VERITRAILBook Governance Baseline

IS — Integrity, Safety & Security

IS-13: Human approval for high-impact actions

Require designated human approval before AI agents execute high-impact, irreversible or external actions.

Priority critical · Owner: Business Owner

Where IS-13 sits among the 64 controls of the library.AO12AO-01 AI governance policy (critical)AO-02 Executive accountability and AI governance forum (critical)AO-03 AI roles, responsibilities and RACI (high)AO-04 AI risk appetite and tolerance (high)AO-05 AI literacy and role-based training (high)AO-06 Escalation paths and decision rights (high)AO-07 Regulatory and contractual obligations register (high)AO-08 Responsible AI principles and objectives (medium)AO-09 Third-party AI accountability (medium)AO-10 Prohibited and restricted AI use policy (critical)AO-11 Stakeholder engagement and feedback (medium)AO-12 Governance resourcing (medium)EI12EI-01 Enterprise AI inventory (system of record) (critical)EI-02 AI risk classification and tiering (critical)EI-03 Accountable owner per AI system (critical)EI-04 Intended purpose and context of use (high)EI-05 Model, prompt and component registry (high)EI-06 Data source and lineage registry (high)EI-07 Agent and tool registry (critical)EI-08 Third-party and vendor AI inventory (high)EI-09 Shadow AI discovery (medium)EI-10 Dependency and integration mapping (medium)EI-11 Inventory attestation and recertification (medium)EI-12 Use-case intake register (high)GL12GL-01 Intake triage workflow (high)GL-02 Pre-deployment risk and impact assessment (critical)GL-03 Risk-tiered approval gates (critical)GL-04 Change management and reassessment triggers (high)GL-05 Pre-release validation and testing (high)GL-06 Technical documentation (high)GL-07 Deployment conditions and go-live readiness (high)GL-08 Post-deployment monitoring (high)GL-09 Retirement and decommissioning (medium)GL-10 Exceptions and risk acceptance (high)GL-11 Vendor AI due diligence and contract controls (high)GL-12 Periodic review and recertification (medium)IS16IS-01 AI threat modelling (high)IS-02 Data governance and quality (high)IS-03 Privacy assessment and data minimisation (critical)IS-04 Sensitive data redaction and masking (critical)IS-05 Bias and fairness testing (high)IS-06 Safety and harm evaluation (high)IS-07 Accuracy and robustness testing (high)IS-08 Transparency and AI disclosures (high)IS-09 Human oversight design (critical)IS-10 Secure model and prompt supply chain (high)IS-11 Least-privilege tool access (critical)IS-12 Runtime guardrails and policy enforcement (high)IS-13 Human approval for high-impact actions (critical)IS-14 Prompt-injection and data-leakage testing (critical)IS-15 Kill switch and rollback (critical)IS-16 Rate, budget and resource limits (medium)SE12SE-01 Control testing programme (high)SE-02 Evidence repository and retention (high)SE-03 Governance metrics and leadership reporting (medium)SE-04 AI incident management and reporting (high)SE-05 Audit-ready evidence packs (high)SE-06 Corrective action and continuous improvement (medium)SE-07 Management review (medium)SE-08 Immutable action logging (critical)SE-09 Post-market monitoring plan (high)SE-10 Independent review and internal audit (medium)SE-11 Stakeholder transparency reporting (medium)SE-12 Model performance and drift monitoring (high)

At a glance

Pillar
IS — Integrity, Safety & Security
Priority
critical
Applies to
Agents and agentic workflows
Typical owner
Business Owner
VERITRAIL modules
Shield, Control
Control library
v1.0.0

Evidence that demonstrates this control

Examples of records an assessor would expect to see. Until a named reviewer has examined them, evidence is labelled User-asserted or Missing evidence in VERITRAIL outputs, never Verified.

  • Approval rules in runtime policy
  • Approval records with approver identity

Indicative framework mappings

FrameworkReferences
NIST AI RMFGOVERN 3.2, MANAGE 2.2
ISO/IEC 42001A.9.2, A.6.2.6
EU AI ActArt. 14

Framework mappings are indicative VERITRAIL interpretations to support governance readiness and evidence collection. They do not constitute legal advice, a certification, or a determination of conformity with ISO/IEC 42001 or compliance with the EU AI Act or any other law.

Where the self-assessment asks about it

  • HUM-Q1 — Are human-review, override, escalation, and appeal mechanisms defined for high-impact AI decisions?

Related critical risk flags

  • High-impact decisions lack defined human oversight — High-impact AI decisions without review, override or appeal raise fairness, safety and accountability exposure.

Last reviewed: