How to set up an AI agent that answers from your live inventory

What is an AI agent in this app?

It is a configured assistant with its own instructions, its own set of permitted Shopify data, and a trigger that decides when it replies.

You can run more than one. A support agent and a product-discovery agent can have different instructions and different data permissions, which is useful when you want tight control over what each one can see.

How do you write instructions that work?

Tell it what it is, what it must not do, and how to sound — then test it against real questions you have actually received.

The most valuable instruction is a prohibition. "Never state a stock level you have not looked up" prevents the failure merchants fear most.

Give it your policies: returns, delivery windows, and the phrasing you do not want used. The app's store profile has a field for phrases to avoid.

Then test with real messages. Take the last twenty questions your team answered by hand and paste them in.

Which data should the agent be allowed to read?

Start with products and inventory, add orders if you want order-status answers, and add customers only if you want it to recognise returning buyers.

Fewer permissions means fewer ways to get an answer wrong in public. Expand once you have watched it behave on the permissions you granted.

The available groups are products, orders, inventory and human-support escalation, with reasoning available for harder questions.

When does the agent reply?

Today, when a customer sends a message. That inbound trigger is the only one merchants can switch on; order, checkout, customer, product and fulfilment triggers are rolling out.

This matters for expectations. The agent answers questions; it does not chase customers. Anything that requires messaging someone who has not written to you first is not available yet.

How do you know it is working?

Read the run timelines. Each run shows which tools the agent called and what came back, so a wrong answer has a visible cause.

Where an answer is wrong because the agent lacked information, fix the product data. Where it is wrong because the agent overstepped, tighten the instructions or remove a permission.