Build a WhatsApp AI Bot for Your Business (India Edition)

Why WhatsApp is the default channel in India

For most Indian businesses, WhatsApp is not one channel among many — it is where customers already are. Building an AI assistant that replies inside WhatsApp, instead of a separate web widget, meets users where they already message you.

The architecture is simple: Meta's WhatsApp Business Platform receives the customer's message via webhook, your backend generates a reply with an LLM, and you send it back through the same API. AICredits handles the LLM half — one API key, INR billing, and access to every major model for the reply-generation step.

Architecture overview

Customer sends WhatsApp message
  → Meta Cloud API webhook → your backend
  → Backend calls AICredits (LLM generates reply)
  → Backend sends reply via Meta Cloud API
  → Customer receives response in WhatsApp

You need a Meta WhatsApp Business Platform account (via Meta Business Suite or a BSP like Interakt, Gupshup, or AiSensy) for the messaging leg, and an AICredits key for the AI leg. This post covers the AI leg — check Meta's or your BSP's current rate card for the messaging leg, since conversation pricing varies by category and changes periodically.

Step 1: Receive the webhook

Meta sends an HTTP POST to your webhook URL for every incoming message. A minimal FastAPI handler:

from fastapi import FastAPI, Request
import httpx
from openai import OpenAI
 
app = FastAPI()
client = OpenAI(base_url="https://api.aicredits.in/v1", api_key="sk-your-aicredits-key")
 
@app.post("/webhook")
async def whatsapp_webhook(request: Request):
    body = await request.json()
    entry = body["entry"][0]["changes"][0]["value"]
    if "messages" not in entry:
        return {"status": "ignored"}
 
    message = entry["messages"][0]
    from_number = message["from"]
    text = message["text"]["body"]
 
    reply = client.chat.completions.create(
        model="openai/gpt-4o-mini",
        messages=[
            {"role": "system", "content": "You are a support assistant for [Business Name]. Be concise and helpful."},
            {"role": "user", "content": text},
        ],
    ).choices[0].message.content
 
    await send_whatsapp_message(from_number, reply)
    return {"status": "ok"}

Step 2: Send the reply

WHATSAPP_TOKEN = "your-meta-access-token"
PHONE_NUMBER_ID = "your-phone-number-id"
 
async def send_whatsapp_message(to: str, text: str):
    url = f"https://graph.facebook.com/v20.0/{PHONE_NUMBER_ID}/messages"
    payload = {
        "messaging_product": "whatsapp",
        "to": to,
        "text": {"body": text},
    }
    headers = {"Authorization": f"Bearer {WHATSAPP_TOKEN}"}
    async with httpx.AsyncClient() as http_client:
        await http_client.post(url, json=payload, headers=headers)

What the AI side costs per conversation

Assume a typical support exchange: a 100-token system prompt, 40-token customer message, and a 60-token reply, repeated 3 times in a conversation thread (so context grows each turn).

| Model | Approx. cost per 3-turn conversation | |-------|----------------------------------------| | Gemini 2.0 Flash | under ₹0.05 | | GPT-4o-mini | ~₹0.06 | | DeepSeek V3 | ~₹0.05 | | Claude 3.5 Haiku | ~₹0.35 |

At 1,000 conversations a month, the AI cost alone stays under ₹100 on a cheap model — the WhatsApp messaging fee from Meta or your BSP will typically be the larger line item, not the AI. Budget accordingly and check your BSP's current per-conversation rate before estimating total cost.

Keeping cost under control

Set a per-key budget cap in the AICredits dashboard so a bug that loops on incoming webhooks (Meta retries on timeout) cannot run away. Use a cheap model like Gemini 2.0 Flash or GPT-4o-mini for the first response, and escalate to a stronger model only for messages your prompt classifies as complex.

Frequently Asked Questions

Do I need Meta approval to send WhatsApp messages with an AI bot?

Yes — any business messaging on WhatsApp goes through Meta's WhatsApp Business Platform, which requires business verification. Choose a BSP (Interakt, Gupshup, AiSensy, or similar) if you want a managed onboarding path instead of going direct to Meta.

Which model should I use for a WhatsApp support bot?

Start with a cheap, fast model — Gemini 2.0 Flash or GPT-4o-mini — for the majority of replies. Both handle typical FAQ and order-status queries well at a fraction of the cost of larger models.

Can I pay for the LLM usage in INR while my WhatsApp billing stays with Meta?

Yes. These are separate billing relationships — Meta or your BSP bills for message delivery, and AICredits bills the LLM calls in rupees via your prepaid wallet. There's no dependency between the two.

How do I handle voice notes sent over WhatsApp?

Download the audio via Meta's media API, then transcribe it with openai/whisper-1 or, for Indian languages, sarvam/saarika-v2 through the same AICredits key. See the audio API docs for the full flow.

Get started

• Create a free account and get an API key for the AI leg of your bot.
• Read the quickstart guide for request formats and streaming.
• Building the backend in FastAPI? See FastAPI + LLM Streaming: Production Backend Template.