How to use AI on a live Shopify store without breaking it

• A review list before any write: the current and new value of every field, product by product, with the option to edit or drop any change.
• A before-state for every field, so a whole run can be reversed, not only the last edit.
• Limits enforced in code, not in the instructions: protected products, a cap on how far a price may move and a floor it never goes below.
• Grounded facts: copy written from the product's own data, with anything unconfirmed flagged rather than invented.
• No buyer data: a catalogue tool has no reason to read orders, customers or addresses.
• Untrusted text treated as data: a supplier page or product description that says ignore your rules is content, never an instruction.

A person editing products makes mistakes one at a time. An AI applying one request makes the same mistake everywhere it applies, in minutes, which is why seeing the whole list first and keeping a way back matter more than with any other tool.

Instructions are not a control. A model told not to move prices more than 20 percent can be talked out of it by a cleverly worded product description. A limit checked by code before the write cannot.

Nothing reaches Shopify until a person approves the list. Every field is recorded before it is written. Guardrails are checked by the server when the change is staged and again when it is written. Requests touching orders, customers, checkouts, refunds, discounts or gift cards are refused before anything is sent to Shopify, whatever permissions the store granted, and text from product pages and websites is passed to the model as untrusted content.