“Keep a human in the loop.”

It has become one of the most common answers to concerns about artificial intelligence.

Let AI do the work—but keep a person involved.

Let the agent make recommendations—but require human approval.

Let the system operate automatically—but make sure someone is watching.

That sounds responsible.

Sometimes it is.

But there is a problem hidden inside the phrase.

A human can be in the loop without actually being in control.

A person can click Approve without understanding what is being approved.

A person can technically have authority to reject an AI recommendation while being pressured never to use that authority.

A person can receive thousands of automated decisions and have no realistic opportunity to examine them.

A person can supervise a system without possessing the knowledge, time, information, or access necessary to challenge it.

In those situations, the human may still exist somewhere in the workflow.

But meaningful human authority has disappeared.

That distinction matters enormously as AI moves from answering questions to taking actions.

The Rise of AI That Acts

Most people's first experience with generative AI involved conversation.

You asked a question. AI answered. You decided what to do with the answer.

That created a natural boundary. The AI produced information. The human acted.

Agentic AI changes that relationship.

An AI agent can increasingly be given a goal and then take steps toward accomplishing it.

  • search information,
  • communicate with other systems,
  • write code,
  • modify files,
  • send emails,
  • update records,
  • interact with business applications,
  • schedule meetings,
  • generate purchases,
  • change configurations,
  • or coordinate other agents.

The question is no longer simply: “Should I trust this answer?”

It increasingly becomes: “What should this system be allowed to do?”

That is a governance question. And it requires more than simply adding a human somewhere in the process.

Presence Is Not Authority

Imagine a company uses AI to review transactions.

The AI examines one thousand transactions each hour and recommends whether each should proceed. A person must approve every recommendation.

On paper, this sounds excellent. A human is in the loop.

But suppose the employee has only a few seconds to examine each recommendation. Suppose the AI's explanation is difficult to understand. Suppose rejecting the recommendation requires completing a lengthy form. Suppose supervisors measure the employee partly on how quickly approvals are processed. Suppose the AI is correct most of the time, so reviewing each decision gradually feels unnecessary.

Eventually the approval button becomes routine.

The human is still in the workflow. But is that person actually governing the system?

Not meaningfully.

The human has become part of the automation.

Approval Must Be a Real Decision

This is why I emphasize human approval gates in The Necessary Evil: AI Orchestration for Real Work.

An approval gate should not merely mean: A human clicked something.

It should mean: A human was given enough information and authority to make a genuine decision about whether the work should proceed.

For approval to be meaningful, several conditions need to exist.

The reviewer needs enough information. The reviewer needs enough understanding. The reviewer needs enough time. The reviewer needs permission to say no. The reviewer needs a practical way to stop or change the action. And the reviewer needs to know what he or she is responsible for approving.

Without those things, approval can become theater.

The Human Must Be Able to Reject the AI

One of the simplest tests of human authority is: Can the person actually say no?

Not theoretically. Actually.

Can the human reject the AI recommendation without extraordinary effort? Can the human request more information? Can the human alter the recommendation? Can the human suspend the workflow? Can the human escalate the issue? Can the human choose another path entirely?

If the answer is no, then the system may technically contain human review while functionally operating without meaningful human authority.

A person who is only permitted to confirm what the machine already decided is not really the final decision-maker.

The Human Must Be Able to Understand the Decision

Authority also requires sufficient understanding.

That does not mean every reviewer must understand every technical detail inside an AI model.

Most people cannot explain every internal operation of the software systems they use today.

But the reviewer does need enough context to understand what is being proposed and why it matters.

Consider an AI agent preparing a software deployment.

A meaningful approval might show:

  • what files are changing,
  • what systems are affected,
  • what tests passed,
  • what tests failed,
  • what risks remain,
  • what rollback method exists,
  • and what action will occur if approval is granted.

Compare that with a button that simply says: Approve deployment?

Those are both technically approval gates. They are not equally meaningful.

The quality of human oversight depends heavily on what the human is actually able to see.

Authority Requires the Ability to Stop the System

Human authority also needs teeth.

If an AI agent begins behaving incorrectly, someone needs the ability to stop it.

This becomes increasingly important as agents move faster than people can review individual actions.

An agent may perform hundreds or thousands of operations before a person realizes something is wrong.

That means human governance cannot rely entirely on manual review.

Real human authority may need to be expressed through technical controls.

  • permission boundaries,
  • transaction limits,
  • restricted credentials,
  • approval requirements,
  • rate limits,
  • rollback mechanisms,
  • audit logs,
  • exception rules,
  • circuit breakers,
  • and emergency shutdown procedures.

These controls are not alternatives to human authority. They are ways human authority is built into the system.

Authority Is Designed Before the AI Runs

Human governance does not begin when someone reviews the AI's final output. It begins much earlier.

It begins when humans decide:

What may this AI access?

What may it modify?

What may it spend?

Who may authorize it?

Which actions require approval?

What happens when confidence is low?

What happens when an action cannot be reversed?

What happens when the AI behaves unexpectedly?

Those are architectural decisions.

