AWS
2.2.1 Risk management for autonomous agents Agentic AI introduces a risk category that traditional IT governance was not designed to address: autonomous systems executing multi-step actions with real-world consequences at machine speed. A single misconfigured agent can approve transactions, modify customer records, or trigger supply chain actions before any human reviews the decision. The impact extends across every system the agent has access to. Therefore organisations deploying agentic AI require three layers of control purpose-built for autonomous systems:
Preventive controls( before any action executes)
• Policy enforcement at the tool level: rules governing what agents can and cannot do
• Permission boundaries: leastprivilege access scoped to the specific task
• Guardrails: content filtering, PII redaction, input / output validation of actions
• Governance gates: human approval before agents execute high-impact / irreversible actions
Detective controls( during operation)
• Real-time observability: full tracing of every agent decision, tool call and data access
• Drift detection: monitoring whether agent behavior deviates from expected patterns
• Cost tracking: per-agent spend visibility to prevent autonomous cost spirals
• Anomaly detection: automated alerts when agents operate outside normal parameters
Responsive controls( when issues arise)
• Automatic rollback: reverting agent actions when thresholds are breached
• Human-in-the-loop escalation: routing complex or high-risk decisions to human reviewers
• Circuit breakers: halting agent execution entirely when systemic issues are detected
• Incident response: documented playbooks for agent failures, analogous to security incident response
Without these three layers, organisations face unbounded risk from autonomous systems operating at scale. With them, Boards can approve agentic AI deployments with confidence that the organisation maintains control over nondeterministic systems. The key principle shall be that the governance framework should not constrain innovation and instead it should enable it by giving teams the confidence to deploy autonomous agents.
aws. amazon. com 19