Email Triage Decision-Routing Workflow (Human-in-the-Loop)

A portfolio case study for Business Automation / Process Automation: route inbound email cases to the right path, escalate uncertainty, and keep human approval at critical decisions.

Problem

Operations inboxes receive mixed requests (billing, access, incidents, vendor messages). Manual triage is slow, inconsistent, and hard to audit.

Solution

This project defines a decision-routing workflow that:

• classifies email intent,
• applies routing and risk rules,
• generates a recommended next step and draft response,
• requires a human gate before final action.

Scope / Non-goals (Safety)

• No auto-send, no deletion, no autonomous actions.
• Draft replies are suggestions only; a human operator is always the final decision maker.
• Examples are synthetic; do not store or commit real customer PII.

Workflow

1. Ingest email from shared mailbox.
2. Build a Decision Context Card (intent, risk, recommendation, draft).
3. Route by rules (standard path vs. edge-case queue).
4. Trigger human review/approval for edge cases or high-risk items.
5. Log decision outcome for audit and continuous improvement.

Inputs / Outputs

Inputs

• Email subject/body + metadata (sender, timestamp, thread).
• Routing rules and policy constraints.

Outputs

• Decision Context Card.
• Suggested route + operator decision record.
• Draft reply (not auto-sent).

Edge Cases Covered

• Low-confidence classification.
• Missing required context (incomplete request).
• Potentially sensitive/high-impact requests.
• Conflicting signals (urgent tone vs low operational risk).

KPIs (Planned)

• Triage cycle time.
• First-pass routing accuracy.
• % cases requiring human escalation.
• Rework rate after operator review.

Note: KPI improvement claims are intentionally omitted until baseline and measured results are collected.

Documentation

• Architecture
• Guardrails
• Evaluation Rubric
• Privacy
• Portfolio Audit
• Agent Reasoning

Quick Demo / Artifacts

This repository is documentation-first (case study + artifacts). It does not include a production runtime yet.

• Synthetic sample emails and expected decision cards in examples/.
• Architecture and controls documentation in docs/.
• Sanity check for examples:

  python scripts/sanity_check.py

Power Platform Implementation Mapping

• Power Apps: intake/review UI for each email case, with operator decision capture.
• Power Automate: orchestration flow for classification handoff, routing rules, approvals, and notifications.
• Outlook/Exchange Connector + Shared Mailbox: monitored intake channel for operational emails.
• SharePoint or Dataverse: persistent case records, labels, and decision logs.
• Teams + Approvals: human-in-the-loop gate for escalations and exceptions.

What This Proves

• Ability to model a credible operations workflow with clear handoffs.
• Human-in-the-loop decision design (not blind automation).
• Governance mindset: guardrails, auditability, and documented edge-case handling.
• Readiness for Junior Business Automation / Power Platform analyst roles.