The average professional receives 121 emails per day. Studies show they spend 28% of their workweek managing email. That's 11 hours a week. Sorting, reading, replying, filing, flagging, and forwarding.
Most of that work is predictable. You know before you open the email whether it's urgent, routine, or noise. The subject line tells you. The sender tells you. The time it arrived tells you.
If you can predict it, an AI agent can handle it.
What Email Triage Looks Like
A properly configured email triage agent processes your inbox in three layers:
Layer 1: Classification — Every email gets categorized:
- Urgent (needs human attention now)
- Routine (can be answered with a template or drafted reply)
- Noise (newsletters, notifications, automated alerts)
- Sensitive (legal, financial, personal — always escalate)
Layer 2: Action — Based on classification:
- Urgent: Push notification to Slack with summary
- Routine: Draft a reply, queue for review (or auto-send if enabled)
- Noise: Archive, label, and move on
- Sensitive: Flag for human review, never auto-respond
Layer 3: Reporting — Daily summary:
- How many emails processed
- How many auto-handled vs. escalated
- Any threads that need follow-up
- Emails from new contacts (potential leads)
The Configuration That Matters
Sender-Based Rules
The most reliable triage signal is who sent the email. You can pre-classify senders into priority tiers:
email_triage:
sender_rules:
vip:
domains: ["biggest-client.com", "investor.com"]
contacts: ["ceo@partner.com", "legal@important.com"]
action: escalate_immediate
known_contacts:
source: crm # Pull from your CRM automatically
action: draft_reply
internal:
domains: ["yourcompany.com"]
action: classify_by_content
automated:
patterns: ["noreply@", "notifications@", "no-reply@"]
action: archive_and_label
unknown:
action: classify_by_content
flag: new_contact
VIP senders always get escalated. You never want an AI drafting a reply to your biggest client without human review. Known contacts get drafted replies. Automated senders get archived. Unknown senders get content-based classification plus a flag so you know about new people entering your inbox.
Content-Based Classification
For emails that can't be classified by sender alone, the agent reads the content:
email_triage:
content_rules:
urgent_keywords:
- "urgent"
- "ASAP"
- "deadline today"
- "invoice overdue"
- "server down"
- "security incident"
routine_categories:
scheduling:
patterns: ["let's meet", "calendar", "available", "free time"]
action: check_calendar_and_draft
follow_up:
patterns: ["following up", "checking in", "any update"]
action: draft_status_update
acknowledgment:
patterns: ["thanks", "received", "got it", "sounds good"]
action: archive
noise_categories:
newsletters:
headers: ["List-Unsubscribe"]
action: label_and_archive
github:
from: ["notifications@github.com"]
action: label_and_archive
marketing:
patterns: ["unsubscribe", "view in browser", "email preferences"]
action: label_and_archive
Sentiment Detection
For emails that don't match keyword patterns, sentiment gives the agent a safety net:
email_triage:
sentiment:
negative:
action: escalate
reason: "Negative sentiment detected — human review recommended"
confused:
action: draft_with_review
reason: "Customer may need clarification"
positive:
action: draft_reply
reason: "Positive interaction — standard response appropriate"
Negative sentiment always gets a human. An angry customer needs empathy, not a template.
The Draft Quality Problem
The biggest failure point in email triage isn't classification. It's the quality of drafted replies.
Bad draft: "Thank you for your email. I'll get back to you soon."
That's what the agent writes when it doesn't have enough context. It's technically correct and completely useless.
Good draft: "Hi Sarah, got your message about the Q2 budget review. I've blocked 30 minutes on Thursday at 2pm — here's the calendar link. I'll bring the updated projections we discussed last week."
The difference is context. The agent needs:
- Recent email history with this person. What have you been discussing? What was the last thing they asked?
- Calendar access. What's your actual availability?
- Your communication style. Do you write formally or casually? Do you use exclamation marks? Do you sign off with "Best" or "Thanks"?
- Business context. What projects are active? What deadlines are coming up?
Feed the agent all of this and its drafts go from "generic template" to "this sounds like I wrote it." Don't feed it context and you'll spend more time editing drafts than you would have spent writing the reply yourself.
Building Your Reply Templates
For the most common email types, pre-write ideal responses:
# reference/email-templates.md
## Meeting Request
Context needed: calendar availability, meeting preferences
Tone: Professional but warm
Example: "Hey [name], Thursday at 2pm works great.
I'll send a calendar invite. Looking forward to it."
## Follow-Up Response
Context needed: last interaction, current status
Tone: Proactive, specific
Example: "Hi [name], good timing on the follow-up.
We finished the review yesterday — I'll send the
updated doc by end of day."
## Invoice/Payment Acknowledgment
Context needed: invoice details, payment status
Tone: Professional, clear
Example: "Got it, thanks. Invoice #[number] is in
the queue — payment goes out on our next cycle
(Friday). Let me know if you need it sooner."
The agent uses these as style guides, not copy-paste templates. It adapts the structure and tone to each specific email.
Rollout Strategy
Don't turn everything on at once. Here's the rollout that works:
Week 1: Classification Only
The agent reads and classifies every email but takes no action. You review the classifications daily. Is it getting the categories right? Is "urgent" actually urgent? Is "noise" actually noise?
Accuracy target: 90%+ correct classification before moving to Week 2.
Week 2: Draft Mode
The agent starts drafting replies for routine emails. All drafts go to a review queue (Slack channel or a daily digest). You review and edit before sending.
Track: How many drafts did you send without editing? Target: 70%+ unedited.
Week 3: Auto-Archive
Enable auto-archiving for noise categories. Newsletters, notifications, and automated emails get labeled and archived automatically. This is the safest automation to enable first because the cost of a mistake is low.
Week 4: Selective Auto-Send
For the highest-confidence categories (scheduling confirmations, acknowledgments), enable auto-send. Start with the categories where the agent's draft approval rate was 90%+.
Month 2+: Expand
Gradually expand auto-send to more categories as confidence grows. Keep VIP senders and sensitive categories on human review permanently.
Measuring Success
After 30 days, check these numbers:
| Metric | Target | Red Flag | |--------|--------|----------| | Classification accuracy | 93%+ | Below 85% | | Draft approval rate | 80%+ | Below 60% | | Average response time | Under 15 min | Over 1 hour | | Emails escalated daily | Under 15% | Over 30% | | False urgents (urgent tag, wasn't urgent) | Under 5% | Over 15% | | Missed urgents (urgent email, not flagged) | 0% | Any |
The last metric is the most important. A missed urgent email is worse than 50 false urgents. Configure for sensitivity, not specificity. It's better to escalate too many emails than to miss one that mattered.
The Inbox Zero Myth
Email triage isn't about achieving inbox zero. It's about achieving inbox relevant. Your inbox should only contain emails that genuinely need your brain. Everything else is handled.
The goal isn't "no emails." It's "only the emails that matter."
For most people, that means seeing 10-15 emails a day instead of 120. The other 105 are categorized, drafted, archived, or auto-responded. Quietly. In the background. While you do actual work.
If you want this running for your inbox, book a call. We'll configure your triage rules, build your template library, and have the agent triaging by end of day. $999, one-time.
Need help with a specific triage configuration? Check our Gmail integration guide or reach out at openclawsetup.dev/meet.