Chapter 12 · companion worksheet

Human-in-the-loop gate mapping

Before any multi-agent or agentic system goes live, identify the actions that are irreversible or expensive to get wrong and place a human gate in front of each one — and only those. Use the table below to map your three most consequential AI actions.

The gate-design principle

A gate does not mean a person watches everything the AI does. It means you identify the specific actions where being wrong is permanent or expensive, and you require human approval there only. Everything else runs and gets logged — visible and correctable, but not blocked.

Tiered gate reference

• ☐ Log only — action runs; record is available for later review. Use below your low threshold.
• ☐ Flag for same-day review — action runs; a named reviewer checks it before end of day. Use between thresholds.
• ☐ Hard stop — action waits for a human signature before anything happens. Use above your high threshold or for any item on the list below.

Actions that always warrant a hard stop

• ☐ Money out the door above your threshold: $ ______
• ☐ Executing or agreeing to any legal document or contractual term
• ☐ Irreversible data changes (deletes, overwrites with no recovery path)
• ☐ Customer-facing commitments: pricing, dispute resolutions, contractual promises
• ☐ Clinical or safety decisions; any regulated outcome

Your three most irreversible AI actions

Threshold settings (set with your CFO or general counsel, not your vendor)

Gate-design checklist

• ☐ Every action in the table above has a named approver — not a role, a person.
• ☐ The gate fires before the action, not after.
• ☐ There is an error log — every gated decision is recorded with what the AI decided and why.
• ☐ We are not gating everything (which turns approvers into rubber stamps) and not gating nothing (which is the Tay mistake in a suit).
• ☐ Thresholds are set by the business owner, not the vendor.
• ☐ Someone reviews the gate log at least ______ (frequency) and can close the loop on near-misses.