Skip to content
AI & Ecommerce Technology8 min read

AI Support Chatbot for D2C — Automate the Repeat Questions

Raghoo Bokam By Raghoo Bokam, Founder & CEOUpdated

An AI support chatbot for a D2C brand should do one job: answer the questions your team already answers the same way every day — order status, returns, sizing, stock — and hand anything about money, damage or an angry customer to a person. Switch it on in stages, starting with 20% of chats, and read its conversations every week. And budget for it: since 1 October 2026, Meta charges for WhatsApp service replies past the first 1,000 a month on each number.

The inbox isn’t busy. It’s repetitive

You know the evening. Dispatch is done and the phone still lights up: “order kab aayega?”, “M fit hoga?”, “return kaise karun?”. Someone opens the courier panel, copies a tracking link, pastes it into the chat, and does it again before they’ve put the phone down.

That’s answering the door, not running a front desk. Most of those messages have one right answer that already sits in your order system or your policy page. Take an example store doing 3,000 orders a month: if one buyer in three asks “where is my order?” once, that’s 1,000 doors opened by hand every month — for an answer a bot can fetch in seconds.

In my experience, the trigger is rarely a wish for AI — it’s a team that can’t keep up with the same few questions.

Find your own mix before you buy anything:

  1. Export the last three months of chats from WhatsApp, Instagram and email.
  2. Read and tag 200 of them by type: order status, return or exchange, size, stock, damaged item, payment, complaint, other.
  3. Count. The types that are both frequent and answerable from data go to the bot first.

What the bot answers, and when it hands over

QueryWhat the bot doesHand to a person when
Where is my order?Asks for the order number or phone, fetches tracking from the courier, sends the status and expected dateTracking is stuck, or it says delivered and the customer says it isn’t
Return or exchangeCollects the reason, checks the return window, opens the return requestThe item is outside the policy, or the customer disputes it
What size?Asks height, weight and usual size, answers from your size chartThe product has no size chart, or they want an alteration
Do you have it in blue?Searches the catalogue and sends product linksIt’s out of stock and they want a restock date
Damaged or wrong itemCollects photos and the order number, opens a ticketAlways — a person decides refund or replacement
Payment or refundSays a person is on itAlways, straight away
Angry customerSpots it and stops answeringAlways, straight away

The cheapest bot reply is the one nobody needs. Send order-confirmed, shipped and out-for-delivery updates as WhatsApp utility templates, and a lot of “where is my order?” never gets typed. As of November 2026, Meta’s India rate for a utility template is ₹0.115 before tax (Meta’s INR list rates, linked from its pricing page); providers add a markup. The setup steps are in my WhatsApp automation guide, and returns themselves are cheaper to fix at the source — see my guide to cutting returns costs.

Before launch: write the answers, then the bot

Week 1 is a document, not software. A simple bot with good answers beats a clever bot with a vague policy page.

  • Write about 50 answers in plain words — shipping times by zone, the return window, COD rules and any prepaid discount you offer to move buyers off COD, size charts, care instructions. The bot can only be as right as this document.
  • Map each top query to a flow — what the bot asks, which system it checks, what it sends back.
  • Connect the data — Shopify orders, courier tracking, your returns app.

One rule if the bot runs on WhatsApp. Meta’s WhatsApp Business Platform terms (section 4.7, last modified 23 September 2026) forbid letting WhatsApp data be used to create, train or improve AI systems — the exception is fine-tuning a model for your exclusive use. In my reading, a person reading old chats to write your answer document is ordinary support work — but the terms also cover “derived” data, so ask your provider in writing what happens to your conversations and anything built from them. Handing a chat export to a vendor to train their shared model is not allowed.

Roll it out in stages

Move up a stage when the checks pass, not when the calendar says so. The check: a full day of bot chats with no wrong order status, no wrong policy and no missed handoff.

  1. Test (week 2). Your team plays customer — in English, Hindi and Hinglish, with typos and half-sentences. Fix every wrong answer.
  2. 20% of chats, about three days. The bot gets one chat in five at random. A person reads every bot reply.
  3. 50%, about four days. People review every chat the bot closed on its own.
  4. 80%, about a week. People see only the escalations, plus a daily sample.
  5. 100%. The bot sends the first reply on every chat; people take the escalations.

Escalation rules — write them before day one

Hand the chat to a person when:

  • The customer asks for one, in any words: “agent”, “human”, “call me”, “kisi se baat karao”.
  • Money is involved: a refund, a payment taken with no order, a COD charge.
  • An item is damaged, wrong or missing — after the bot has collected the photos, the same evidence you need to fight return fraud.
  • The customer is angry: abuse, capitals, a complaint threatened in public.
  • The bot has failed twice on the same question.
  • The same order comes back for the second time in two days.

