Guide
AI Chatbot for Ecommerce: How It Works and What It's Worth
How an AI chatbot for ecommerce is grounded, where it earns its keep, how to measure it honestly, and when a chatbot is the wrong tool.
8 min read
read
·
An AI chatbot for ecommerce is a website or messaging assistant that answers shopper questions from your catalogue and policies instead of a scripted menu. It pays off, through assisted conversions and deflected tickets, when you have enough traffic to generate real questions, clean product data to answer them from, and a person to hand off to.
How an AI chatbot for ecommerce works
Four parts. The first is where most of the quality comes from.
Grounding in your catalogue and policies
The model does not know your products. It is given them: the product feed (titles, variants, stock, size charts, materials), the shipping and returns policies, the FAQ, and often a sample of past tickets. When a shopper asks something, the system retrieves the relevant product and policy text and generates an answer constrained to it. That is the difference between "the medium fits a 38 to 40 inch chest" and a confident guess. If the size chart is a JPEG nobody has updated, the bot will be confidently wrong.
Intent and context
Each message is classified: pre-sale question, order status, complaint, return request, or "I want a human". Good implementations also see context: the page the shopper is on, what is in the cart, whether they have ordered before. "Does this fit my car?" means nothing without the product page it was asked on.
Handoff
Every bot needs a rule for when to stop. The usual triggers are a low confidence score, certain words (refund, chargeback, allergic, damaged), a repeated question, or the shopper asking for a person. What matters is what travels with the handoff. A transcript plus the cart and order number, landing in Gorgias, Zendesk, Shopify Inbox or Slack, means the person picks up mid-conversation. A bare "a customer needs help" means the shopper repeats themselves, which they hate.
Channels
The same grounded model can sit behind a website widget, Instagram DMs, Messenger and WhatsApp. The website is where pre-purchase questions happen. Instagram and Messenger are where questions arrive after someone sees an ad, and where the reply expectation is minutes. WhatsApp carries order status and returns where it is the default channel. The Instagram chatbot page covers the DM side, and the wider picture, including voice and SMS, is in the guide to conversational AI for ecommerce.
Rule-based chatbot vs AI chatbot vs AI sales agent
Three products get called "chatbot". They differ in what they can answer and what they do without being asked.
Capability | Rule-based chatbot | AI chatbot | AI sales agent |
|---|---|---|---|
Answers off-script questions | No, falls back to "I didn't understand" | Yes, from grounded content | Yes, from catalogue and policies |
Uses catalogue and stock | Only through hard-coded buttons | If connected | Yes, per product and variant |
Intervenes at exit intent | No | Rarely | Yes, on the product page and at exit |
Handles objections | No | Partly, if asked directly | Yes, "does it run small", "is it worth it" |
Hands off to a person | To a form or inbox | To an inbox, often without context | To a person with transcript and context attached |
Attributes revenue | No | Rarely | Every sale it closes tied to the conversation |
Setup effort | Days of flow building | Hours to connect data, then tuning | Hours to connect data, then tuning |
Typical cost basis | Flat monthly or per seat | Per conversation or tiered monthly | Tiered monthly, judged on attributed revenue |
The middle column is what most brands mean by "AI chatbot": an answer engine that waits to be asked. The right-hand column adds the behaviour of a sales assistant, covered below.
Where it earns its keep
Pre-purchase questions
Sizing, shipping to a specific country, compatibility, "is the navy more blue or more grey". These are the questions that end sessions when they go unanswered, and they happen on the product page at 10pm.
The Classic runs about half a size small. Most people who wear a 9 in Nike take a 9.5 here. Exchanges are free if it's not right.
Order status
"Where is my order?" is the largest ticket category for most stores. With an order lookup by email or order number, the bot resolves it without a person, including the first step of "it says delivered but I don't have it".
Returns triage
Not processing the return, triaging it. Checking the window, explaining the policy, starting the portal flow, and flagging exceptions (damaged, wrong item sent) to a person who can offer a replacement instead of a refund.
Lead capture
When the question cannot be answered now (out of stock, custom order, a B2B enquiry), the bot collects contact details and the specific need, so the follow-up goes out with context.
Upsell at cart
Someone with a tent in the cart asking about pegs gets the footprint suggested. Relevant and rare, or people learn to close it.
After-hours coverage
Much DTC traffic arrives when nobody is on shift. The bot answers the questions that would otherwise become morning tickets or, more often, lost sessions.
What it's worth: how to measure it
Vendors quote ROI figures. Ignore them and run your own, with four measures.
Assisted conversion rate. Conversion rate of sessions with a chat against sessions without, same landing pages, same traffic source, same week. Beware selection bias, since shoppers who chat are already more engaged. The clean version is a holdout: hide the widget for a share of traffic and compare.
