How much should you authorise your AI agent to do?

Agentic AI signifies a shift where leaders must determine the distribution of knowledge, authority, and action between people and machines. This evolution moves beyond earlier generative AI assistance, which focused on tasks like drafting and summarization, to agentic systems capable of operating across workflows, coordinating tasks, making recommendations, triggering actions, and persisting over time. The core idea is that AI is transitioning from a tool people use to a collaborator they work alongside.
A teaching case involving Maya Mercer, head of Cyber Defence at Meridian Water & Power, illustrates this shift in a cybersecurity context. Ms. Mercer’s team moves from a fragmented security operations center to one supported by three custom multi-agent systems. The dilemma presented is whether to allow the incident-analysis agent to move from making recommendations to taking limited containment actions without immediate human approval, or to maintain human oversight while organizational foundations mature.
The case emphasizes that responsible leadership in this era involves defining boundaries for agents—what they can do and learn—establishing guardrails, measuring risk, addressing job anxiety, and determining the necessary scope of human judgment. The metaphor of agents as teammates forces leaders to govern hybrid systems of human and artificial agency rather than simply managing tool users.

* Agentic AI requires leaders to decide how knowledge, authority, and action are distributed between people and machines.
* The first wave of generative AI focused on assistance (drafting, summarizing, coding).
* Agentic AI involves acting across workflows, connecting to tools, coordinating tasks, making recommendations, triggering actions, and working with persistence.
* A case involves Maya Mercer, head of Cyber Defence at Meridian Water & Power.
* Ms. Mercer’s team transitioned from a fragmented security operations center to one supported by three custom multi-agent systems over two years.
* The agents include systems for incident analysis, detection engineering, and threat intelligence.
* A decision point is whether the incident-analysis agent should move from "recommend" to "act," taking limited containment actions without human approval.
* The framework suggests agents are treated as teammates with job descriptions, access controls, and review processes.
* The knowledge co-production occurs through human-agent systems where AI agents triage incidents, ingest intelligence, and generate hypotheses.

The shift described moves leadership from managing human expertise and resources to governing hybrid systems of human and artificial agency, challenging traditional assumptions about where judgment resides. The core tension lies in balancing the demonstrable operational gains offered by autonomous action (faster triage, reduced false positives) against the necessity of accountability and retained human oversight. The significance extends beyond technology to algorithmic discourse, suggesting that AI is reshaping how organizations define and authorize knowledge, moving it from being solely held by experts to being co-produced within learning loops.
The framing of agents as "teammates" serves a crucial organizational function by making abstract governance questions—what to trust, what to supervise, and where accountability lies—tangible leadership challenges rather than purely technical ones. The ultimate finding is that successful navigation requires conditional delegation: autonomy must be achieved through narrow, reversible steps based on verifiable confidence thresholds and explainability, ensuring human judgment remains the final site of oversight.
The underlying pattern suggests a systemic pressure to treat AI capabilities as compounding organizational capital, leading to a new leadership imperative: securing the knowledge generated by human-agent collaboration against leakage into external systems. The risk is not merely technological failure but the potential erosion of governed expertise if the learning loop and resulting value are not explicitly captured and controlled by the organization. What questions remain unanswered regarding the durability of this "teammate" model across long-term organizational structures?