Generate empathetic customer replies with Claude AI and auto-escalation

🧠 Overview

Generate empathetic, professional reply drafts for customer or user messages.
The workflow detects sentiment, tone, and risk level, drafts a concise response, sanitizes PII/links/emojis, and auto-escalates risky or low-confidence cases to human review.

⚙️ How It Works

1. Input — Manual Test or Webhook Trigger
2. AI Agent (Empathy) — returns { sentiment, tone, reply, confidence, needs_handover }
3. Post-Process & Sanitize — removes URLs/hashtags, masks PII, caps length
4. Risk & Handover Rules — checks confidence threshold, risk words, and negativity
5. Routing — auto-send safe replies or flag to Needs Review

🧩 Setup Instructions (3–5 min)

• Open Set Config1 and adjust:
  • MAX_LEN (default 600)
  • ADD_FOLLOWUP_QUESTION (true/false)
  • FORMALITY (auto | casual | polite)
  • EMOJI_ALLOWED (true/false), BLOCK_LINKS (true/false)
  • RISK_WORDS (e.g., refund, lawsuit, self-harm)
• Connect Anthropic credential to Anthropic Chat Model
• (Optional) Replace Manual Trigger with Webhook Trigger for real-time use

> Tip: If you need to show literal angle brackets in messages, use backticks like <example> (no HTML entities needed).

📚 Use Cases

1) SaaS Billing Complaints

• Input: “I was billed after canceling. This is unacceptable.”
• Output: Calm, apologetic reply with refund steps; escalates if refund is in RISK_WORDS or confidence < 0.45.

2) Product Bug Reports

• Input: “Upload fails on large files since yesterday.”
• Output: Acknowledges impact, requests logs, offers workaround; routes to auto-send if low risk and high confidence.

3) Delivery/Logistics Delays

• Input: “My order is late again. Should I file a complaint?”
• Output: Empathetic apology, ETA guidance, partial credit policy note; escalates if language indicates legal action.

4) Community Moderation / Abuse

• Input: “Support is useless—you’re all scammers.”
• Output: De-escalating, policy-aligned response; auto-flags due to negative sentiment + risk keyword match.

5) Safety / Self-harm Mentions

• Input: “I feel like hurting myself if this isn’t fixed.”
• Output: Immediate escalation, inserts approved resources; never auto-sends.

🚨 Auto-Escalation Rules (defaults)

• Negative sentiment
• Message matches any RISK_WORDS
• confidence < 0.45
• Mentions of legal, harassment, or self-harm context

🧪 Notes & Best Practices

• 🔐 No hardcoded API keys — use n8n Credentials
• 🧭 Tune thresholds and RISK_WORDS to your org policy
• 🧩 Works on self-hosted or cloud n8n
• ✅ Treat outputs as drafts; ship after human/policy review

🔗 Resources

• GitHub (template JSON): https://github.com/yskmtb0714/n8n-workflows/blob/main/empathy-reply-assistant.json