AI chatbots for ecommerce: what to automate and when to hand over

Product questions versus order questions

Product and policy questions are the same for everyone: what sizes are available, how long shipping takes, what the returns policy says. A chatbot can answer these from approved content.

Order questions are about one customer: where is my parcel, can I change my address, why was I charged twice. These involve personal information and often need a person. Treat them as a separate, more careful part of the design.

Use approved knowledge sources and keep them current

A chatbot should answer from content you have approved: your shipping, returns and warranty policies, product pages and a short list of common questions. It should not invent policies or make promises your store cannot keep.

Platform tools increasingly generate suggested answers. Shopify Inbox can suggest instant answers based on store policies (opens in a new tab), and Shopify says merchants must review generated content for accuracy before publishing. Apply the same rule to any chatbot.

Assign someone to update the chatbot's sources whenever a policy, price or shipping arrangement changes. Out-of-date answers are one of the most common chatbot problems.

Verify customers before showing order details

Never show order information to someone just because they typed an order number. Order lookups should require verification, such as an order number together with the email address used, or signing in to an account.

Shopify's documentation describes this approach: its track-my-order answer uses an order number and email, and customers must sign in to review an existing order. If your chatbot connects to order data through another tool, ask how it verifies the customer before revealing anything.

Define when to hand over

Handover rules decide whether customers feel helped or trapped. business.gov.au notes that if an AI tool cannot solve a customer's query, it can refer them to a person (opens in a new tab). Make that easy and visible.

• The chatbot is uncertain or the question is outside its approved sources
• The customer is unhappy, makes a complaint or asks for a person
• The request involves a refund, a payment problem or a change to an order
• The message mentions safety, injury or a faulty product
• The customer asks something the chatbot is not permitted to answer, such as legal or medical questions

When handing over, pass the conversation so far to the person, and tell the customer honestly when they can expect a reply. Do not promise instant responses outside business hours.

Be open that it is AI and protect personal information

The OAIC's guidance on commercially available AI products (opens in a new tab) says public-facing AI tools such as chatbots should be clearly identified as AI. Label the chatbot plainly so customers know when they are talking to software.

The same guidance recommends against entering personal information into publicly available generative AI tools. Ask any chatbot provider where conversations are stored, who can access them and whether they are used to train models. Our guide to AI and privacy for Australian businesses explains the questions to ask.

Pre-launch test checklist

Test with realistic questions before customers see the chatbot, and repeat the tests after any major policy or product change.

• Common shipping, returns and product questions return answers that match your approved policies
• Questions outside the approved content lead to a polite handover, not a guess
• Order lookups require verification and reveal nothing to an unverified user
• Complaints, refund requests and safety issues reach a person
• The chatbot is clearly labelled as AI
• Out-of-hours handovers tell customers honestly when to expect a reply
• Someone can switch the chatbot off quickly if it misbehaves

Write the chatbot's scope down

Before setting anything up, write a one-page scope. List the topics the chatbot may answer, the sources it may use, the topics it must always hand over and the hours a person is available to take handovers.

This document becomes your test plan and your guide when something goes wrong. It also makes conversations with a specialist or software provider much clearer, because you can ask whether their tool can follow your rules rather than adapting your rules to their tool.

Review the scope after the first few weeks using real conversations. Questions the chatbot could not answer may point to missing content; repeated handovers on the same topic may point to a policy that customers find confusing.

Next steps

If you are not sure a chatbot should be your first project, compare it with other options in which ecommerce tasks to automate first. When you are ready to plan, discuss your chatbot workflow with an ecommerce specialist or see how the free consultation works.

If your chatbot will eventually answer stock questions, read the ecommerce inventory integration checklist first, because accurate answers depend on accurate stock data.