Who should hear about an AI mistake and when?
Match the response to the consequence. A corrected internal draft and a wrong message that already reached a customer need different handling. Before deployment, work out who an error could affect and which cases go to a supervisor, a security specialist, legal review or straight back to the recipient as a correction.
Treat an internal draft you've fixed and a message already sent to a customer as the same AI error, and employees are left making big calls with no guidance. I'd sort the cases before anyone deploys the tool. Use your organization's existing requirements and skip inventing an informal standard of your own.
Give employees a short reporting route. The route should cover what happened and what's already paused, plus what may have been affected. Nobody should have to prove the cause before they raise a concern.
Keep the correction and the investigation apart. A customer may need accurate information before anyone knows whether a source document, the tool or a review failure caused the problem.
Afterward, look at why the process let the mistake travel, then update the relevant control and test it. Protect the people affected first, and then fix the failure so it doesn't happen again.