A single WhatsApp chatbot works well for simple, linear interactions — ordering food, answering FAQs, capturing basic lead info. But real businesses have complex interactions: a DHA property inquiry needs an agent who knows Karachi Phase 6 pricing, not one who handles Lahore. A hospital's WhatsApp line gets both appointment bookings and medical emergency questions — a bot handling appointments should never attempt to answer a medical query. Multi-Agent WhatsApp systems solve this through intelligent routing: the right conversation gets to the right handler, whether that's a specialized bot, a department, or a specific human agent.

Section 1: The Multi-Agent Architecture

A Multi-Agent WhatsApp system has three layers:

Layer 1 — The Router Bot: The first point of contact. Every incoming message hits the Router first. It identifies: what department does this belong to, what is the urgency level, and does this need a bot or a human? The Router should NOT answer content questions — it only classifies and routes.

Layer 2 — Specialized Agents (Bots or Humans): Each department/function has its own agent:

• Sales Bot: product info, pricing, lead capture
• Support Bot: order status, complaints, refunds
• Booking Bot: appointments, scheduling, availability
• Human Agents: complex queries, complaints, high-value leads

Layer 3 — Escalation Handler: When any agent cannot resolve an issue (3 failed attempts, explicit customer request, keyword triggers like "complaint," "refund," "manager"), the Escalation Handler takes over, assigns a ticket, and routes to the appropriate human.

Section 2: Building the Routing Logic

Step 1 — Intent Classification: Use a lightweight AI call to classify incoming messages. This runs on every message before routing:

def classify_intent(message: str) -> dict:
    """Classify WhatsApp message intent using Gemini Flash"""
    from google import genai

    client = genai.Client(api_key=os.getenv("GEMINI_API_KEY"))

    prompt = f"""Classify this WhatsApp message from a Pakistani customer.
Return JSON only:
{{
  "intent": "sales|support|booking|complaint|general|emergency",
  "urgency": "high|medium|low",
  "language": "english|urdu|mixed",
  "escalate_to_human": true|false,
  "reason": "one sentence"
}}

Message: "{message}"
Pakistani business context: This is a WhatsApp line for a multi-specialty hospital in Karachi.
"""

    response = client.models.generate_content(
        model="gemini-2.5-flash",
        contents=prompt
    )
    return json.loads(response.text)

Step 2 — WATI Department Routing: WATI supports agent assignment and inbox routing. When your classification engine returns an intent, route accordingly via WATI's API:

import httpx

async def route_to_department(wa_number: str, intent: dict):
    wati_headers = {"Authorization": f"Bearer {WATI_TOKEN}"}

    if intent["escalate_to_human"] or intent["intent"] == "emergency":
        # Assign to human agent immediately
        await httpx.post(
            f"{WATI_BASE}/api/v1/assignConversation",
            headers=wati_headers,
            json={"whatsappNumber": wa_number, "assignedTo": "on_call_agent"}
        )
        # Send priority message to team lead
        await send_emergency_alert(wa_number, intent)

    elif intent["intent"] == "sales":
        await activate_sales_bot_flow(wa_number)

    elif intent["intent"] == "support":
        await activate_support_bot_flow(wa_number)

    elif intent["intent"] == "booking":
        await activate_booking_bot_flow(wa_number)

    else:
        await activate_general_faq_bot(wa_number)

Step 3 — Escalation Triggers: Build escalation triggers at every bot level. If a bot detects:

• Customer has asked the same question 3+ times with no resolution
• Customer has explicitly said "manager," "complaint," "refund," "not working," "very upset," "koi kaam nahi"
• The conversation has been going for 15+ minutes without completion
• Customer sends an all-caps message (frustration signal)

Then immediately: send a handoff message ("Connecting you to our team, just a moment"), assign to available human agent in WATI, notify agent via WhatsApp with conversation summary.

