All insights

Responsible AI

Where AI agents still need people

Human review is not a failure of automation. It is part of a well-designed operational system.

Human review is a design decision

The goal of an agent is not maximum autonomy. It is useful work performed within a level of authority the organisation can justify.

Let agents handle bounded, reversible work

Agents are strongest when actions are easy to verify, consequences are limited and the source information is clear. Preparing a draft, retrieving a record or classifying a request can often be performed safely with monitoring.

Require approval for consequence

Financial commitments, contractual changes, sensitive advice, exceptions to policy and irreversible actions normally warrant a person. The agent should present the context, source and proposed action so approval is efficient rather than ceremonial.

Escalate with context

A human handoff should include the original request, relevant records, actions already attempted, uncertainty and the specific decision required. Making a person reconstruct the case defeats much of the value.