Merchandising Agent security and governance

The Merchandising Agent operates inside the Coveo Merchandising Hub (CMH) rather than as a separate, general-purpose AI tool. Because it works within CMH, it stays inside the same boundaries and governance patterns as the rest of your merchandising workspace.

A few principles shape the experience:

Each of these principles is covered in more detail in the sections that follow.

The Merchandising Agent is designed to help you prepare and understand work, not to act on its own. When you describe a goal in natural language, the agent interprets your intent and, where applicable, prepares a proposed change for you to review.

Nothing that affects your storefront is applied until you review the proposal and confirm it. You remain the gatekeeper: if a proposed result doesn’t match your intent, you can refine it before anything changes, or decline it.

Not all requests carry the same level of risk, so the Merchandising Agent treats them differently. Actions fall into three categories.

This model lets the agent move faster on low-risk requests while keeping a clear, deliberate step between a proposed change and a confirmed one. When a request falls outside the supported action set, the agent explains that the action isn’t available, so you know to handle it manually in the CMH or the Coveo Administration Console.

Among the supported write actions, filter rules deserve the closest review. A filter rule that’s broader than you intended can leave a search or product listing page returning no products at all, so confirm the scope of a proposed filter rule before you apply it.

The agent acts only through the merchandising capabilities that the CMH itself exposes. It works with search, product listing pages, product recommendations, and facets for the properties and locales you manage.

It doesn’t reach into the Coveo Administration Console. Configuration such as query pipelines, sources, and fields stays outside the agent’s scope, and the agent declines any request to change it. This boundary matters in practice: during the early access program, some users asked the agent to adjust query pipelines, which it can’t do.

The agent is also bound by your own privileges. It can’t read or change anything in the CMH that you couldn’t read or change yourself.

The Merchandising Agent can create and edit supported configurations, but it has no delete access. There’s no request that causes the agent to remove a rule, a facet collection, or any other configuration.

Anything the agent creates is an ordinary CMH configuration. It appears in the manager that owns it, and you edit or delete it there exactly as you would a configuration you created yourself, through Search manager, Product listings manager, Recommendations manager, or Facet manager.

Each manager’s list includes a Last edited column. For a configuration the agent created or changed, that column shows Agent together with the name of the person who confirmed the change.

Every agent-assisted change therefore carries a named owner. You can see which configurations came through the agent, and which merchandiser approved each one, without leaving the manager you’re already working in. From there you manage them like any other configuration: open one to review it, edit it, or delete it.

The Merchandising Agent grounds its responses in relevant, trusted context from the CMH and supported Coveo systems, rather than treating its output as free-form guesses. When it surfaces an insight or prepares an action, it draws on the configurations, catalog information, and metrics available in your merchandising workspace.

You should still review what the agent returns. grounding improves the relevance and reliability of responses, but it doesn’t remove the need for you to validate important recommendations before you rely on them.

The Merchandising Agent can accelerate your workflows, but you stay in control of the results. To get the most value while keeping your storefront safe: