# Build: Customer Support Autopilot

You are an expert AI agent product engineer. Scaffold and implement an MVP for the following agent product. Use modern best practices, TypeScript, and production-ready patterns.

## Product Overview

**Category:** Customer Service
**Target customer:** SaaS companies with 500+ monthly support tickets
**Difficulty:** Intermediate
**Estimated time to MVP:** 4-6 weeks
**Market score:** 9.4/10

## Problem

Support teams drown in repetitive tickets about billing, password resets, and feature questions. Response times slip and churn rises.

## Solution (Agent Behavior)

An AI agent that triages tickets, resolves tier-1 issues autonomously, drafts responses for complex cases, and escalates with full context when needed.

## Revenue Model

Per-seat SaaS ($49-199/mo) or per-resolution pricing ($0.50-2.00/ticket)

**Monetization tips:** Start with draft-only mode to reduce risk. Charge per resolved ticket once you hit 70%+ auto-resolution on tier-1 issues.

## Recommended Tech Stack

- Next.js
- OpenAI API
- LangChain
- PostgreSQL
- Zendesk/Intercom API

Prefer **Next.js 15 (App Router)** for the web app unless the stack above specifies otherwise. Use the **Vercel AI SDK** for streaming agent responses and tool calling.

## Required Integrations

- Help desk (Zendesk, Intercom, Freshdesk)
- Knowledge base
- Stripe for billing context

## Key Features (MVP scope)

- Auto-triage by urgency and topic
- Knowledge base sync
- Draft mode for agents
- Escalation with context bundle
- Resolution analytics

## Build Plan — follow these steps in order

1. Connect to help desk API and ingest historical tickets for training context
2. Build RAG pipeline over docs, FAQs, and past resolved tickets
3. Create agent workflow: classify intent, retrieve context, generate response, confidence score
4. Add human-in-the-loop approval for responses below confidence threshold
5. Implement escalation rules and handoff with conversation summary
6. Ship analytics dashboard: resolution rate, CSAT proxy, time saved

## Implementation Requirements

1. **Agent loop:** perceive context → plan → call tools → observe → iterate until task complete.
2. **Tool use:** wrap each integration as a typed tool with Zod-validated inputs.
3. **RAG (if applicable):** chunk domain docs, hybrid search, cite sources in outputs.
4. **Human-in-the-loop:** draft/approve mode for high-stakes actions (send email, update records).
5. **Auth & multi-tenancy:** isolate customer data per account from day one.
6. **Observability:** log every agent run (prompt, tools, latency, cost) for debugging.
7. **UI:** clean dashboard showing task status, agent steps, and edit-before-send controls.

## Deliverables

- Project scaffold with README and `.env.example`
- Core agent workflow implementing the happy path
- At least one integration wired end-to-end
- Simple UI to trigger the agent and view results
- Instructions to run locally (`npm install && npm run dev`)

Start by proposing the folder structure, then implement step by step. Ask clarifying questions only if a requirement is ambiguous.