Tickets deflected. Conversations that ended without a handoff, tagged by intent, set against helpdesk tickets per thousand orders before and after launch. Watch for deflection that is really abandonment, which shows up as the same shopper contacting you through another channel.
Response time. Median time to first useful answer, especially outside business hours and on Instagram.
Revenue per conversation. Orders placed by the same session or email within a fixed window after a chat. Pick the window (24 hours, 7 days), publish it, do not move it. The conversion attribution guide covers the trade-offs.
Then the cost side: the tool, plus the hours someone spends each week reviewing transcripts and fixing answers. If assisted revenue and deflected tickets do not clear that within a quarter, it is a support tool, not a revenue tool, and should be budgeted as one.
When a chatbot is the wrong tool
Low traffic. Under a few hundred sessions a day you get a handful of chats and no signal from the measures above. Write a better FAQ and use Shopify Inbox or a Tidio-style free tier for the occasional question.
Complex B2B quoting. Custom pricing, minimum order quantities, approvals, samples. A bot can qualify and collect, it cannot quote. Route to a person early.
No catalogue data to ground it in. If sizing lives in a PDF and stock lives in someone's head, the bot will answer wrong and correcting it will cost more time than it saved.
Brands that will not let it hand off. If the rule is "never bother the team", the bot becomes the wall customers hate, and every measure above moves the wrong way.
Regulated claims. Supplements, medical devices, anything where an efficacy answer needs sign-off. Restrict it to logistics and policy questions, or keep a person in the loop on product answers.
Chatbot or AI sales agent?
A chatbot waits to be asked, answers, and ends the conversation. For order status and returns triage that is all you need. An AI sales agent is the same grounded model pointed at conversion rather than deflection. It opens with a relevant prompt on the product page ("choosing between the 40L and the 55L?"), steps in at exit intent with the objection most likely to be stopping the sale, handles "is it worth it" with the catalogue and reviews rather than a discount, and ties each order it influenced to the conversation.
Exerta's on-site sales agent is an example of the second kind. It answers sizing, shipping, "does it run small" and "is it worth it" from the brand's catalogue and policies on the product page and at exit intent, hands edge cases to a person with the context attached, and attributes every sale it closes to the conversation. The AI sales agent guide covers the mechanics and when the simpler chatbot is the better buy.
Setting one up in a week
Day one. Pull the last 500 tickets and DMs from your helpdesk, tag them by intent, take the top 20 questions. That is the bot's job description.
Day two. Connect the product feed, size guides, shipping and returns policies, and the FAQ. Delete anything out of date first, because the bot will quote it.
Day three. Write the handoff rules (trigger words, confidence threshold, hours) and where the handoff lands, transcript attached. Test with 50 real tickets and score the answers yourself.
Day four. Soft launch to a share of traffic, or on one channel. Website first, Instagram once the answers are stable.
Day five. Read every transcript. Fix wrong answers at the source (the product data, not the prompt) and add the missing questions.
Days six and seven. Full launch. Set up the holdout, the ticket baseline and the attribution window first, or you will never know what it was worth.
FAQ
What is the best AI chatbot for ecommerce?
There is no single best one. For support deflection on Shopify, helpdesk-native tools (Gorgias, Zendesk, Shopify Inbox) are the safe choice. For low-traffic stores, a Tidio-style free tier. For conversion on the product page and at exit intent, an AI sales agent. Pick by the job, then test on your own tickets.
How much does an AI chatbot for ecommerce cost?
From free tiers on helpdesk and widget tools to monthly plans priced by conversation volume, seats or attributed revenue. The tool is rarely the largest cost. Budget for the hours someone spends each week reviewing transcripts and fixing product data, which is what keeps answers accurate.
Does an AI chatbot work with Shopify?
Yes. Most tools connect to the Shopify catalogue, inventory and order data through the app store, which is what allows order lookups and stock-aware answers. Check that variants, size guides and metafields come through, not only titles and prices, or the bot cannot answer fit questions.
Can an AI chatbot handle returns?
It can triage them: check the window, explain the policy, start the portal flow and collect the reason. Whether it should complete refunds without a person depends on your margin and fraud exposure. Most brands let it handle exchanges and route refund requests above a threshold to a human.
Should shoppers be told they are talking to a bot?
Yes, at the start and plainly. Shoppers forgive a bot that says it is one and hands off cleanly. They do not forgive one pretending to be Sarah from support. Disclosure also sets the right expectation for what it can do, which reduces the frustrated "just give me a person" messages.
Other useful guides
Get started today
If order status and returns triage are the problem, a grounded chatbot with a clean handoff will do. If sessions are ending on the product page with a question unanswered, read the AI sales agent guide or book a demo to see one on your own catalogue.
