Why an AI agent should never publish on your Shopify store

Author and reviewer are two jobs

Every organisation that publishes anything serious separates the person who writes from the person who signs off. Newspapers have editors, finance teams have a second signature on payments, developers have code review. Authors aren’t careless, but nobody reliably catches their own mistakes: they read what they meant to write.

An AI agent is an author with that weakness at its most extreme. It has no doubt about its own output and no way to know which of its sentences are guesses. When the same agent can also publish, you have merged the two jobs into one actor, and the check simply disappears.

Tools now make that merge easy. Shopify’s own connector for Claude, for example, can update a product’s title, description, status, images and variant prices from a chat, within the access and permissions you grant it. That is useful. It also means the question “should the agent publish?” is no longer theoretical: it is a setting you choose, often without noticing.

Errors at scale

A person who misreads a supplier sheet gets one product wrong. An agent that misreads the same pattern gets it wrong everywhere it applies the pattern. Take a hypothetical store selling kitchen knives: the agent decides that “dishwasher safe” is a standard selling point and adds it to forty product descriptions, including the carbon steel ones that rust in a dishwasher.

With direct publishing, you discover this through returns and reviews. With a review step, you discover it on the first proposal, send it back with a note, and the next batch is fixed. The cost of a review is a few minutes per version. The cost of cleanup is finding every affected page, reconstructing what was there before and answering the customers who already bought.

An agent doesn’t make more mistakes than a person. It makes the same mistake more times, faster.

Somebody has to be accountable

When a claim on your store turns out to be wrong, the questions are always the same: who wrote this, who approved it, and when. An agent can answer none of them in any useful sense. It has no memory of the conversation that produced a given paragraph and no responsibility for it.

Accountability needs a name on the approval. That person doesn’t have to write anything; they have to read the change against what is live and decide. If your process has no such person for AI-written changes, then in practice nobody is accountable for a growing share of your catalogue.

Claims that carry legal weight

Some product copy is marketing. Some is a claim a buyer relies on, and in many markets it is regulated: health benefits of a supplement, the safety of a toy for a given age, a waterproof rating, an organic or recycled label, a certification. Models produce this kind of language easily because it is common in their training data, whether or not it is true of your product.

A reviewer who knows the catalogue catches “clinically proven” on a product that was never tested, or “BPA-free” on a bottle nobody checked. Treat any change that adds or modifies such a claim as needing a second look from someone qualified to confirm it, not just a proofreader.

What mass low-quality changes do to SEO

Google is clear that it cares about the quality of content rather than whether AI helped produce it. It is equally clear about the other side. Its spam policies define scaled content abuse as generating many pages for the primary purpose of manipulating rankings and not helping users, and give as an example using generative AI tools to generate many pages without adding value for users. Its guidance on generative AI content also asks site owners to pay attention to the accuracy and quality of generated metadata such as titles and meta descriptions.

An agent with publish rights can rewrite hundreds of product pages and SEO fields in an afternoon, all in the same voice, all saying roughly what competitors say. Even when that stops well short of a spam action, you have replaced pages that had specific detail with pages that have none, across the whole store at once, with no easy way back.

What “propose, don’t publish” looks like in practice

The alternative is not to ban agents. It is to give them the same door as a junior copywriter:

• The agent reads the catalogue and drafts a new version of a product, collection, page or article.
• The draft is submitted as a proposal and does not touch the live store.
• A reviewer compares it field by field with what is live, then approves, requests changes with a note, or rejects it.
• Only an approved version reaches the storefront, and it keeps a record of who proposed and who approved it.

This is how Vetta’s MCP server is designed: Claude or ChatGPT can read content and propose versions, and there is no publish tool for them to call. Proposals go into the same approval workflow as work from your team or a CSV import, with a word-by-word diff of titles, descriptions, SEO fields, tags and metafields, including translations. Agents can only be authors; approving and publishing stay with people.

Rules for agent permissions

Whatever tools you use, these rules hold up:

1. Agents draft; people approve and publish
2. No agent tool can write directly to live content
3. Connect agents through accounts with minimal permissions
4. Keep agents away from orders, customers and discounts
5. Label every agent version as agent-written
6. Name the reviewer on every approved version
7. Send health, safety and certification claims to a qualified reviewer
8. Start with small batches and widen scope gradually
9. Keep every previous version so you can roll back
10. Review agent permissions whenever you add a tool

None of these rules slows the agent down. They only slow down the part that should be slow: the moment a change becomes something your customers read and your store stands behind.