Human versus AI decision rights
‘A human makes the final decision’ sounds safe, but it is not a design. If nobody knows which human, which evidence, when approval occurs, or what AI can do before then, the control exists only on paper.
A usable decision model
Decision rights state who may recommend, act, approve, veto or stop, escalate an exception, and remain accountable for the outcome.
Decision Rights Matrix
Tune authority to decision risk, reversibility, and observability.
Low consequence · easier to reverse
High consequence · easier to reverse
Low consequence · harder to reverse
High consequence · hard to reverse
Tune authority through five dimensions
Replace a single human-control rule with boundaries that fit the characteristics of the decision.
Consequence
What is the impact of a wrong decision on people, customers, finances, or reputation?
Uncertainty
How incomplete, novel, or open to interpretation is the context?
Reversibility
Can the action be undone cheaply and in time?
Observability
Can failure be detected before it compounds?
Time pressure
Does human approval add safety, or delay the action until it loses value?
Decision-boundary case
Where may an expense agent act?
The agent sees a 450 TRY claim from a known vendor clearly inside policy, a 7,500 TRY in-policy claim from an unusual vendor, a 28,000 TRY claim near the policy limit, and three split transactions from one person.
What a decision contract needs
Recommendation
Which options and rationale may AI provide?
Action
Within which threshold may it execute directly?
Escalation
At which signal must it stop, and to whom does it hand over?
Override and accountability
Who may change the decision, and who owns outcome and policy quality?
Misconception
“A nominal human approver resolves accountability.”
A person approving hundreds of recommendations without context or time becomes a rubber stamp. Real control requires information, authority, time, and capacity to intervene.
Take it to work
For one AI decision, fill in Recommend / Act / Stop / Accountable owner.
Use at least one action threshold and one escalation signal, and name the human owner by role.
Private reflection · not saved
Keep this
Human-AI governance is not about whether a human exists. It is about whether the right authority operates at the right risk with the right evidence.