AI chatbots for ecommerce: which questions should they answer?

A chatbot can help shoppers find information outside service hours. It becomes counterproductive when it invents stock, delivery timing or a return exception.

Start with questions customers actually ask

List repeated requests: where is my order, which size fits, when will delivery arrive, how do I return an item? Connect each answer to an updated store page. Shopify Inbox can use catalog, policy and published storefront content. Check those sources against the real offer before enabling automated answers.

Define limits and handoff

Uncertain sizing, complaints and failed payments need a route to a person. Do not let a bot promise a discount or change an order without controls. Shopify documents conversation management and staff involvement. Test both languages when a store serves more than one market; a correct English answer does not guarantee a natural French one.

Measure actual resolution

Review anonymized conversations: was the answer correct, sufficient and clear? Track handoffs, repeated questions and complaints rather than just chats handled. For simple questions, Shopify instant answers may be enough without generating a different response each time.

Start with ten approved questions and expand only after reviewing mistakes.

A short acceptance test

Ask the bot a question answered by the return policy, then an intentionally ambiguous question about a product with no dimensions listed. The first answer should point to the exact rule; the second should request clarification or offer human help. Repeat with an out-of-stock item and a delivery address outside the service area. Keep wrong answers in a correction log and update the knowledge source before expanding the bot’s scope.

Define what the bot knows and when it hands off

Classify a week's customer requests. Order tracking, sizing and box contents may be suitable for an assistant when its data is reliable and access is appropriate. Disputes, safety issues, legal questions and complex customer-specific cases need a path to a person. Do not let a model improvise discounts or warranty promises.

Illustrative example: “Does this filter fit my model X?” The bot should consult a verified compatibility table. If X is not listed, a good answer is “I cannot confirm that” with a human contact option. A plausible but wrong answer can create a return and damage trust. The knowledge base matters more than a cheerful bot persona.

Test before opening the channel

Prepare twenty anonymized real requests: ten straightforward, five ambiguous, three with missing information and two designed to push the bot beyond its scope. Score each reply for accuracy, source, invention and appropriate handoff. Then test changing data such as price, stock and shipping status. If the bot cannot obtain current values, it should send the customer to the right page or team.

Track conversations resolved without correction, transfers, complaints and affected orders. A high automation rate is no success if customers must contact support again to repair a bad answer.

Define the questions it can handle

Start with a narrow scope: support hours, order tracking through an authorized tool, verified product facts and delivery steps. Give each answer an up-to-date source and review date. The chatbot should say when information is unavailable rather than inventing stock, a warranty or return exception. Keep a visible path to a person for disputes, refunds and unusual cases.

Test twenty real support questions and ten difficult cases: unavailable variant, unsupported country, expired code, lost parcel and personalized item. Compare answers with the current policy and public product page. Record correct, incomplete and dangerously false responses. If the bot repeatedly fails on one topic, remove it from scope until fixed.

Check what changes after launch

Assign someone to update information when prices, delivery or rules change. Review what personal data the tool receives, how long it is retained and who can access it. Do not send customer conversations to an outside service without checking relevant permissions and obligations.

Measure useful answers, handoffs to people, repeated questions and reported errors. A bot that deflects requests without resolving them can reduce visible tickets while increasing frustration. The goal is an accurate answer or quick handoff, not a long imitation of support.

A support-bot boundary test

Write ten real customer questions, including ordinary delivery queries and harder cases about a damaged order, a missing parcel, a return outside the stated window and a product compatibility claim. For each, record the approved source and whether a person must review the answer. Test the bot on a product that has just changed price and one that is out of stock. Save its full replies and links.

Mark a failure whenever the bot invents a policy, promises a refund it cannot authorize or hides a route to human support. Revise the knowledge source and escalation rule, then retest the same cases. After launch, review a sample of conversations every week, especially unresolved ones. A quick automated response is useful only when the buyer can act on accurate information or reach someone who can resolve the issue.