Configure AI Customer Service

#Start from Shop configuration

Open Shops, select the Shop you want to connect, then choose AI Customer Service. This page controls how Customer Service runs for that Shop. Cross-Shop conversations and escalations are managed from Customer Service in the main navigation.

Configure the page in this order: confirm the service and device first, select the run profile and model, then design business instructions and escalation delivery together. This establishes reliable automation before routing the smaller set of authorization decisions to people.

APP CAPTUREThe Shop-level AI Customer Service settings in the English light interface. Check service status, this-device assignment, Agent permissions, and the Customer Service model from the left navigation.

#1. Confirm that the service is active

The service status at the top shows whether AI Customer Service is available and enabled for this Shop. If it is not active, return to the Shop overview and check the service switch, Shop authorization, and account entitlement.

An active service does not mean that Customer Service is already running. The Shop must also be assigned to a Desktop device.

#2. Bind a device that stays online

Enable Handle Customer Service on this device to assign new conversations to the current Desktop. Keep this device signed in and online, and avoid moving the same Shop between devices frequently.

When migrating to another device, sign in and verify the model on the new device before switching the assignment. Then use the conversation inbox to confirm that new messages are still being handled.

#3. Choose a run profile and model

The Agent run profile controls which Customer Service tools AI may use. Keep the built-in profile for the first launch. Select a custom profile only when your team already maintains explicit tool permissions and behavior constraints.

The Customer Service model inherits the global setting by default. A Shop-specific override is useful when:

• the Shop uses a different language and needs a model that performs better in it;
• high-value orders or complex aftersales cases need stronger reasoning;
• the team must keep the model fixed so a global change does not alter Customer Service behavior.

Model changes affect speed, cost, and decision quality. Evaluate them with Customer Service performance analysis instead of relying on one conversation.

#4. Write operational business instructions

Business instructions are not brand copy. They are the operating rules for this Shop. Prioritize facts and actions that can be verified:

1. Product and fulfillment facts: dispatch times, where delivery status comes from, and how return addresses are obtained.
2. Actions AI may take directly: such as in-policy cancellations, standard returns, or predefined compensation.
3. Decisions that require escalation: such as refunds above a limit, reshipments, exceptional compensation, or conflicting rules.
4. Promises AI must not make: such as an unconfirmed delivery date, stock level, or platform-review result.
5. What to collect when information is missing: ask the buyer for an order number, photo, or problem description before proceeding or escalating.

A useful instruction looks like this:

When delivery is past the promised window but the parcel is not confirmed lost, explain the current status and ask for the latest tracking screenshot. Do not promise a replacement. If loss is confirmed or the buyer requests a replacement, create an escalation so a manager can decide.

It defines what AI should do, what it must not do, and when it should request a human decision.

#5. Configure escalation delivery

In Customer Service escalation delivery on the same page, select a connected channel and recipient. Configure this together with business instructions: instructions determine when to escalate; delivery determines who receives the decision.

APP CAPTUREThe English light-theme advanced AI Customer Service settings. Escalation delivery selects the messaging channel and responsible recipient.

• Feishu is recommended: the recipient gets a structured escalation form and can submit guidance and resolution status directly.
• WeChat or Telegram: the recipient can reply to the escalation notification. Reply to or quote the original message so the escalation request remains in context.
• No delivery route: the escalation remains in the in-app queue, but no proactive channel notification is sent.

After configuration, use a test case to verify that the intended recipient receives the message. See escalation and human review for the full operating procedure.

#6. Configure unpaid-order outreach

If the Shop follows up on unpaid orders, enable proactive outreach here and define the message stages. Start with fewer stages and conservative delays. Confirm the tone, timing, and order-state checks before increasing automation.

Message templates can reference values such as order ID, product count, and Shop name. If an experiment is enabled, retain control traffic and monitor payment progress and experiment maturity. See proactive outreach for unpaid orders for details.

#Go-live checklist

• Service status is active.
• The current Desktop handles Customer Service and will stay online.
• The Agent run profile and Customer Service model are available.
• Business instructions define what AI may handle, must not promise, and must escalate.
• A Feishu, WeChat, or Telegram channel and recipient are configured.
• One routine conversation and one human escalation have been tested.
• If unpaid-order outreach is enabled, its stages, delays, and message templates have been reviewed.

When these checks pass, open the conversation inbox for daily operations.