CONCEPT / REVIEWED JULY 22, 2026
Human checkpoint
Where should a human checkpoint sit in an AI workflow?
Place an active human approval immediately before the workflow sends, publishes, pays, changes access, records a decision, makes a promise, or affects a person.
What the operating concept requires
- The reviewer is named by role and has authority to reject or rewrite the output.
- The original source, draft, checklist, and known uncertainty are visible at review time.
- Approval is an explicit action, not the absence of an objection.
- The automation cannot bypass the hold when a check fails or a case is unusual.
What this concept does not prove
- A human click is not meaningful oversight if the reviewer lacks time, evidence, authority, training, or a usable rejection path.
- Human review does not cure a prohibited input, invalid purpose, or unqualified high-impact use.