What is an AI chatbot for ecommerce?
An AI chatbot for ecommerce is a shopping and support assistant connected to your store's catalogue, orders and policies, so it can answer in natural language and act on real data instead of reciting a fixed FAQ.
Older store chatbots worked on buttons: “Track order”, “Talk to us”, “Offers”. They were reliable but useless the moment a shopper typed “will the navy kurta in L fit a 40 inch chest and reach before Saturday in Nagpur?”. A language model can read that sentence. On its own, though, it knows nothing about your kurta, your stock or your courier. The engineering work is connecting it to those facts and preventing it from inventing any.
Done properly, the chatbot sits across the whole shopping journey: before purchase (questions, recommendations), at checkout (doubts about delivery, payment, COD), and after purchase (tracking, returns, exchanges). Each of these reads different data and carries different risks, which is why we treat them as separate modules.
If your enquiries are mostly bulk or wholesale and involve quotes rather than carts, an AI sales agent is the better model. For a store selling to consumers, read on.
How do online stores use an AI chatbot? Five jobs worth automating
Online stores use an AI chatbot for five main jobs: answering product questions, recommending items, tracking orders, handling returns and exchanges, and recovering abandoned carts. A sixth, COD confirmation, is especially useful in India.
1. Product questions
Size, material, dimensions, care, compatibility, warranty. These block purchases when unanswered at 11 pm.
2. Recommendations
“Something for a wedding under my budget” turns into a short list of in-stock items with links, not a generic bestseller carousel.
3. Order tracking
The single most common support question in most stores. Automating it frees staff for real problems.
4. Returns and exchanges
Checking the window, collecting reason and photos, and logging the request consistently by your policy.
5. Cart recovery
Reminding opted-in shoppers about what they left behind, on WhatsApp where they actually read messages.
6. COD confirmation
Confirming cash-on-delivery orders and addresses before dispatch to catch fake or mistaken orders early.
Most stores should launch with jobs 1 and 3, which carry the least risk and the highest volume, then add the others once answers are trusted.
Catalogue-grounded answers: how the chatbot stays accurate
The chatbot answers product questions only from your catalogue data, such as titles, descriptions, variant fields, metafields, size charts and policy pages, and refuses to guess when the data does not cover the question.
We sync the catalogue from your store on a schedule, or on every product update where the platform allows it. Structured fields like price, stock and variant options are read directly at the moment of the question, so the bot never quotes yesterday's price. Long text, such as fabric notes, care instructions and size guides, is indexed for retrieval so the relevant passage can be found quickly.
Size charts deserve special mention. Many Indian fashion stores keep them as images. A language model cannot reliably read measurements from a picture of a table, so we convert charts into data once, with your team checking the numbers. That single step often removes a large share of “will this fit me?” handovers.
Before launch we run a question set built from your real chat and email history, including questions about products you do not sell. Any wrong answer is traced to missing or conflicting data and fixed there. The method is covered in depth on RAG chatbot development.
Product recommendations that respect stock, size and budget
Good chatbot recommendations ask two or three questions, filter your catalogue by stock, size, budget and category, and then let the language model explain the choice briefly. The filtering is done by code; the model only writes the words.
This split matters. If a model picks products freely, it may suggest an item that is out of stock in the shopper's size or outside their budget. When the shortlist comes from a database query, every suggestion is buyable.
Useful signals include the product being viewed, items already in the cart, the occasion the shopper mentions and any preference stated (“no polyester”, “only cotton”). You can also add business rules: prefer items with healthy stock, avoid showing clearance items to someone asking for gifts, or push a new collection first. We never add false urgency such as invented “only 2 left” messages; the stock figure shown is the real one or nothing.
On WhatsApp, Meta's developer documentation says multi-product messages can show up to 30 products from your catalogue, organised into sections, which is ideal for a short list. On the website, suggestions appear as cards with an add-to-cart button.
Order tracking through store APIs: what the chatbot needs
Order tracking works by verifying the customer, fetching the order from your store platform's API, reading fulfilment and tracking details, and translating the status into a plain answer like “shipped yesterday, expected Thursday”.
On Shopify, the Admin API's Order object exposes fulfilments and the fulfilment status, and Shopify's documentation notes it requires the read_orders access scope; only the last 60 days of orders are accessible by default, with older records needing the read_all_orders scope. On WooCommerce, the REST API has order and refund endpoints authenticated with a consumer key and secret. Custom stores need an order lookup endpoint, which we build if it does not exist.
Verification is not optional. The bot should never reveal an address or order contents to someone who merely types an order number. On WhatsApp the sender's phone number can be matched to the order; on the website a one-time code to the registered phone or email does the job.
Where your courier or shipping aggregator offers a tracking API, the bot adds live movement details. Where it does not, the bot shares the tracking link. Delays and failed deliveries are handed to your team with context. For Shopify-specific WhatsApp flows, see Shopify WhatsApp integration.
