AI Agents Can Act. Who Sets the Limits?

Imagine asking an AI agent to compare suppliers for a new project. It finds offers, tests them against your criteria and prepares a recommendation.

Can it then contact the suppliers? Share your internal budget? Accept a proposal? At each step, it moves from analysis to an action that could commit the organisation.

An agent can plan, use tools, check a result and keep going. That is what makes it useful. It is also why an instruction such as “find the best supplier” is not enough.

A task is not yet a mandate

Before assigning the work, define the intended result: by what criteria, using which data and tools? Which decisions may the agent make on its own?

The intention remains human. The agent receives an objective and a scope of authority. That scope should be chosen, rather than inherited from whatever technical access it happens to have.

If the steps are known in advance, a simple automation may be enough. Autonomy should serve a real need, because it can also introduce costs and errors.

Choose the stopping points

An agent can explore options and prepare a decision without asking for approval at every step. But some thresholds change the nature of the action: contacting a third party, sharing sensitive information, committing funds or changing an important system.

At those points, leadership must decide who authorises what. Friction is useful when it protects a judgment or a commitment; elsewhere, it may simply slow the work.

Check what autonomy achieves

A mandate also needs limits on time, actions and spending, along with a way to see what the agent did and stop it. Otherwise, a series of plausible tasks can continue long after it stops serving the original purpose.

Counting tasks completed or costs incurred does not tell you whether the agent created value. Look at the quality of the result, the risks avoided, and what people can now decide or accomplish better.

The leadership decision

Before deploying an agent in an important process, ask three questions:

• What outcome do we expect, and how far can the agent act alone?
• Which actions require human approval before they happen?
• How will we know when to stop it, correct it or expand its mandate?

An agent’s autonomy is not simply a technical decision. It is a delegation of authority that needs to be designed.

The intention belongs to the leader. The agent gets a mandate.