Agent #1 — Customer Service (Mafalda)

The Highest-ROI Agent to Deploy First

If you deploy only one agent, make it this one. Customer service is: - High volume (most brands handle 30-200+ tickets/day) - Highly repetitive (80% of tickets fall into 10-15 categories) - Time-sensitive (customers expect responses in minutes, not hours) - Directly revenue-impacting (slow CS = lost sales, returns, bad reviews)

What the CS Agent Does

Tier 1: Fully Autonomous (70-80% of tickets)

These are tickets the agent handles end-to-end without human intervention:

• "Where is my order?" → Pulls tracking from Shopify/3PL, sends personalized update with ETA
• "I want to return this" → Checks return eligibility (date, condition rules), initiates return, sends label
• "What size/variant should I get?" → Analyzes customer's past orders, product specs, generates recommendation
• "Do you have X in stock?" → Real-time inventory check across all locations
• "My discount code isn't working" → Validates code, checks conditions, applies manually if legit
• General product questions → Answers from product catalog + brand knowledge base

Tier 2: Agent Drafts, Human Reviews (15-20% of tickets)

• Complaints about product quality
• Requests for exceptions to policy
• Multi-issue tickets (return + exchange + complaint)
• High-value customer issues (lifetime value > €500)

Tier 3: Escalated to Human (5-10% of tickets)

• Legal or compliance issues
• PR-sensitive situations (influencers, media)
• Emotional/angry customers requiring empathy
• Fraud suspicion

Configuration Blueprint

The SOUL.md (Personality)

# CS Agent — Mafalda

You are Mafalda, the brand's customer service agent. Empática, resolutiva, directa, cercana.

## Voice
- Warm but efficient. Not robotic, not overly casual.
- Match the customer's energy: if they're frustrated, acknowledge it before solving.
- Use the customer's first name.
- Never use corporate jargon ("we apologize for any inconvenience")
- Instead: "I'm sorry this happened — let me fix it right now."

## Languages
- Primary: Spanish (Spain)
- Secondary: English
- Detect language from customer message and respond in same language

## Decision Rules
- If return is within 30 days and item is unworn: approve automatically
- If return is 31-45 days: offer store credit, not refund
- If return is > 45 days: escalate to human
- Discount codes: never create new ones without human approval
- Shipping issues: if delivery is > 5 days late, proactively offer €10 credit
- Always check customer lifetime value before deciding escalation threshold

## Escalation Triggers
- Customer mentions lawyer, legal, or lawsuit
- Customer is a verified influencer (> 10K followers)
- Ticket involves more than €200 in dispute
- Customer has expressed strong negative emotion 2+ times
- You are less than 80% confident in your response

Key Integrations

Integration	What For
Shopify	Orders, products, customer data, returns
Richpanel/Gorgias	Ticket management, conversation history
3PL API (Amphora, ShipStation)	Real-time tracking
Klaviyo	Check if customer is in a segment (VIP, first-time, etc.)
Internal knowledge base	FAQs, policies, product details, specs & guides

Workflow Example: "Where is my order?"

Customer: "Hi, I ordered 3 days ago and haven't received any shipping confirmation"

Agent Process:
1. Extract customer email/name from message
2. Query Shopify: GET /orders.json?email={email}&status=any
3. Find most recent order → check fulfillment status
4. If unfulfilled:
   - Check if item is in stock at fulfillment warehouse
   - If yes: "Your order is being prepared. You'll receive tracking within 24h."
   - If no: alert Ops Agent, tell customer: "We're checking on this and will update you today."
5. If fulfilled:
   - Pull tracking number from Shopify/3PL
   - Check tracking status with carrier API
   - Generate response with tracking link + estimated delivery
6. Set confidence level, send or queue for review

Real Metrics

Metric	Before	After	Impact
Average response time	4.2 hours	12 minutes	-95%
First-contact resolution	34%	71%	+108%
Tickets handled/day	45	180+	+300%
CSAT score	3.8/5	4.4/5	+16%
CS team hours/week	40+	12	-70%
Cost per ticket	€4.20	€0.60	-86%

The CS team member didn't lose their job — they moved from answering "where is my order?" 50 times a day to handling VIP relationships, managing the agent's learning, and focusing on cases that genuinely need human empathy.

Implementation Checklist

• [x] Set up agent with CS-focused SOUL.md (includes confidence scoring framework)
• [ ] Connect to Shopify (read access to orders, products, customers)
• [x] Connect to CS platform — automated ticket monitoring every 15 min
• [ ] Build knowledge base with top 20 FAQs and their answers
• [ ] Import return/exchange policy as structured rules
• [ ] Import product sizing guides
• [ ] Configure escalation thresholds
• [ ] Run in "shadow mode" for 1 week (agent drafts, human sends)
• [ ] Review shadow mode results, adjust thresholds
• [ ] Go live with Tier 1 autonomy
• [ ] Weekly review of escalated tickets to identify new autonomous patterns

Common Mistakes

1. Starting with 100% autonomy. Always start in shadow mode. Build trust gradually.
2. Not maintaining the knowledge base. If you launch a new product and don't update the agent's knowledge, it will hallucinate answers. Keep it current.
3. Ignoring edge cases. The 5% of tickets that need human handling will define your customer's perception of ythe brand. Handle them exceptionally.
4. Using the wrong LLM. For CS, you need the best reasoning model you can afford. Cheaper models make more mistakes on nuanced customer situations. We use Claude Opus for CS. It's worth it.