Guide · AI-assisted editorial
Before an AI workflow goes live: six checks that matter
A practical acceptance checklist for a workflow that reads text, proposes an answer or moves information.
Check the inputs and evidence
Use fictional or appropriately authorised test data. Include missing fields, contradictory information and irrelevant text. Ask for the source behind important claims. If the system cannot support an answer, design a clear route to say so rather than guessing.
Check permissions and approvals
Grant only the access the workflow needs. Separate drafting from sending, and preparing a change from applying it. Make a person approve consequential outputs during the trial. Record who can stop the automation and how to revoke its access.
Check failures and the bill
Disconnect a dependency and confirm the workflow reports failure. Retry the same input and check that it does not create duplicates. Put limits on usage and inspect costs with a realistic workload. Keep a short rollback procedure beside the workflow instructions.
Define success before measuring
Choose a small measurable target, such as reducing the time spent preparing a weekly report while preserving the required checks. Record correction time as well as generation time. This checklist is an editorial testing framework, not a certification or security audit.
Sources & reading
Features and terms can change. This is editorial guidance, not a claim of independent product testing. Editorial policy