33 lines (27 loc) · 1.83 KB

PRD: AI-Powered Support Triage

1. Problem Statement

What is the problem? Support agents spend 30% of their time manually categorizing and routing incoming tickets.
Who is experiencing it? Tier 1 Support Agents.
Why solve it now? Ticket volume is growing 15% MoM, and we are missing our 2-hour initial response SLA on 40% of tickets.

2. Proposed Solution

High-level description: An LLM-based service that reads incoming tickets, predicts the category, urgency, and best-fit team, and automatically routes the ticket.
Value proposition: Reduces manual triage time to zero, improving SLA compliance and agent satisfaction.

3. Key Features (MoSCoW)

Must Have: Auto-categorization (Billing, Tech, Account), Urgency prediction (Low, Med, High), Auto-routing to correct queue.
Should Have: Suggested initial response draft for the agent.
Could Have: Confidence score visible to the agent.
Won't Have: Fully automated replies to the customer (human-in-the-loop only for V1).

4. User Experience / Flow

Step-by-step flow:
1. Customer submits ticket.
2. System processes ticket in background (<5s).
3. Ticket appears in the correct agent queue with tags applied.
Key states: If confidence is <80%, route to a "Needs Manual Triage" queue.

5. Success Metrics

North Star Metric: % of tickets correctly routed on the first try (Target: >85%).
Supporting Metrics: Average time to initial response (Target: <2 hours).
Counter Metrics: % of tickets re-routed by agents (Target: <10%).

6. Open Questions / Risks

Technical risks: Latency of the LLM call delaying ticket creation.
Business risks: Incorrect routing of high-urgency Enterprise tickets.
Open decisions: Which model to use (Claude 3.5 Haiku vs Sonnet) for the optimal cost/latency/accuracy tradeoff.