Can people delegate without giving up control?

The human question.

Automation offers little freedom if someone must constantly watch it. Delegation should make room for other things while keeping authority understandable and limited.

What we expect—and need to test.

Our working hypothesis is that explicit limits and approvals can support useful delegation at an acceptable supervision burden.

How we would investigate.

First compare legitimate and adversarial requests under several spending-control designs using an offline fault corpus. A separate study with people would test comprehension, review effort, and the ability to change or refuse a request.

Compare Wallet Kernel with a simpler quota and approval system. Measure legitimate task completion alongside unauthorized actions, unnecessary denials, recovery, and operator effort.

What we would look for.

• Useful task completion
• Unauthorized or duplicate spending
• Understanding of delegated authority
• Approval effort and recovery time

What would change our thinking.

A simpler approach delivering comparable control with less effort would weaken the case for added complexity. A system that blocks every action would fail the usefulness test.

The study has not yet run.

This is a proposed study, with no participant results or measured effects to report. The sampling plan, final measures, and analysis need to be specified before the study begins.

Next step: Prepare a comparative study using the existing offline engineering evidence.

When AI saves time, who gets that time back?