Human-in-the-loop AI workflows: put approval where it matters.

Approval is not a vague safety promise

A message that says 'please review' at the end of an opaque process does not create meaningful control. The reviewer needs to know what changed, which sources were used, what remains uncertain and what will happen after approval.

The approval should be attributable, time-stamped and tied to the exact version of the proposed action.

Good places to stop

• Before sending an external message or publishing content.
• Before spending money, issuing an invoice or changing a financial state.
• Before granting access, connecting an account or exposing private data.
• Before changing rights, contractual terms or authoritative records.
• Whenever evidence conflicts or confidence falls below a defined threshold.

What the human should receive

• A plain-language summary of the proposed action.
• The source evidence and its freshness.
• Known gaps, conflicts and assumptions.
• The consequence of approving, rejecting or deferring.
• A reliable way to reverse or correct the action where possible.

Do not put a person in every step

If every routine action requires a click, the workflow has shifted labour rather than removed it. Deterministic, low-risk steps should proceed automatically once their conditions are satisfied.

Human attention should be reserved for exceptions, material decisions and actions whose consequences cannot be safely inferred or reversed.