Let's say the task is automatically aggregating customer support info.

First step is collecting incoming emails

Second step is summarizing each one

Third step is batching by issue/severity

Notice how there is tolerance for deviance/error. "An error" looks like coding a ticket red instead of yellow, or slightly misrepresenting what a client said. The overall workflow can still be net positive.