Overview

What it is: An AI-powered inbox triage + response drafting + routing system.

Goal: Classify every inbound email, extract key details, draft a safe high-quality reply, and route to the right workflow (sales/support/billing/etc.).

Why it matters: Creates “lead coverage” for email so messages get acknowledged fast and high-value threads escalate instantly.

Architecture (5-layer system)

1. Intake (email ingestion + normalization)
   • Ingests from Gmail/Outlook, shared mailboxes, and form-to-email.
   • Normalizes each message into a structured record:
     • thread_id, message_id, sender, subject, cleaned body, attachments metadata, timestamp, mailbox
     • Prior thread context (last N messages)
   • Cleans content by stripping signatures, quoted replies, tracking pixels, and optional forwarded chains.
2. Pre-processing + safety filters (deterministic rules)
   • Spam / threat / noise filters
     • Known spam patterns, blacklisted domains, suspicious formatting, excessive links, sender mismatch.
   • Compliance / risk guardrails
     • Detect payment/sensitive-info requests, legal threats, refund disputes.
     • Blocks auto-send and forces human review on risky categories.
   • Identity + relationship checks
     • Existing customer, CRM known sender, reply to outbound, net-new inbound lead.
   • Produces a preliminary safety tag:
     • safe_to_draft, requires_human_review, spam_ignore
3. AI understanding (intent classification + extraction)
   • Uses GPT as a structured classifier, returning strict JSON:
     • intent category (sales/support/billing/partnership/hiring/spam)
     • urgency, sentiment, action required
     • key fields (company, product, issue type, budget indicator, timeline)
   • Generates a 1–2 sentence human summary.
4. Response composer (draft replies + templates)
   • Template-first routing by intent (sales/support/billing/partnership/not-a-fit).
   • AI-enhanced personalization fills templates with constraints:
     • short, professional, minimal fluff
     • references the exact ask
     • clear next step
     • no hallucinated promises
     • no legal/financial commitments
     • no passwords/payment links
   • Link + scheduling injection
     • Sales: calendar link only after qualification.
     • Support: create ticket + request missing info (account email, screenshots).
5. Routing, handoff, and logging (ops layer)
   • Executes routing actions and creates operational artifacts:
     • CRM create/update + owner assignment + tags
     • Helpdesk ticket creation + categorization
     • Finance routing for billing
     • Slack notifications (summary + draft)
   • Logs everything for auditability:
     • classification result, draft version, routing outcome, response time, resolution time (optional)

End-to-end flow

1. New email arrives → webhook triggers ingestion + cleaning + thread context.
2. Safety filters run first (spam/risk/relationship).
3. AI classifies + extracts structured fields + recommended action.
4. Draft generated via template + constrained personalization.
5. Routing executes (CRM, Slack, tasks, tickets).
6. Optional human review gate:
   • Safe items can be auto-sent (optional mode).
   • Risky items go to Drafts with an explanation.

Key components (conceptual)

• inbox_agent/email_ingest.py
• inbox_agent/text_cleaner.py