Skip to content
← Frameworks

AI Autonomy Levels Model

A four-level classification system that maps AI agent capabilities to governance requirements. Every AI deployment sits somewhere on this spectrum – and governance intensity must match.

L1

Assisted

Human decides, AI informs

AI provides recommendations, analysis, or data summaries. All decisions remain with a human operator. The system has no authority to act.

Governance

Standard review. Periodic audit.

Decision Rights

Human retains full authority.

Example

CV screening suggestions, risk score dashboards.

L2

Supervised

AI acts, human approves

AI generates outputs or takes preliminary actions, but requires human approval before execution. Human acts as gatekeeper with override authority.

Governance

Approval workflows required. Audit trail mandatory.

Decision Rights

Shared – AI proposes, human disposes.

Example

Auto-drafted contracts pending review, AI-generated shortlists.

L3

Delegated

AI acts within defined boundaries

AI operates autonomously within predefined guardrails, policies, and thresholds. Exceptions trigger human escalation. Monitoring is continuous.

Governance

Real-time monitoring. Escalation protocols. Regular governance review.

Decision Rights

AI holds operational authority. Human retains strategic override.

Example

Automated loan pre-approvals within risk parameters.

L4

Autonomous

AI acts independently, reports after

AI makes and executes decisions without prior human approval. Humans receive post-hoc reports. Governance is structural, not operational.

Governance

Board-level oversight. Full observability stack. Regulatory reporting.

Decision Rights

AI holds full authority. Human role is governance, not operation.

Example

Autonomous trading agents, self-scaling infrastructure, agentic workflows.

Governance Principle

As autonomy increases, governance must shift from operational approval to structural accountability. The question moves from "Did a human approve this?" to "Is the system designed so that accountability is never ambiguous?"

Sources: EU AI Act risk classification (2024), NIST AI RMF 1.0, ISO/IEC 42001:2023