Create an automation

Use a test project with appropriate access and available credits. The sample store is a good place to start. You can also ask the agent in the chat to set up an automation for you; it proposes one for your approval.

Configure the task

Open AI → Automations → New Automation. Enter Catalog completeness review and choose visibility appropriate to the audience. Space visibility requires selecting a space you can access; personal and project visibility have different audiences.

Use these initial instructions:

Inspect the products and report records without a category.
Include each affected record's ID and SKU.
Do not change any records. If you cannot access a field, report that limitation
rather than treating the field as empty. State how many records you inspected.

Select a model. Keep Mutations → Require approval. The form shows Memories → Always on; previous context is separate from your instructions and attached skills.

Attach only what is needed

Attach the catalog review skill if you created it. A read-only catalog inspection does not need the formatting tool from the custom-tool tutorial. If you add instructions requiring that tool, attach it explicitly after discovery. A tool being visible in Dashboard’s catalog, or a plugin being connected, does not make it callable by this automation; add it under Tools.

Save and run manually

Save the automation and use its manual run action. New automations start Inactive; adding and enabling recurring triggers is a later step. Open the run and compare its summary with the actual tutorial data in Desk or Explorer.

A useful result identifies the records, distinguishes missing data from inaccessible data, and states its inspection scope. A green status without those results is not sufficient verification.

Test a proposed change

In disposable data, change the instructions to propose one narrowly identified update. Run with Require approval, inspect the affected record and before/after values, and approve or reject through Approvals. Read the record afterward to confirm the outcome.

Only then configure a schedule, entity event, or incoming webhook. Check costs, access, and loop protection before enabling unattended work.