Return and exchange flows the chatbot can run by your policy
A returns chatbot checks whether the item is eligible under your policy, collects the reason and photos, offers exchange before refund where you prefer, and logs a request your team approves, so every customer is treated by the same rules.
We turn your written policy into explicit rules: return window in days from delivery, categories that cannot be returned, conditions for exchange, and whether a size exchange needs the new size in stock. The bot reads the delivery date from the order, applies the rules and explains the outcome in simple language.
For eligible requests the bot asks for a reason from a short list, photos where damage is claimed, and a preferred pickup slot if you offer pickup. It then creates the return in your store or a shared sheet. Refund amounts and approvals remain with your team; the bot never promises a refund on its own. For exceptions, such as a customer asking for a return outside the window, it hands over politely with the whole history attached.
The benefit is consistency as much as speed. Customers get the same answer at 2 am as at noon, and your team spends time only on the cases that need a decision.
Abandoned cart recovery on WhatsApp and email
Cart recovery sends a timely reminder of the exact items a shopper left behind, only to people who opted in to messages, and stops as soon as they order or ask you to stop.
WhatsApp is usually more effective than email for Indian shoppers, but it has rules. Meta's WhatsApp Business Messaging Policy says businesses may only message people who have shared their number and opted in, and must honour opt-out requests. Messages outside a 24-hour window after the customer's last message must use approved templates, and Meta's pricing page says these have been charged per delivered message since 1 July 2025. So an opt-in checkbox at checkout, worded clearly, is the first requirement.
A common sequence is one reminder after an hour or so and a second the next day. Adding a chatbot makes recovery conversational: if the shopper replies “is COD available?” or “will it come by Friday?”, the bot answers from store data instead of leaving the question unread.
We avoid invented discounts. If you want to offer a code in the second reminder, you set the rule and the code; the bot does not make one up. WooCommerce stores can see the setup on WooCommerce WhatsApp integration.
Should your AI chatbot for ecommerce live on WhatsApp or the website?
Put it on both if you can, with the same brain behind them: the website widget catches shoppers while they browse, and WhatsApp handles post-purchase questions and follow-ups where customers read messages.
Each channel suits different moments. On the website, a shopper is looking at a product and wants an answer before scrolling away; a light widget that opens only when tapped keeps pages fast. On WhatsApp, customers ask about orders days later, send photos for returns and respond to reminders. Meta's catalogue features let the bot share products as native WhatsApp messages rather than plain links.
Page speed is a real concern on the website. Many chat widgets load heavy scripts on every page, which hurts Core Web Vitals and mobile shoppers on budget Android phones. We load the widget after the page is interactive, and only its launcher at first. For stores worried about speed already, Shopify speed optimisation is worth reading.
If budget allows only one channel, choose by where your questions already arrive. Many Indian stores find that most support traffic is already on WhatsApp, so starting there gives the quickest return.
COD confirmation and fewer returned-to-origin parcels
A chatbot can confirm cash-on-delivery orders on WhatsApp before dispatch, asking the customer to confirm the order and address, so obviously fake or mistaken orders are caught before the courier charges you for a failed trip.
The flow is short. When a COD order is placed, the customer receives a utility message with the items, total and address, and two buttons: confirm or cancel. Replies are written back to the order as a tag. Your dispatch team ships confirmed orders first and calls or holds the rest by your rule. If the address looks incomplete, the bot asks for a landmark or the correct PIN code.
Some stores also offer a prepaid switch in the same chat, sending a payment link where the customer prefers to pay online. We connect this to your existing payment setup; we do not change your payment provider.
This is one of the rare chatbot features with directly visible savings, because you can count failed deliveries before and after. It is also a gentle introduction to the bot for customers who later use it for tracking. The WhatsApp ordering system page covers taking orders fully inside WhatsApp.
Does an AI chatbot for ecommerce increase sales? How to measure it
It can, but you should prove it with your own numbers: track chat-assisted sessions and orders as analytics events, compare them against similar sessions without chat, and count recovered carts and support hours saved separately.
Google's GA4 documentation lists recommended ecommerce events such as add_to_cart, begin_checkout and purchase. We send a custom event when a chat starts and tag the session, so you can compare conversion rates of shoppers who chatted against those who did not. That comparison is biased, since shoppers who ask questions are often more serious, so for a cleaner answer we can show the chat launcher to only part of your traffic for a few weeks and compare the two groups.
Other metrics worth tracking:
- Share of support conversations resolved without a person
- Recovered carts: reminder sent, then order placed within a set window
- COD orders confirmed, cancelled and failed at delivery
- Top handover reasons, which point to missing catalogue data
- Average first-response time before and after launch
We set up a small dashboard with these figures, alongside store revenue, so you can judge the bot on results rather than chat counts.
How much does an AI chatbot for ecommerce cost in India?