Section 3: Pakistan-Specific Routing Considerations

Language Detection: Pakistani customers mix English, Urdu, and Roman Urdu freely. Your routing must handle:

• "I want to check my order status" (English — route to support bot)
• "mera order kahan hai" (Roman Urdu — route to same support bot, but activate Urdu template)
• "Complaint hai mujhe" (Roman Urdu complaint signal — escalate to human)

Business Hours Routing: Pakistani businesses typically operate 10 AM–8 PM PST. Build time-based routing:

• Business hours: Route to bots first, escalate to humans if needed
• After hours: Route all to bots, flag high-priority issues for morning follow-up
• Eid/Ramadan: Adjust greeting templates, notify customers of reduced hours

Section 4: Agent Routing Decision Table

MULTI-AGENT ROUTING MATRIX (Pakistan Business Context)
┌──────────────────────────────────────────────────────┐
│  MESSAGE TYPE           │ ROUTE TO       │ URGENCY   │
│  ──────────────────────────────────────────────────  │
│  "Price for product X"  │ Sales Bot      │ Low       │
│  "My order is late"     │ Support Bot    │ Medium    │
│  "Book appointment"     │ Booking Bot    │ Low       │
│  "Refund chahiye"       │ Human Agent    │ High      │
│  "Manager se baat karo" │ Human Agent    │ High      │
│  "Chest pain" (hospital)│ Emergency+Human│ Critical  │
│  "Same-day delivery?"   │ Sales Bot      │ Low       │
│  "Wrong item received"  │ Support Bot    │ Medium    │
│  "Kal delivery hogi?"   │ Support Bot    │ Low       │
│  "Discount milega?"     │ Sales Bot      │ Low       │
└──────────────────────────────────────────────────────┘

Exercise 1: Design the routing logic for a Pakistani e-commerce business (selling clothing on Daraz/Shopify). Create a classification table: list 10 types of incoming messages and how each should be routed (which bot, which human department, what urgency level). This is your routing specification — the blueprint before you build in WATI.

Exercise 2: Build the intent classification function using Gemini Flash. Test it with 10 real WhatsApp message examples (write them yourself or use sample messages). Evaluate: does the intent classification match what a human agent would do? Fix any prompts where classification is incorrect.

Exercise 3: Configure WATI's agent routing for at least 2 departments (e.g., Sales and Support). Create a test conversation for each. Verify that a "price inquiry" message correctly flows to the Sales bot and a "my order is late" message correctly routes to the Support bot. This end-to-end test confirms your routing system works before going live.

Pakistan Case Study: Karachi Multi-Brand Agency — 500 Chats Per Day

A Karachi digital agency managing WhatsApp for 3 clients (clothing brand, restaurant, skincare clinic) was running all 3 accounts on separate WATI instances with separate teams. When one client had a surge, the team was overwhelmed; others under-utilized.

They rebuilt with a unified multi-agent router:

Architecture:

All 3 client numbers → WATI Shared Inbox
                           │
                    Router Bot classifies:
                    - Which client?
                    - What intent?
                           │
        ┌──────────────────┼──────────────────┐
        ▼                  ▼                  ▼
  Clothing Bot        Restaurant Bot     Skincare Bot
  (sales/catalog)     (orders/menu)      (bookings/FAQs)
        │                  │                  │
        └──────────────────┴──────────────────┘
                           │
                    Escalation Pool:
                    3 human agents shared
                    across all 3 clients

Results after 60 days:

The multi-agent router cut staffing costs by 67% while tripling capacity. The 3 human agents now handle only what bots genuinely cannot — and they do it far better because they're not burned out answering "What are your rates?" 200 times a day.

Key Takeaways

• Multi-agent routing is the feature that scales a WhatsApp business from 50 to 500+ daily conversations — without it, a single bot tries to handle everything and handles nothing well
• Intent classification using Gemini Flash is fast (< 1 second) and accurate enough for production use — it doesn't need to be perfect, just better than random routing
• Escalation triggers based on conversation patterns (3 failed attempts, frustration keywords) are more reliable than waiting for customers to explicitly ask for a human
• Language detection in a Pakistani context requires handling three languages (English, Urdu, Roman Urdu) — build templates for all three since customers will use all three in the same conversation
• The shared human agent pool across multiple bots/clients is an efficiency multiplier — agents handle only escalations, which means fewer agents handle more volume at higher quality
• Business hours routing is a mandatory feature in Pakistan: after-hours bots must set clear expectations ("We'll respond tomorrow at 10 AM") to avoid customer frustration with bot-only interactions late at night