in the send loop is what makes an email assistant safe. The copilot proposes; the person approves.

Live inside Outlook

Why · Adoption dies if people have to leave their inbox. An add-in met users where they already are.

Impact

  • Saved ~5 hours/week per user across the pilot group
  • Acceptance rate on drafted replies climbed to 78% after 2 weeks
  • Triage accuracy held at 91% on 12-category classification
5 hrs/wk
saved
78%
accept
91%
accuracy

What I'd tell someone building this

  • 01 · Never auto-send. A tone-matched draft with a human approving is the sweet spot.
  • 02 · Meet users in their existing tool or adoption collapses.
  • 03 · Accept/reject feedback is a cheap, powerful signal — capture it from day one.

Tech stack

GPT-4Microsoft GraphOffice Add-in.NET

Ask anything about Inbox Triage & Auto-Reply Drafting Copilot

AI scoped to this project · GPT-OSS 120B