Reviewing AI-generated workflow output before approval

Treat generated output as a draft until it passes a human review. Compare it with the source for facts, names, dates, numbers, tone, and any omitted caveat.

Check for invented details and private data that should not leave the workflow. Inspect unusual cases, not only a polished sample. Decide who can approve, revise, or stop distribution, and retain enough context to trace a correction later.

What claim would be most damaging if wrong?