Get in touch
← Hello / Perspectives
Governance & trust

Automation does not settle responsibility.

As agents gain the ability to act, leaders need a clearer account of authority, judgement and ownership.

An agent can prepare a recommendation, communicate a decision or change a record. To the recipient, those actions may all look like the business speaking. Internally, however, they involve different levels of authority and different consequences.

The introduction of agentic systems therefore creates a management question: whose responsibility is being exercised, and under what conditions? Leaving that implicit is a weakness even when the technology performs well.

Hello / Operating model

Permission follows the action.

Preparing information, recommending a decision and making a commitment require different authority.

Explore the modelPrepare information

Can be delegated within approved source and access boundaries.

Conceptual model · Select a component to explore its responsibility.

Separate capability from permission

A system's ability to use a tool does not establish its right to take every action that tool permits. Access should follow the responsibility assigned to the agent. Preparing a supplier comparison and committing expenditure are different mandates.

This distinction should be visible in the operating model. Define what the system may prepare, what it may execute within limits and what must reach a responsible person. Include what happens when the available information is incomplete or contradicts itself.

Make human review meaningful

An approval step can become ceremonial if the reviewer receives a persuasive conclusion without evidence, has insufficient time, or is expected to approve a large volume of routine requests. A human in the process is not automatically effective oversight.

Consider a service exception with financial implications. The reviewer needs to see the original commitment, relevant policy, proposed resolution and uncertainty. The system should make disagreement and correction easy. It should also record what was actually approved, rather than treating approval of one proposal as indefinite authority.

Keep accountability operational

Controls require owners who can maintain them. Someone must review failures, change permissions, investigate unusual behaviour and decide when a workflow should stop. These are ongoing responsibilities, not a document completed before launch.

The depth of control should be proportionate to the consequence. A low-risk draft and an irreversible external action do not need identical treatment. Excessive approval can destroy the usefulness of a system; insufficient control can expose the business to avoidable harm.

Trust grows when people can understand the system's remit and intervene effectively. The leadership task is to make responsibility clearer as action becomes more automated. That clarity allows the business to use agentic AI with ambition while retaining ownership of the decisions that matter.

Build the capabilityExplore Hello OS