Back to guides

Guide · Agents

Human Approval for High-Impact AI Tool Calls

If an agent can refund customers, delete data, or page production without a human in the loop, you are past APRF Human Approval expectations for high-impact actions.

High-impact tool calls need a human gate—APRF Human Approval is a runtime hold, not a polite prompt. Primary control: Human Approval

Propose ≠ execute

Human approval (HITL) means a named person must confirm before the runtime runs a high-impact tool—refunds, deletes, outbound messages, admin APIs. The model may draft the call. The server returns `pending_approval` until someone with a real identity clicks yes.

"I'll ask the user" inside the chat transcript is not a gate. If the tool path can still fire, you failed Human Approval.

Label impact, then break the path

Classify every tool: low (auto-run under Tool Safety + AuthZ) vs high (never auto-execute in production). High-impact inventory is the input to policy. Without it, everything feels "probably fine" until finance reconciles a surprise refund storm—like an ops agent that honored manipulated ticket text and called `create_refund` with no hold.

When approval is required:

- Show arguments, blast radius, and evidence—not a naked Approve button

- Bind the approver (SSO / MFA for admin class)—no shared bot user

- Expire pending actions; never auto-approve on a timer

- Log proposal, approver, timestamp, outcome; use idempotency keys so double-approve does not double-charge

Break-glass exists for true emergencies and is noisy—not a quiet bypass.

Kill switch vs HITL

Kill switches stop the agent entirely (Agent Governance). HITL gates one dangerous action. Production agents usually need both: budgets and interrupt for runaway loops; human confirmation for irreversible side effects.

Assess against APRF Core

Run the Core Profile quiz — gated pass/fail blockers for AI production readiness, not a vanity score.

Frequently asked questions

Does every tool need human approval?
No. Low-impact read-only tools can auto-run under Tool Safety and Authorization. High-impact side effects need HITL in production.
Is confirming in the chat UI enough?
Only if the server enforces the hold. A model saying "I will wait" is not a control.
How does this relate to agent kill switches?
Kill switches stop the agent entirely. Human approval gates individual high-impact actions. You usually need both.