Saying that people remain in charge of AI is a commitment. Making that commitment real requires arrangements that people can use, question, and test.
In his article Models as Insider Risks in the Super Intelligence Era, published on 10 October 2026, Satya Nadella argues for separating the provision of intelligence from authority over its use. He proposes controls outside the model, independent evidence and validation, and the ability for an authorized person to interrupt its work.
That is a useful starting point for governance. It connects a claim about human authority to the design of systems through which authority is exercised. The same questions matter for an agent handling routine organizational work. We do not need to settle what counts as superintelligence before asking who may act, within which limits, and who must answer when something goes wrong.
Delegation leaves obligations with people.
When an organization lets an AI system act, it makes a choice about how work will be done. That choice needs a purpose, a defined scope, and someone with authority to make it. The system’s capability does not establish whether a particular use is appropriate.
Nadella argues that provider assurances do not transfer responsibility away from those deploying a system. His proposed safeguards include diverse models, activity records, continuous testing, independent controls and audit, containment, and disclosure when incidents occur. Source
For those who govern, the task extends across the life of the arrangement. They need to assess the proposed use, establish its conditions, and revisit those conditions as circumstances change. Assigning daily work to an operator does not settle every obligation of the people who authorized it.
Responsibility includes the continuing duty to notice, judge, and act. Accountability requires people to answer for how authority was exercised and for its consequences. Both need clear ownership and practical support.
Authority needs practical means.
Consider an agent allowed to prepare and send messages to customers. A written instruction might tell it to seek approval before making a commitment. The governing question is whether the system actually prevents an unapproved commitment, including when the model misunderstands an instruction or encounters misleading content.
A workable arrangement would distinguish drafting from sending. It would limit access to the information required, define which actions need approval, and enforce those limits through controls the model cannot change. The approver would see the proposed message and enough context to judge it.
The ability to stop work also needs a precise meaning. Pausing further messages does not retract those already sent. A stop procedure must therefore connect to investigation, correction, and recovery. Someone needs authority to carry out each step, including deciding when work may resume.
These are design questions with governance consequences. They determine whether a person’s authority has an effect at the moment it matters.
Oversight needs evidence people can examine.
A reviewer needs to know what the system was asked to do, what information it used, what actions it attempted, and what actually happened. An explanation generated after the event may help someone investigate. It should not be the only record available.
The records used for oversight should be protected from alteration by the system whose conduct they describe. They should connect actions to permissions and approvals, while distinguishing an attempted action from a completed one. A confident statement that a message was sent is different from evidence supplied by the sending service.
More records do not automatically produce better oversight. People need relevant information in a usable form, with a clear route for responding. Access and retention also need limits: collecting sensitive material without a defined purpose can create another problem to govern.
Using a second model to review the first may help. It does not, by itself, establish independence. The review process needs checks that can identify failures shared by both models, and evidence beyond their agreement.
The people exercising control also need governance.
This is where the discussion must extend beyond containment. A system can remain within its assigned permissions and still serve a poor purpose. An operator can have effective technical control without having legitimate authority to make the decision at hand.
We therefore need to ask who grants authority, what bounds apply to it, and how its exercise can be challenged. The person who can stop an agent may not be the person entitled to approve a new use. A supplier’s control over a platform may differ from an organization’s authority over its work.
People affected by decisions need a way to raise concerns and obtain a considered response. That requires an identifiable person or body able to investigate and provide a remedy. A complaint channel has little value if everyone receiving the complaint lacks the authority to act.
Shared responsibility needs explicit relationships. It should make clear who prepares, who decides, who operates, and who reviews. It should also make clear where disagreements go and who remains answerable.
Test whether authority works.
A governance arrangement should be tested through situations in which it is needed. Can an agent act without a required approval? Can the designated person stop it while a task is running? Does the record show what completed before the stop? Can someone investigate without relying on the agent’s own account?
Tests should include unavailable reviewers, missing evidence, and failures in connected services. They should also test the human arrangements: whether people understand the information presented, have time to assess it, and know how to escalate an unresolved concern.
Review should continue when tools, permissions, models, or working practices change. A control demonstrated once may no longer fit the work it is meant to govern.
For AI Governance Center, the practical standard is this: human authority must be built into how AI operates, supported by evidence, and open to challenge. Good governance connects the purpose of the work to the people entitled to direct it, and keeps those relationships effective as the work unfolds.
Source and perspective
Satya Nadella, Models as Insider Risks in the Super Intelligence Era, 10 October 2026.
This article develops the author’s interpretation and governance recommendations. The customer-message example is illustrative. It does not describe a reported incident or claim that a particular product implements these controls.