A refund request is more than a message. Someone has to check the order, gather the facts, make a decision, and complete the next step. AI customer service should help with that work too.
SmartReply turns the request into an action. One AI talks to the customer and prepares the task. Your team can make the decision, or a second AI can decide under your rules. Your agents can spend less time carrying each request through a long chat.
What is an AI customer service action?
An action is a task tied to a customer request. A refund action, for example, puts the request and its needed facts in one place. It gives the next person or AI a clear decision to make.
In a manual ticket workflow, an agent often has to read the chat, ask more questions, check the order, and do the task. SmartReply can do that first part and hand the team a prepared action. This changes the agent's job from working through every message to reviewing the cases that need a decision.
For a refund, the action may be named refund_order. The name describes the task. It does not mean a refund has already happened. The action still has to pass the right checks and complete.
Example: an item arrives damaged
- The customer asks for help. They say, “My item arrived broken. Can I get a refund?”
- The conversation AI checks the request. It finds the order and asks for a photo if that is required by the brand's policy.
- The AI prepares a refund action. The order, reason, and evidence stay with the request.
- The action gets a decision. A person can approve or decline. With AI approval enabled, a second AI checks it against the brand's rules.
- The approved action runs. The result is checked. The customer gets an update based on what actually happened.
Your team does not have to start the conversation again just to review the action. They can look at the gathered facts and make the call. If something is unclear, they can take over and help.
Keep customer updates short. Before a decision, “I have your photo. Your refund request is ready for review.” After a confirmed refund, “Your refund has been issued.” Do not tell a customer that money has been sent while the action is still pending or has failed.
Choose team approval or AI approval
Team approval: AI prepares, a person decides
This is useful when you want a person to make the final call. The AI handles the customer conversation and gathers the needed details. A teammate reviews the action and clicks to approve or decline. The teammate can still open the conversation if they need more context.
This can save work even before you allow AI to make decisions. The person is reviewing a prepared task rather than sorting through each message and asking for the same details again.
AI approval: a second AI decides under your rules
The approval AI has a different job from the conversation AI. It checks the pending action, order details, customer history, and the rules you set. It can approve, decline, or send a case for manual review.
You choose where to use it. Keep people involved for cases that need their judgment. Start with a common request type and expand after checking the decisions. Available actions depend on the store and services you connect.
Set rules before you turn on AI decisions
Write your policy in plain language. Include the facts the AI needs, the limits it must follow, and when a person should review the case. A clear rule is easier to test than a vague instruction such as “keep customers happy.”
- Evidence: What does this request need? Is a photo required? What happens if it is missing?
- Order details: Which dates, items, or order values affect the decision?
- Exceptions: When should a person make the call?
- Customer updates: What should the customer hear while waiting and after the action completes?
SmartReply's action approval test mode shows what the approval AI would decide without approving or declining the action. Use it to check common cases and edge cases. Test mode is a way to review decisions before live use. It is not proof that every future case will be handled correctly.
Keep your sources current too. If a return window changes, update the policy the AI uses. Our knowledge base guide explains how to organize those sources.
Watch for pending actions and failed steps
An action can be ready for review and still leave the customer waiting. Set up notifications so the right people know a decision is needed. Add reminders for actions that remain pending.
SmartReply supports action notifications through channels such as Slack, Gmail, and Telegram. You can allow authorized recipients to approve or decline through supported replies or controls. Set the recipients and permissions before relying on those decisions.
If an action fails, the request still needs attention. Check the error, route the case to the right person, and keep the customer updated. An approved request is not the same as a completed task.
Learn the exact controls in the action notification guide. Carrier outreach has its own setup. A carrier accepting delivery instructions does not prove delivery or finish a separate refund.
Test five cases before launch
Use test orders and made up customer details. Try the same cases in team approval and AI approval modes. Check the whole result, not just whether the reply sounds good.
- A request that fits the policy. Are the facts right, and is the decision supported?
- Missing evidence. Does the AI ask for what it needs before treating the action as ready?
- A request outside the policy. Does it follow the decline or review path you chose?
- A case that needs a person. Does the right teammate receive the facts and the review request?
- An action that fails. Is the failure visible, and does the customer avoid getting a false success message?
Download the AI action test worksheet. Record the expected decision before the test. Then add the actual decision, action result, and customer update. Do not treat a sample test as a real payment or claim.
Measure how much time your team spends per resolved case. Also watch wrong decisions, repeat contacts, pending actions, and customer satisfaction. A lower message count is useful only if customers still get the help they need.
Start with the work your team repeats most
Bring a common request to a SmartReply demo. Ask to see the conversation, the action, the decision, and the completed result. Try a policy exception too. That shows where a person still needs to help.
Use the same standard for Gmail, website chat, and connected social channels. SmartReply is built for customer service across those channels. Confirm the actions you need with your own connected store.
Read the main AI customer service guide for setup, use cases, and costs. Explore SmartReply AI customer service to see how the work fits together.
For the full setup steps, see the SmartReply AI action approval guide.