Shopify AI Support Agent Setup Checklist: What to Connect, Which Policies to Write and How to Test Before Going Live

This guide walks you through those five essential steps. It is for merchants who have already picked a tool and want to go live without embarrassing replies or wrong refunds.

What an AI support agent actually needs to do its job

An AI support agent is software that reads a customer message, finds the right order, applies your rules and replies. If you allow it, it also takes the action: a refund, a promo code, a cancellation or an address change.

That means it needs three inputs:

• Data: your orders, customers and products.
• Rules: your policies, written clearly enough for a new hire to follow.
• Permissions: a list of what it may do alone and what it must not touch.

Relay is built around this idea. It works like an autonomous support employee, so the setup looks like onboarding a person: give access, hand over the handbook, set limits.

Step 1: What to connect

Connect only what the agent needs to answer questions and act on orders.

• Your Shopify store. This lets the agent find an order from an email address or order number, see its status, items and shipping details, and act on it.
• The ticket center. Tickets should live in one place, so the agent and your team see the same thread.
• Your knowledge base. Gather your FAQ, sizing guides, care instructions and shipping pages. If you have none, start with your ten most repeated questions.

A quick check before moving on: open a recent order in Shopify, then search for it from a test ticket. If the agent cannot find it, fix the connection now, not after launch.

Step 2: Which policies to write

This is where most of the quality comes from. An agent that follows vague rules gives vague answers. Write each policy as a short rule with a condition, an outcome and an exception.

• Returns and exchanges — return window, item condition, who pays shipping, what is final sale.
• Refunds — when a refund is allowed, full or partial, to original payment or store credit.
• Cancellations — until which order status a customer can cancel.
• Address changes — until which status you can still edit an address.
• Shipping delays — what to say when a parcel is late, when you reship, when you refund.
• Damaged or wrong item — whether you ask for a photo, replace or refund.
• Discount codes — when you offer one, for which situations, with what limits.
• Sensitive cases — topics a human should handle: legal threats, chargebacks, VIP customers, anything unusual.

How to write a rule the agent can follow. Compare two versions. Vague: "We accept returns on most items." Clear: "Returns are accepted within the return window if the item is unused. Final sale items cannot be returned. If the item arrived damaged, offer a replacement or refund without asking for a return." The second version covers the condition, the exception and the outcome. Write them all like that.

Think about tone too. Write down three or four lines on how you speak to customers. Formal or casual? First name or not? Do you apologize first or solve first? Keep it short. A tone note works better than a long brand essay.

Step 3: Set permissions before you test

Permissions decide what the agent can do without asking. Relay applies the permissions you grant, so treat this as a safety dial. A sensible starting point:

• Allow first: answering order status and policy questions, sending tracking information.
• Allow next: address changes and cancellations for orders that have not shipped.
• Allow last: refunds and promo codes, ideally once you have seen the agent apply your rules correctly in tests.

Anything involving money deserves the most scrutiny. You can widen permissions later. It is harder to undo a wrong refund.

Step 4: How to test before going live

Do not test with made-up questions only. Real tickets are messier, and that is what you need to see. Take real past tickets from your inbox and check how the agent would handle them.

Build your test set. Pull 30 to 50 tickets from your past inbox. Make sure they cover:

• Where is my order?
• Return and refund requests, including ones outside your window.
• Cancel or change my order.
• Damaged or wrong item.
• Questions about products and sizing.
• Angry or unclear messages.
• Messages that mix two requests in one email.
• Messages that your own team should handle.

Check each answer against four questions: Did it find the right order? Did it apply the correct policy, including exceptions? Did it take the right action, or correctly refuse to act? Does the reply sound like your brand? Keep a simple list of failures. Most will trace back to one of two causes: a policy that was unclear, or a missing knowledge base article. Fix the source, then run the same tickets again.

Test the edge cases on purpose. Write a few tricky tickets yourself:

• A refund request one day after the window closes.
• A cancellation on an order that already shipped.
• A customer who asks for a discount without a reason.
• A message with a wrong order number.
• A request that is not about your store at all.

The agent should handle these calmly and stay inside your rules. Any answer that looks like a guess tells you a policy needs more detail.

Step 5: Go live gradually

Going live does not have to mean everything at once.

• Start with the simplest ticket types and keep sensitive ones with your team.
• Read the agent's replies daily for the first week.
• Use your stats: resolution rate and time saved show where the agent works and where it needs better rules.
• Add one new permission at a time, and review after each.

Relay's ambition is to automate the large majority of support over time, but your own results depend on your tickets and your policies. Let your numbers guide how far you go.

The setup checklist

• Shopify store connected and order lookup tested
• Tickets arriving in the ticket center
• Knowledge base gathered or written
• Return, refund, cancellation, address and shipping policies written with exceptions
• Sensitive cases defined
• Tone notes written
• Permissions set, with refunds and codes enabled last
• 30 to 50 real past tickets reviewed
• Edge cases tested
• Gradual launch with daily review in week one

FAQ

How long does it take to set up an AI support agent for Shopify? Writing clear policies and testing take the most time, so plan for that. A trial period helps, since you can check the agent on your own tickets before committing.

Should I let the AI issue refunds from day one? No. Start with questions and order information, then enable actions like cancellations and address changes. Turn on refunds once tests show the agent applies your rules correctly.

What if the agent does not know the answer? Treat it as a gap in your setup. Add the missing answer to your knowledge base or tighten the policy, then run the same ticket again.

Do I need a knowledge base before I start? You need something. Even a short document with your most common questions, shipping times and product care instructions gives the agent solid ground. You can grow it as you review tickets.