Human Approval Gate for Consequential AI Decisions
Require a qualified human to review and approve AI-generated recommendations before they produce irreversible or high-stakes outcomes.
Objective
Preserve meaningful human control over AI outputs in contexts where errors cause significant harm to individuals or the organization.
Maturity Levels
Initial
Approval requirements are undefined; humans review AI outputs inconsistently with no documented criteria.
Developing
High-risk use cases are identified informally; approval steps exist in some workflows but are not standardized.
Defined
Approval gates are documented in process workflows with clear criteria, mandatory triggers, and defined review SLAs.
Managed
Approval rates, override frequency, and turnaround times are tracked and reported to governance bodies.
Optimizing
Override decisions feed back into model improvement; approval thresholds are tuned based on observed outcomes.
Evidence Requirements
What an auditor or assessor would expect to see for this control.
- —Approval workflow configuration showing trigger conditions, gated decision types, and review SLA
- —Audit logs for a sample period showing reviewer ID, timestamp, decision (approve/modify/reject), and rationale code for each gated decision
- —SLA compliance report showing actual review turnaround times vs. defined SLA
- —Override rate report disaggregated by reviewer, team, and use case over a defined period
- —Deployment configuration or integration test results confirming the gate applies to all systems meeting trigger criteria
Implementation Notes
Key steps
- Define which risk tiers (per HOC-001) require approval gates — without tiers, everything gets escalated or nothing does.
- Capture reviewer identity, timestamp, any override rationale, and the final decision alongside the AI output for audit purposes.
- For agents: approval gates must be synchronous — async approval creates race conditions in multi-step workflows where the agent can act before approval is received.
- Track automation bias: a near-zero override rate is a red flag that reviewers are rubber-stamping, not reviewing.
Example Implementation
Health insurer routing AI-assisted claims triage for adjuster review
Approval Gate Configuration — Claims Triage System
Trigger conditions (any of the following requires human approval before action):
- AI confidence score < 0.85
- Claim value > $50,000
- Claimant has open dispute on record
- AI recommendation = Deny
Review SLA: 4 business hours
Reviewer role: Licensed claims adjuster, Level II or above
Required log fields: reviewer_id · review_timestamp · decision (approve / modify / reject) · rationale_code · modified_output (if applicable)
Rationale codes: CONF_LOW | MISSING_DOC | POLICY_CONFLICT | FACTUAL_ERROR | OTHER
Control Details
- Control ID
- HOC-002
- Domain
- Human Oversight
- Typical owner
- AI Governance Team / Business Line Risk Owner
- Implementation effort
- Medium effort
- Agent-relevant
- Yes