AI customer service for ecommerce: what it should handle, what it should not, and how to set it up
UPDATED 7 MIN READBY THE SCENOX COMMERCE TEAM
SHORT ANSWER
AI customer service works well for repeat, factual questions in an online store: order status, delivery times, returns policy, sizing and product details. It works badly for refund disputes, upset customers and unusual cases. Connect it to your real order and product data, design a clean handoff to a person, and measure resolution, not just deflection.
ON THIS PAGE
Most store support inboxes are dominated by a small set of questions: where is my order, how do I return this, will this fit, is it in stock. Those answers already exist in your order system, your policies and your product pages. AI customer service is software that reads that information and answers in plain language, at any hour, in the customer's channel. It does this well within limits, and badly outside them. This guide covers where the line is, how to set one up, and how to judge it without flattering yourself.
What AI customer service is, in practice
In an ecommerce store, it is usually a chat assistant on your site, in WhatsApp or in email, backed by a language model. It is given three things: your policies and product information, live access to order and customer data, and rules about what it may and may not do. When it cannot answer or should not, it passes the conversation to a person.
The difference from old rule-based chatbots is that customers can phrase things freely. The difference from a human is that it only knows what you have given it, and it does not have judgement or accountability.
What it should handle
These are factual, high-volume and low-risk, which is why they are the right place to start.
Order status and tracking
"Where is my order" is often the single largest category. With a connection to your store and carrier tracking, the assistant can look up the order, share status and the tracking link, and explain what a status means. Without a live connection it can only guess, which is worse than no automation.
Delivery and shipping questions
Costs, regions, estimated delivery times, cut-off times and international shipping, all taken from your published shipping policy. The assistant should quote your policy, not estimate.
Returns and exchanges
It can explain the return window and conditions, check whether an order is eligible, and start a return where your system supports it. Setting an exchange in motion is a good use. Deciding a borderline case is not.
Sizing and fit
With size charts and product data loaded, it can recommend a size from the measurements a customer gives. Be careful here: fit advice that is wrong costs you returns, so keep the guidance conservative and based on your own charts.
Product questions
Materials, dimensions, compatibility, care instructions, what is in the box. The quality of these answers depends on the quality of your catalog data, which is the reason clean product data matters before you automate anything.
Stock and basic account questions
Whether an item is in stock, how to reset a password, how to change an address before dispatch, how to apply a discount code.
What it should not handle
Keep a person in charge of:
- Refund disputes and chargebacks. Money and goodwill are at stake, and rules often need interpretation.
- Damaged, wrong or missing items. The customer is already annoyed and may need an immediate remedy.
- Angry or distressed customers. A scripted tone makes it worse. Detect frustration and hand off early.
- Legal, safety and medical questions. Product safety, allergies, warranty claims and regulatory matters should not be answered by a model improvising.
- Exceptions and special offers. Anything that means bending policy, giving discounts or making promises.
- Account security and payment problems. Changing details, suspicious orders, failed payments.
- High-value or wholesale customers. The relationship is worth more than the saved minutes.
A useful test: if a wrong answer could cost you more than the time saved on a right one, it should not be automated.
What data it needs
An assistant is only as good as what it can see.
| Data | Why it needs it | Source |
|---|---|---|
| Orders and fulfilment status | To answer status and tracking questions | Your store platform, carrier tracking |
| Product catalog | For product, sizing and stock answers | Your store, clean product data |
| Policies | Shipping, returns, warranty, payment | Your help pages, written clearly |
| Size charts and guides | For fit advice | Your product content |
| FAQs and past tickets | To mirror real questions and your tone | Support inbox history |
| Escalation rules | To know when and to whom to hand off | You define them |
Two practical points. First, write down policies that exist only in your head, as the assistant cannot follow an unwritten rule. Second, be careful about customer data: give it only what it needs to answer, and check how your provider stores and uses conversations. We do not use client data to train third-party models.
Setting it up, step by step
- Pull a month of support tickets and sort them. Count by type. You will find the top five reasons account for most of the volume.
- Choose what to automate first. Start with two or three categories from the list above, usually order status, returns policy and product questions.
- Write or tidy the source content. Shipping, returns and sizing pages should be complete and consistent. Fix contradictions between pages before the assistant repeats them.
- Connect live data. Link the store for orders and the catalog for products. Test with real order numbers.
- Set the rules. What it may do (look up, explain, start a return), what it may never do (promise refunds, offer discounts, give legal advice), and the tone.
- Design the handoff (see below).
- Test with real questions. Run 50 to 100 past tickets through it and read every answer. Include rude messages, vague ones and questions it should refuse.
- Launch to a slice. One channel or a share of traffic first.
- Review conversations weekly and fix the source content when answers are wrong.
Designing the handoff
The handoff is where most automation is won or lost. Good design:
- Offer a person early. A visible "talk to a person" option, and automatic escalation when the customer asks twice, uses strong language or the assistant is unsure.
- Pass the context. The human should see the full conversation, the order details and what the assistant already tried, so the customer does not repeat themselves.
- Set honest expectations. If your team is offline, say when someone will reply and collect an email or phone number.
- Route by type. Returns issues to one person, payment issues to another.
- Do not trap people. A loop of "I did not understand, please rephrase" is the fastest way to lose a customer.
Measuring it honestly
The easiest number to report is deflection: the share of conversations that did not reach a human. It flatters the system, because a customer who gave up is also deflected. Measure instead:
- Resolution rate. Of conversations the AI handled, how many ended with the issue solved, judged from customer confirmation or no repeat contact.
- Handoff rate and reasons. Why people were escalated, which tells you what to fix.
- Repeat contact. Customers who write again about the same issue within a few days.
- Satisfaction on AI conversations, compared with human ones.
- Response time and out-of-hours coverage. Often the clearest real gain.
- Wrong answers. Count them from your weekly review. Even a low rate matters if the errors involve money.
- Time saved by your team, estimated from actual ticket counts, not from a vendor's claim.
Read a sample of full conversations every week. Numbers will not show a polite, confident, wrong answer.
Risks to plan for
- Confident wrong answers. Limit the assistant to approved sources and require it to say when it does not know.
- Promises you cannot keep. Block it from offering refunds, discounts or delivery dates not drawn from your systems.
- Stale content. When policies change, update the source the same day.
- Tone mismatch. Write a short style guide and test against it.
- Disclosure. Tell customers they are talking to an automated assistant, and check requirements that apply in your market.
When to do it yourself and when to hand it off
If you get a few dozen tickets a week, a clear set of help pages, saved replies and a well-written FAQ will often do as much good as an AI assistant, at lower risk. Start there, and consider automation when volume or out-of-hours coverage becomes the problem.
If volume is high, questions are repetitive and your policies and catalog are in reasonable shape, setup is worth doing properly. Our AI customer support service builds the assistant on your policies and live order data, with handoff rules you approve, and the customer service automation for ecommerce page covers the wider workflow. We give a fixed quote after reviewing your tickets and systems, and you can ask for the free 10-product sample if the project is a fit.
