Scenario A: Metadata Triage
Pass 1 Only
Email can be classified with high confidence using only metadata (sender, subject, headers). No body fetch needed - saves tokens and latency.
Your package is on the way! Track your shipment and view delivery updates...
Pass 1 - Metadata
✅ Sufficient Confidence
Known transactional sender + shipping subject pattern = high confidence auto-processing
sender: shipment-tracking@amazon.com
subject: "order...has shipped"
pattern: TRANSACTIONAL_SHIPPING
Pass 2 - Body
⏭ Skipped
Not needed - confidence threshold met in Pass 1
status: BYPASSED
tokens_saved: ~800
Triage Decision
Auto-classified and processed
PROCESS
Action
transactional
Category
low
Priority
0.96
Confidence
💰 Token Cost Comparison
85% Savings
on this email
💡
Two-Pass Architecture
Pass 1 uses only metadata (sender, subject, headers) which costs ~150 tokens. Pass 2 fetches the body (~850 additional tokens) only when needed. For transactional emails like order confirmations, Pass 1 is sufficient 90%+ of the time.
150
Tokens Used
1
Pass Count
~200ms
Latency
Scenario B: Body Analysis
Pass 1 + Pass 2
Ambiguous email that requires body fetch to determine intent. Pass 1 confidence is below threshold, triggering Pass 2 analysis.
Hi, I wanted to circle back on our conversation from last week. I think we might be able to help with your upcoming project. Would you have time for a quick call this week?
Pass 1 - Metadata
❓ Low Confidence
Generic subject "Re: Following up" provides minimal signal. Unknown sender domain.
sender: unknown domain
subject: generic reply pattern
confidence: 0.42 (below 0.70 threshold)
Pass 2 - Body
✅ Analyzed
Body reveals sales outreach from consulting firm. Contains scheduling request.
intent: SALES_OUTREACH
signals: "help with your project", "quick call"
confidence: 0.88
✉
New Email
Metadata only
→
🤔
Pass 1
Confidence: 42%
→
📝
Pass 2
Body fetch
→
✅
Classified
Sales outreach
Triage Decision
Classified after body analysis
ARCHIVE
Action
sales_outreach
Category
low
Priority
0.88
Confidence
💡
Confidence Threshold
When Pass 1 confidence is below 0.70, the system automatically triggers Pass 2. This balances token efficiency with classification accuracy. For this email, the extra ~850 tokens were necessary to correctly identify it as low-priority sales outreach.
1,000
Tokens Used
2
Pass Count
~600ms
Latency
Scenario C: Clarification Queue
User Feedback
Email remains uncertain even after body analysis. Surface a card to the user for manual classification, then learn from their feedback.
Hey! Hope you're doing well. I was wondering if you'd be free sometime next week to catch up? It's been a while!
Pass 1 - Metadata
❓ Uncertain
Generic Gmail sender, vague subject. Could be personal or professional.
sender: personal gmail
confidence: 0.35
Pass 2 - Body
❓ Still Uncertain
Casual tone, but unclear relationship. No prior conversation history found.
intent: MEETING_REQUEST?
relationship: UNKNOWN
confidence: 0.52
Help Me Understand This Email
I'm not sure how to handle this one
Who is Sarah Johnson to you? This will help me handle similar emails in the future.
📚 What I Learned
Sarah Johnson is a work contact
Future emails from sarah.johnson@gmail.com will be categorized as work-related
Personal Gmail can be work-related
Don't assume personal email domains mean personal context
Reclassified
Based on your feedback
Thanks! I've marked this as a work contact meeting request with medium priority. I'll draft a polite response suggesting available times from your calendar.
Next time Sarah emails, I'll know to treat it as work-related.
💡
Human-in-the-Loop Learning
When both passes fail to reach the confidence threshold (0.70), the email enters the clarification queue. User feedback is stored with a 90-day TTL and used to improve future classifications. This creates a personalized triage model over time.
0.52
Initial Conf.
0.95
After Feedback
90d
Learning TTL
Scenario D: Reply Draft
Style Matching
Generate a reply that matches the user's writing style for this recipient. Style is determined by analyzing past sent emails and recipient context.
Hi,
Could you review the attached Q1 report and let me know your thoughts by EOD Friday? Specifically looking for feedback on the growth projections in Section 3.
Thanks,
David
Could you review the attached Q1 report and let me know your thoughts by EOD Friday? Specifically looking for feedback on the growth projections in Section 3.
Thanks,
David
✉
Original Email
Parse context
→
🎨
Style Resolver
Match voice
→
📝
Draft Generator
Compose reply
🎨
Style Resolution
Formality
Professional
Tone
Friendly-Professional
Length
Concise
Greeting Style
"Hi [Name],"
💡
Style Guide Builder
The StyleGuideBuilder analyzes your past emails to this recipient and similar work contexts. It extracts patterns like greeting style, sign-off preferences, emoji usage, and formality level. The draft matches your voice, not a generic template.
23
Past Emails Analyzed
75%
Formality Score
0.91
Style Match
Scenario E: Reply-All Thread
Group Thread
Reply to a group email thread while preserving CC list, adjusting formality for the audience, and maintaining thread context.
Team,
Attached is the final creative for the Q2 campaign. Please review and confirm approval by tomorrow morning so we can hit our launch deadline.
@[Your name] - Can you confirm the landing page is ready?
Thanks,
Lisa
Attached is the final creative for the Q2 campaign. Please review and confirm approval by tomorrow morning so we can hit our launch deadline.
@[Your name] - Can you confirm the landing page is ready?
Thanks,
Lisa
Thread Context
Extracted from conversation history
5
Messages in Thread
4
Participants
campaign
Topic
you
Mentioned
👥
Group Reply Style
Formality
Semi-Formal
Audience
Cross-Functional Team
Reply Type
Reply All (preserve CC)
Action Required
Confirm status
Preserved
CC Recipients
All 4 recipients maintained from original thread
Adjusted
Formality Level
Slightly less formal than 1:1 with manager (60% vs 75%)
Added
Team Context
Note for broader team about UTM parameters
Maintained
Thread Headers
In-Reply-To and References headers preserved for threading
💡
Group Thread Intelligence
Reply-All drafts automatically: preserve all recipients, adjust formality for group context (slightly less formal than 1:1), include relevant context for the broader audience, and maintain thread headers for proper email client threading.
4
Recipients
60%
Formality
✅
Thread Preserved