“Where is my order?” is a good candidate for support automation. The answer should come from an order record and a tracking service. A teammate shouldn’t need to type it from scratch every time.
“Can you make an exception to the return policy?” needs a different process. A confident, friendly reply can still promise something the business hasn’t agreed to do.
Before choosing a support bot, decide which questions it may answer and which actions require a person.
Start with information you already maintain
Order status, delivery links and published policy details are useful first cases. Product facts can work too, provided the underlying catalog is accurate.
Have the system retrieve the relevant record or policy. It should not invent a delivery date, fill a missing measurement from a similar product or improvise a return window.
Check the customer is authorized to see any order or account information before returning it. The convenience of a chat interface doesn’t change that requirement.
For less certain cases, use AI to classify a message or draft a reply for review. Shopify Inbox’s suggested replies are one example of assistance that leaves a teammate in the conversation.
Write the handoff before the greeting
Decide what happens when the system doesn’t have an answer.
The customer needs a clear route to a person, a realistic expectation for response time, and continuity. Pass the conversation and relevant order details along so they don’t have to explain everything again.
Useful handoff triggers include:
- A request for a refund, discount or policy exception
- Conflicting information in the available records
- A product safety question the approved material doesn’t answer
- Repeated attempts that haven’t resolved the issue
- A direct request to speak to someone
An unanswered message sitting in a queue isn’t a completed handoff. Assign ownership and check that someone receives it.
Separate answers from actions
Reading an order status is different from changing an address or issuing money.
Give each action explicit rules and permissions. Decide whether it needs customer confirmation, agent approval or both. Keep a record of the action and the data used to make it.
For a first rollout, it is often enough to let AI prepare a proposed response or action. The team can learn where it helps before granting more authority.
Clean up the source material
A support system quoting an outdated policy can be wrong without inventing a word.
Check that return windows, final-sale exclusions, delivery regions, warranty terms and contact details agree across your site and support tools. Assign owners to those pages.
The same cleanup improves how shopping assistants describe your products. Customers and automated systems both need a consistent answer.
Review real conversations
Start with a narrow question category and read its transcripts regularly. Record errors by cause: missing source, stale information, misunderstanding, failed lookup or a handoff that didn’t reach anyone.
Fix the cause before expanding to more categories. A few successful demonstrations don’t show how the system handles a busy Monday or an unusual order.
Include the support team in the review. They know which apparently simple questions tend to become complicated.
Measure whether the customer got help
A lower ticket count can mean less repetitive work. It can also mean customers gave up.
Compare resolution, repeat contacts, handoff waiting time and customer satisfaction. Look for refunds or complaints connected to an incorrect answer. Review a sample of conversations classified as resolved.
Once a narrow category performs reliably, you can consider automating more of it. Keep an easy way to return to drafts or human handling.
If the difficult part is joining order, warehouse and support data, the project may be a custom integration as much as an AI rollout. Get those records working together before expecting a model to compensate for the gaps.