And three rules for how:

  • Say it’s a bot in the first message. A bot that pretends to be a person loses the customer the first time it gets something wrong.
  • Pass the whole chat to the person, so the customer never types the order number twice.
  • Mind the 24-hour window. On WhatsApp, free-form replies are allowed only within 24 hours of the customer’s last message (Meta). If a person picks up the escalation more than 24 hours after the customer’s last message, all you can send is an approved template. Staff the handoff queue for the hours your customers write, and outside them have the bot say when a person will reply.

Hindi and Hinglish: test it, don’t trust the language list

Your customers don’t write in one language. They write “mera order kab aayega”, “COD available hai?”, “size chart bhejo” — Hindi in English letters, mixed with English words. Some write in Devanagari; in the South, Telugu or Tamil, often in English letters too. A vendor’s language list tells you whether Hindi is supported. It does not tell you whether the bot understands “parcel abhi tak nahi aaya”.

  • Pull 50 real Hinglish and regional-language messages from your export and put them in the week-2 test set.
  • Reply in the language and the script the customer used.
  • Write your top 20 answers in Hindi as well, so the bot isn’t translating your return policy on the fly.
  • Add Hinglish phrases to the handoff triggers: “baat karni hai”, “refund nahi mila”, “galat product”.
  • In the weekly review, read the Hinglish chats separately. That’s where the wrong answers hide.

The weekly review — 30 minutes, every week

A bot goes stale because your store changes: a new collection, a new courier, a Diwali return window.

  1. Read 20 chats the bot closed alone, and every chat it escalated.
  2. List the questions it couldn’t answer, and add the answers to your document.
  3. Track four numbers week on week: chats handled, the share handed to a person, repeat contacts on the same order, and complaints about the bot.
  4. Update the answers before a sale or a policy change goes live, not after.
  5. Check the WhatsApp bill against Meta’s free tier of 1,000 service messages a month per number.

What it costs, as of November 2026

Two tools as worked examples — Interakt for WhatsApp, Tidio for website chat. Others price differently; I compare more of them in my AI chatbot tools post.

WhatPrice, as of November 2026Notes
WhatsApp service replies, Meta India rateFirst 1,000 a month per number free, then ₹0.115 eachSince 1 October 2026; before tax (Meta’s INR list rates, linked from its pricing page)
WhatsApp utility templates, Meta India rate₹0.115 eachCharged even inside the 24-hour window since 1 October 2026
Interakt Growth₹2,799 a month + taxesService replies ₹0.128 and utility templates ₹0.150 each on this plan (Interakt)
Interakt AI agents₹2,499 a month + taxes, 100 AI messages included, then ₹0.50 eachNeeds Growth or Advanced. The same page also lists a ₹3,000 AI add-on for those plans — ask which applies (Interakt)
Tidio Free$0Website chat (Tidio)
Tidio Starter$29 a month, or $24.17 a month billed annuallyBilled in US dollars (Tidio)
Tidio Lyro AI agentFrom $39 a month, or $32.50 a month billed annuallyFirst 50 Lyro conversations free (Tidio)

Take an example store doing 3,000 orders a month on Interakt Growth with AI agents. Say it gets 2,000 customer chats a month at four replies each — 8,000 replies, 5,000 of them from the bot — and sends three shipping updates per order:

  • Interakt Growth: ₹2,799
  • AI agents: ₹2,499
  • Bot replies past the first 100: 4,900 × ₹0.50 = ₹2,450 (assuming each bot reply counts as one AI message — confirm with Interakt)
  • Service replies past the free 1,000: 7,000 × ₹0.128 = ₹896
  • Shipping updates: 9,000 × ₹0.150 = ₹1,350

That’s about ₹10,000 a month before GST, or about ₹10,500 if the ₹3,000 add-on applies instead. These are round illustrative numbers, not a client’s bill. Set them against the hours your team spends typing order-status replies today — only your own export can tell you that.

When not to bother

If your inbox gets a few dozen messages a day, don’t buy an AI agent yet. Answer them yourself, send shipping updates automatically, and use quick replies in the WhatsApp Business app — here’s which WhatsApp setup fits at which size. The bot earns its bill when you’re doing hundreds of orders a month and your team answers the same questions every day.

If that’s you and you want the bot, the escalation rules and the weekly review set up properly, that’s what our AI and automation work for ecommerce covers. Book a strategy call — 30 minutes with my team, and if your volume doesn’t need a bot yet, we’ll say so.

Share
LinkedInX

The service for this problem

Intelligence

AI & Personalisation

Smart recommendations & predictions

Operator notes

30 minutes, free

Apply this note to your own store.

A 30-minute call — we'll map what works for your store, not a generic playbook. No sales pitch.

See all success stories

Calls Monday to Friday, in Telugu, English, Hindi or Kannada.

30-minute call, free