From assistance to agency
Traditional software waits for a user. An agent can interpret an objective, choose steps, call tools and continue until a condition is met.
That shift introduces operational questions familiar from people management: authority, supervision, escalation and performance review.
Access defines risk
An agent with no access is only a conversational interface. An agent with access to email, files, systems and payment processes can create value—and consequences.
The control model therefore matters as much as model quality. Permissions should be narrow, observable and reversible.
Supervision must be designed
Human oversight is often invoked as a principle but left undefined. Effective supervision specifies which actions require approval, what evidence appears to the reviewer and how quickly intervention is possible.
Measure the full workflow
A fast agent can still create slower work if employees must repeatedly correct it or investigate unclear actions. Performance should include rework, exception rates, user confidence and downstream impact.
A colleague, not an employee
The metaphor of a digital colleague is useful because it encourages leaders to think about delegation. It becomes dangerous when it implies trust without accountability. Organisations remain responsible for what their systems do.