Ecommerce support requests are surprisingly repetitive. Different customers, different products — but the questions circle the same few themes.
Classify the incoming questions first
Before automating anything, open the last month of support tickets and sort them roughly into four groups:
1. Information questions. "How long does shipping take?", "Is this waterproof?", "What is the return window?" — The answer is fixed and should already be written on your site.
2. Account-specific questions. "Where is my order?" — The answer varies per customer and requires a system lookup.
3. Decision questions. "Which size should I get?", "Which of these two suits me?" — Requires product knowledge and a little guidance.
4. Problem reports. "It arrived broken", "You sent the wrong item" — Requires a human and usually a goodwill decision.
AI handles groups one and three well. Group two is an integration problem. Group four should stay with a person, not only for technical reasons but for the sake of the customer relationship.
Group one: the fastest win
This is usually the largest slice of support volume and the easiest to automate. When the AI crawls your site it already picks up what is on your shipping, returns and product pages.
There is one condition: the information has to exist on the site. A common situation is shipping terms living only in staff members' heads. If it is not written down, the AI cannot know it either. The fix is simple — add it as a knowledge card by hand.
Group three: sizing and product choice
AI does well here, but the setup matters. With page context enabled, the AI can see which product page the visitor is on and answer accordingly. "What are the measurements for this size?" gets answered without the visitor having to name the product again.
Do not skip adding size charts to the knowledge base. In most stores that information lives inside an image — the AI cannot read text in a picture, so you need to enter it as text too.
Group four: getting the handoff right
When someone reports a problem, there is only one thing the AI should do: pass them to a human, quickly and politely. Write it explicitly into the system prompt:
If the customer reports a damaged item, a wrong delivery, a return request or a complaint, do not attempt to resolve it. Apologise, take brief details, and transfer the conversation to an operator.
In these situations an AI "trying to help" usually makes things worse.
Proactive messages: use sparingly
Auto-opening chat boxes become irritating fast. A few uses that genuinely work:
- Stalled at checkout. A visitor idle on the payment page gets "Something not working? Can I help?"
- Lingering on shipping terms. Someone reading your delivery page for a while is probably looking for a specific answer.
- Returning visitor. Someone who has viewed the same product several times is close to deciding.
What does not work: messaging everyone three seconds after they arrive.
Measure it
The way to know whether automation is working is measurement, not intuition:
- Share of conversations the AI escalates to an operator — should fall over time
- Support email and phone volume — should fall
- Sessions that convert to a sale after a chat
- Questions received and answered outside working hours
That last one is the most striking number in most stores: traffic that was previously lost entirely.
A practical order to start in
- Sort the last month of tickets into the four groups.
- Check whether the answers to group one are written on your site.
- Add the missing ones as knowledge cards.
- Write the handoff instruction for group four into the system prompt.
- Read the conversations for the first two weeks and extend the knowledge base.
Do not skip step five — it is the step that determines how good the automation turns out to be.