Human judgment needs room to act

A human approval is meaningful only if the person can understand the choice and change what happens. A button at the end of an automated process doesn’t establish that. For someone leading a team that uses AI agents, the responsibility starts earlier: choosing the purpose, deciding which constraints matter and giving people a workable way to intervene.

I read this principle as a commitment about authority. It leaves plenty of work for agents, including finding flaws in the objectives people have set.

The support schedule

In a hypothetical team, an agent proposes a support schedule. The stated goal is to shorten response times. The agent finds a schedule that covers every busy period, but repeatedly assigns inconvenient hours to the people whose availability records have the fewest restrictions.

That schedule may satisfy the target it was given. Whether it is acceptable is a different question. Perhaps some people haven’t entered their restrictions because they assume an informal arrangement still applies. Perhaps the schedule distributes an unwanted burden in a way nobody agreed to. Those possibilities require a conversation with the team members who would cover those hours.

A responsible lead sets the scheduling rules with their input: when people are available, how unpopular shifts are shared and which changes require agreement. The agent can then search for schedules within those rules, explain conflicts and compare alternatives. If coverage is impossible, it should show the conflict rather than quietly weaken a constraint.

The bad version presents one recommended schedule with a green score and an approval button. The lead sees that all shifts are covered and approves. The people affected have no opportunity to correct the assumptions before the schedule takes effect. A human has clicked, but the unresolved question of whose time matters has passed through the process unnoticed.

Authority has to be usable

The person reviewing a consequential recommendation needs access to the evidence and alternatives. They also need enough time to examine the parts that matter, and permission to reject the recommendation without being treated as an obstacle to an otherwise completed process.

That can change the design of the system. A scheduling tool might highlight departures from agreed availability, show who bears the difficult shifts and let a reviewer compare an alternative. Asking for approval after commitments have already been sent gives the reviewer much less room to act.

Human judgment can also be mistaken or unfair. Replacing an agent’s answer with a manager’s preference does not, by itself, improve the decision. The NIST AI Risk Management Framework’s discussion of human–AI interaction recognizes human bias, different needs for oversight and the importance of feedback from affected people. Each team still needs to test whether its own review process gives people a usable way to challenge an answer.

The relevant people include more than the person operating the tool. A manager’s authority over a schedule does not erase the interests of those expected to work it. A process that calls itself human-directed should make room for those interests to be heard and for material errors to be corrected.

What can be delegated

Taken literally as a ban on agents making decisions, the principle would be unworkable. We don’t need a person to approve each search query, sort order or routine assignment inside a schedule whose rules have already been agreed. Requiring that would consume attention better spent on disputed goals and consequential exceptions.

Delegation should specify the decisions an agent may make, the limits it must respect and the conditions that return a choice to a person. In the scheduling example, an agent could assign shifts within the agreed rules. It could not invent availability or waive an agreement to improve its score.

Before delegating, I would ask who can change the goal, who can stop or correct the action, and how a person affected by a mistake gets it addressed. If nobody can answer the last question, assign responsibility for handling those mistakes before treating the review process as complete.