Setting up the AI agent takes about 10-15 minutes if you know your products well.
1. Add knowledge
The AI agent answers questions based on what it knows. Before it can help customers, you need to give it knowledge.
Navigate to AI Agent → Knowledge and add your store information:
Help articles - Create articles in the Help Center with answers to common questions. The AI will reference these when responding. Good starting articles:
-
Shipping and delivery times
-
Return and refund policies
-
Product care instructions
-
Common troubleshooting steps
Text snippets - Short knowledge pieces covering policies, FAQs, and specific topics. You can also save good responses from conversations directly to AI knowledge.
Files - Upload PDF, DOCX, CSV or TXT documents, up to 10 MB each, in the Files tab.
Guidance is not knowledge. It tells the AI agent what to do in one situation, such as a return request or a late delivery, and you write it in AI agent → Guidance. Rules for every reply, like "Never promise delivery dates", go in General rules (step 4).
The more knowledge you add, the better the AI performs. Start with your most common questions.
2. Configure actions
Actions let the AI do things, not just answer questions. Navigate to AI Agent → Actions and review what's available:
Shopify actions - The AI can look up orders, check tracking status, update shipping addresses, and more. These connect directly to your Shopify store. If you click on Edit on the Shopify action, you can specify condition under which that action will activate.
Custom actions - Create your own actions with a URL and instructions. For example, you can let the AI check a custom loyalty points system.
You can check our 15 Custom Action Templates on this link here.
Toggle on the actions you want the AI to use.
3. Clarifying questions
When the AI doesn't have enough information to help, it can ask the customer a clarifying question.
For example, if a customer asks about their order but doesn't provide an order number, the AI can ask for it.
The AI agent can also ask your team instead of the customer. An Administrator switches on Ask human agent in AI agent → Settings. A teammate's answer is saved as a text snippet in AI knowledge, so the AI agent doesn't need to ask the same question again. What a customer answers is not saved to knowledge.
4. Set AI behavior
Navigate to AI Agent → Settings and configure:
Inbox assignment - Choose which inboxes the AI monitors. The AI only responds to conversations in selected inboxes.
Use pending status - While the AI agent waits for a customer's answer, the conversation is Pending. With If no reply within and Then you choose what happens when the customer doesn't answer: keep it pending, hand it over, resolve it, or send a follow-up (email only). The AI agent resolves a conversation by itself once the request is done. There is no setting to switch that off.
General rules - Short brand and compliance rules the AI agent follows in every reply, for example "Always use British English" or "Never mention competitors by name". You edit them in AI agent → Guidance: click Add or Edit on the General rules card. The field holds 1,000 characters. Keep catch-all handovers such as "hand over when unsure", ticket status, greetings and tone out of it.
Tone of voice - How replies sound: length, formality, emoji, greetings and sign-offs. An Administrator sets it in AI agent → Settings → Tone of voice, separately for Emails & DMs, Social-media comments and Chats, with 1,000 characters each. Browse presets gives you a starting point.
Words to avoid - Words and phrases you don't want in replies are a matter of tone, so list them in Tone of voice. The AI agent takes the list as an instruction. Replies are not checked against it afterwards, so a listed phrase can still appear now and then. Keep the list short, because it shares the 1,000 characters with the rest of your tone text. If replies have to be checked before they go out, switch on Require approval before AI actions. It does not cover social media comments.
Handoffs to humans - To make the AI hand off specific topics to a human, you have these options:
-
Turn on Restrict AI responses to enabled guidances only (AI agent → Settings). On email and social inboxes the AI agent then answers only conversations that match a guidance with the status Enabled. It leaves everything else for your team, and the customer gets no message from the AI agent. Chat and WhatsApp inboxes are not restricted.
-
Create guidance for the topic (e.g., refunds, complaints) and instruct the AI to greet the customer and hand off to a human agent.
-
Hand over without a reply. Tick Hand over silently on email in AI agent → Settings so email handovers send the customer nothing. For one topic on any channel, tell the AI agent in that topic's guidance to hand over without replying.
5. Test before going live
Use the Playground (AI Agent → Playground) to test the AI before it talks to real customers.
Send test messages and see how the AI responds. If a response isn't right, you can:
-
Add or update knowledge to fill gaps
-
Adjust guidance to change behavior
-
Save the correct answer as a text snippet in AI agent → Knowledge → Text
6. Enable the AI
When you're happy with the AI's responses, go to AI agent → Settings:
- Click Select inboxes and pick the inboxes the AI agent should answer.
- Switch on Enable AI agent.
- Click Update settings.
The AI agent now joins new conversations in those inboxes. Conversations that were already there are not picked up. On email it waits up to 3 minutes before it starts on a reply. Chat and DMs are answered right away.
Tips for better performance
Start narrow - Enable the AI for one inbox first. Learn how it performs before expanding.
Review conversations - Check conversations where the AI responded. Look for gaps in knowledge or wrong answers, and fix them.
Use CSAT feedback - If customers rate interactions poorly, investigate what went wrong and improve the knowledge base.
Refine over time - The AI improves as you add more knowledge and examples. Regular maintenance pays off.
What the AI won't do
The AI is designed to be helpful and safe. It will not:
-
Make promises you haven't approved in guidance
-
Take actions you haven't enabled
-
Respond to inboxes you haven't selected
-
Continue a conversation after a teammate replies or is assigned to it. The AI agent doesn't come back by itself. To hand a conversation back, choose the AI agent as assignee.