A custom AI chatbot for ecommerce from BtechWaleTech starts at ₹40,000 (US$600) for catalogue answers and order tracking on one channel; WhatsApp, returns, cart recovery and COD confirmation are quoted as additional modules.
The main drivers are the catalogue and the policies, not the chat interface:
- Number of products and variants, and how much detail sits in images rather than text
- Store platform: Shopify and WooCommerce have documented APIs; custom stores may need new endpoints
- Number of channels: website only, WhatsApp only, or both
- Complexity of return and exchange rules by category
- Courier tracking integration, which depends on what your shipping partner offers
- Languages supported and the review effort for each
Running costs are separate and paid directly: AI usage, WhatsApp template messages and hosting. Compare chatbot budgets more broadly on chatbot development cost in India, or see WhatsApp chatbot price in India for the WhatsApp side. Maintenance is free for two months after launch, then optional from ₹8,000/mo.
The chatbot design stays the same across platforms; what changes is how it reads products and orders and how it writes returns and tags back.
Shopify
Products, variants, inventory and orders are available through Shopify's APIs with scoped access. Metafields often hold size and material details worth indexing. Older orders beyond the default window need extra access approval, so plan for that if customers ask about old purchases.
WooCommerce
The REST API covers products, orders and refunds, secured with consumer keys. Plugins sometimes store important data in custom fields, which we map during planning. Hosting performance matters because the bot calls the API often.
Custom stores
We add read-only endpoints for product search and order lookup if they do not exist, with rate limits and authentication. This adds time but gives full control.
Marketplace-only sellers
If you sell only on large marketplaces, their own messaging rules apply and a separate chatbot is rarely allowed or useful. A direct store changes that; see the best ecommerce platform in India to choose one.
Hindi, Hinglish and budget phones: building for Indian shoppers
Indian shoppers often type in Hinglish, ask in Hindi, and browse on low-cost Android phones with patchy data, so an ecommerce chatbot here must understand mixed language and stay light on the page.
Language models handle messages like “ye saree ka blouse piece milega kya?” well, and the bot replies in the same style. Product names, sizes and prices stay exactly as in your catalogue. For WhatsApp templates, each language version needs separate approval from Meta, and we draft Hindi and English versions for your review. For other Indian languages, we build the flow and you supply or approve the wording.
On the website, the widget loads only its launcher icon at first and fetches the rest when tapped. Replies are kept short, product cards are compressed images, and the chat works even when the connection drops for a moment. Voice notes are common on WhatsApp; the bot can transcribe short ones or ask the customer to type, depending on your preference.
For stores where Hindi is the main language of support, the Hindi AI chatbot page goes into more detail.
Risks and red flags with ecommerce chatbots
The biggest risks are wrong stock or price answers, invented offers, order details shown to the wrong person, and messages that break WhatsApp policy. Good design prevents each.
- Prices or stock taken from a stale copy instead of the live store
- The model allowed to mention discounts or delivery dates you never set
- Order lookup by order number alone, with no phone or email verification
- Cart reminders sent to shoppers who never ticked an opt-in
- Fake urgency such as invented stock counts or countdown timers
- A heavy widget script slowing every page on mobile
- Chat logs containing addresses stored without access control
- No route to a person, which Meta's policy requires when automation replies
Ask any builder how each of these is handled before you sign. A clear answer for each is a good sign; vague reassurance is not.
Worked example: an AI chatbot for a hypothetical ethnic-wear store
Say a hypothetical ethnic-wear brand in Jaipur sells around 600 products on Shopify, gets most questions on WhatsApp about sizes, blouse stitching and delivery dates, and loses a noticeable share of COD orders at the doorstep.
A sensible first phase would convert image-based size charts into data, index product descriptions, and launch a WhatsApp bot that answers size, fabric and stitching questions, tracks orders after matching the sender's phone number, and confirms COD orders before dispatch. Phase two would add a website widget with guided recommendations for occasions and budgets, plus cart reminders to shoppers who opted in at checkout. Phase three would add exchange flows for size issues, checking whether the new size is in stock before offering an exchange.
Each phase would be quoted separately from ₹40,000, with GA4 events set up in phase one so results could be compared from the start. This is a planning illustration only, not a real client or a promised outcome.
Launch checklist for an ecommerce AI chatbot
Tick these before launch and the chatbot will start on accurate data with clear rules.
- Product data complete: sizes, materials and care as text, not only images
- Written return, exchange and COD policies, with exceptions listed
- API access set up for the bot with the minimum scopes it needs
- Customer verification method chosen for order lookups
- WhatsApp Business Platform number owned by the business, display name approved
- Opt-in checkbox at checkout with clear wording
- Templates drafted and approved for tracking, COD and cart reminders
- Handover rules and the team inbox that receives them
- GA4 events for chat start, assisted add-to-cart and purchase
- A test set of 50–100 real customer questions
Send us what you have, even if half the list is open. We reply on WhatsApp, work remotely in English and Hindi, and send an itemised quote in about two working days.