The Four Properties of Human in Control
Human in control is an operating model for AI agents in which autonomous agents execute the work while a human retains final decision rights. But how do you know it is working? These four properties make the model observable, testable, and auditable.
Why properties matter
"We have a human in the loop" is not a testable claim. A person might be in the org chart without being in the decision path. The four properties turn a vague commitment into a checklist you can audit.
1. Visibility
You can see what every agent is doing, why it is doing it, and what it plans to do next.
Visibility means the overseer has a real-time view of agent actions, reasoning traces, and planned next steps. Without visibility, oversight is watching a black box blink. With it, a human can form a genuine judgment about whether an agent is operating within its mandate.
2. Decision Rights
A named human holds the formal authority to approve, change, or cancel any agent action.
Decision rights are the formal authority over what gets approved, changed, or cancelled. In a human-in-control model, decision rights stay with the person even when execution is delegated to agents. The agent proposes; the human decides. This is not a workflow checkbox - it is a governance guarantee.
3. Intervention
You can step in at any point - redirect an action, change the objective, or stop the agent entirely.
Intervention means the human can step in at any point: redirect an action, change the objective, or stop the agent entirely. The mechanism must work in real time, not just in theory. A kill switch that takes three business days to execute is not intervention - it is theater.
4. Attribution
Every consequential action is linked to a named person who approved it, with an automatic audit trail.
Attribution means every consequential action is linked to a named person who approved it. The audit trail is automatic: what the agent did, with what inputs, and under whose authority. When something goes wrong, you do not need a committee to find who was responsible - the chain of authority is explicit.
How they work together
The four properties are not independent. Visibility without decision rights is surveillance. Decision rights without intervention are a title on a business card. Intervention without attribution is chaos - anyone can stop anything, but no one owns the outcome.
Human in control only exists when all four properties are present and connected: the overseer can see (visibility), has authority (decision rights), can act (intervention), and is accountable (attribution). Remove one and the model collapses into its component failure modes.
Audit questions
- Can the overseer see what every agent is doing right now?
- Does a named person hold the authority to approve, redirect, or stop?
- Can that person intervene in under 60 seconds?
- Is every consequential action linked to a named approver in an audit trail?
- Has the kill switch been tested in the last quarter?
Read the full canonical definition:
What is Human in Control?