Autonomy in AI refers to the degree to which a system can make decisions and take actions without direct human intervention. In assurance contexts, autonomy is not merely a feature — it represents a shift in responsibility, risk, and oversight that must be carefully assessed, particularly when AI systems are used in safety-critical domains such as defence, disaster response, or infrastructure protection.

An autonomous system typically includes components for perception, decision-making, and action, all of which operate based on learned models or programmed logic. Examples include unmanned aerial vehicles, autonomous ground robots, and decision-support tools that trigger actions in response to sensor data or contextual changes.

AI assurance of autonomous systems involves evaluating whether the system behaves predictably under various conditions, whether it respects operational boundaries, and whether it has fail-safes or override mechanisms to ensure human control when necessary. Autonomy increases the complexity of assurance because it introduces uncertainty in behaviour that may not have been explicitly programmed or foreseen.

Key assurance activities include:

• Scenario-based testing to simulate edge cases and operational stress.

• Evaluation of decision logic for consistency and safety.

• Verification of escalation protocols when the system is uncertain or encounters novel inputs.

• Human–machine teaming assessments to ensure meaningful human control or supervision.

Importantly, autonomy must be assessed not only technically but also from a governance and ethical perspective. Systems that act independently must be subject to policy oversight, legal compliance, and societal expectations about accountability and control. This includes ensuring transparency of autonomous decisions and the ability to explain or justify those decisions to human stakeholders.

Autonomy is often categorised in levels (e.g., in autonomous vehicles), with each level requiring different degrees of assurance and oversight. As systems become more autonomous, the burden of assurance increases, requiring more robust testing, traceability, and documentation.