3. How Can CIOs Set Boundaries for AI Agent Autonomy?
The decision: Assign autonomy according to business risk, not according to what the technology is capable of doing.
AI is probabilistic: it reasons, predicts, and recommends based on available information. Many enterprise processes are deliberately deterministic. They require consistent rules, traceable decisions, and auditable outcomes.
When payroll runs, the books close, or compliance is at stake, 'almost right' is not good enough.
CIOs need an operating model that combines the strengths of both systems. AI can support reasoning and dynamic orchestration, while deterministic systems enforce policy, permissions, and transaction integrity.
Moorthy put it simply in The Australian Financial Review: “When you have those deterministic rails, it actually makes the probabilistic side of the AI more powerful because it can be more trustworthy.”
Autonomy can be considered across three modes:
- Advise: The agent provides information or recommends an action, but a person makes the decision.
- Act with approval: The agent prepares or initiates the action, but a person authorises execution.
- Act within policy: The agent completes routine work autonomously within predetermined permissions, controls, and exception thresholds.
The appropriate mode should reflect the consequence of an error, whether the action is reversible, the sensitivity of the data, and the financial, regulatory, or workforce impact.
High-risk actions, such as payroll changes or material financial commitments, should make appropriate use of deterministic calculations where necessary, and remain tightly bound by approvals and audit controls.
Lower-risk, reversible tasks may justify greater autonomy after the agent’s performance has been demonstrated.
The principle is simple: autonomy should be earned, not assumed.
As one CIO of an Australian healthcare organisation observed, the opportunity is to rethink what no longer needs to be done manually and what AI can 'take off the load'.
The goal? To remove low-value administration and drive scaled insights, while preserving human judgement where consequences and accountability demand it.