By the time an AI agent is operating, much of its practical authority has already been determined by the permissions and policies humans gave it.

That means system design is governance.

Not Every Action Needs Human Approval

This does not mean every AI action requires a human click.

That would defeat much of the purpose of automation.

If an AI agent has to ask permission before every trivial step, the system may become slower than simply doing the work manually.

The better question is: Where is human authority actually necessary?

The answer depends upon risk.

A low-risk agent summarizing public documents may need very little direct oversight. An AI generating a draft email may simply require the person to send it. An agent modifying production infrastructure deserves much stronger controls. An agent spending company money may require transaction limits and explicit approvals. A system affecting health, employment, security, legal rights, or public safety deserves even greater scrutiny.

The amount of human involvement should be proportional to consequence, reversibility, uncertainty, and risk.

Reversibility Matters

One useful governance question is: If the AI gets this wrong, how difficult is it to undo?

Some actions are easy to reverse.

Generate a draft. Reformat a document. Suggest three ideas. Create a test file.

Other actions are much harder.

Send an email to ten thousand customers. Delete production data. Transfer money. Change someone's access. Reject an applicant. Publish a statement. Shut down infrastructure.

The harder an action is to reverse, the stronger the human approval requirement should usually become.

That is not fear of automation. It is responsible system design.

Approval Fatigue Is a Real Governance Risk

There is another problem with human-in-the-loop systems.

Too many approvals can make approvals meaningless.

If a person must approve hundreds of routine AI actions every day, attention declines.

The reviewer begins trusting the system by default.

The action becomes: approve → approve → approve → approve

rather than: review → understand → decide

Eventually the human approval mechanism exists mostly to satisfy policy.

That is dangerous because everyone can point to the human checkpoint after something goes wrong.

“The human approved it.”

But if the process made meaningful review practically impossible, the existence of that approval does not necessarily mean the system was well governed.

Good governance therefore requires deciding which decisions genuinely deserve human attention.

Human Authority Includes the Right to Ask Why

Another test is whether the human can obtain enough evidence to evaluate what the AI did.

AI systems often produce conclusions much faster than people can reconstruct the reasoning behind them.

That creates a temptation to trust the result because examining it is inconvenient.

A strong workflow should instead preserve evidence.

What information did the AI use? What assumptions did it make? What tool calls occurred? What actions were taken? What checks passed? What uncertainty remains? What exceptions occurred?

For significant decisions, a human should not be asked to approve a mystery.

The evidence needed for approval should travel with the work.

Authority and Accountability Belong Together

One of the strangest possible AI arrangements would be this:

The AI has the practical authority to make the decision.

The human has responsibility for the result.

That creates accountability without control.

It is not sustainable.

If people remain accountable for AI-assisted outcomes, they need sufficient authority over the system to fulfill that responsibility.

And if an organization delegates authority to an AI system, the organization cannot pretend responsibility disappeared with the delegation.

AI cannot accept legal discipline. AI cannot answer to a family whose life was affected. AI cannot stand before a board. AI cannot accept moral responsibility.

Human beings and human institutions remain responsible for the systems they deploy.

Delegating execution does not automatically delegate accountability.

AI Should Have Bounded Authority

This is why I prefer to think of AI authority as bounded rather than unlimited.

An AI agent may have permission to act. But within clearly defined limits.

The agent may draft the email but may not send it.

The agent may restart this service but may not modify firewall rules.

The agent may spend up to $100 for approved categories but must request permission above that amount.

The agent may remediate known low-risk problems but must escalate anything outside the documented pattern.

This allows AI to be useful without pretending every decision deserves the same amount of autonomy.

The Human Must Remain the Governing Layer

A well-designed AI workflow may contain many layers.

Models. Agents. Tools. APIs. Databases. Automation. Verification. Monitoring. Policies.

But somewhere above those mechanisms needs to remain a human governing layer.

That does not mean a person manually performs every action.

It means humans determine:

  • purpose,
  • acceptable risk,
  • permissions,
  • boundaries,
  • escalation rules,
  • approval requirements,
  • success criteria,
  • and accountability.

The system operates inside those boundaries. The boundaries do not originate from the system itself.

A Six-Question Human Authority Test

1. Can the human understand what is being proposed?

If not, approval may be ceremonial.

2. Can the human reject it?

If no realistic path to rejection exists, authority is weak.

3. Can the human change the proposed action?

A meaningful reviewer should have more than a binary confirm button when circumstances require modification.

4. Can the human stop the system?

There should be an effective intervention mechanism when something goes wrong.

5. Does the human have enough evidence?

Important approvals should be based on visible information rather than blind trust.

6. Does accountability match authority?

The person expected to answer for the outcome must have sufficient power to govern the process.

If those six conditions exist, the human is not merely “in the loop.” The human remains in authority.

Human Authority Does Not Mean Human Perfection

There is an obvious objection. Humans make mistakes too.

Absolutely.

Human judgment is imperfect. Human reviewers can be biased. They can misunderstand evidence. They can overlook problems. They can make terrible decisions.

But AI governance is not based on the claim that humans are infallible.

It is based on responsibility.

Someone must remain answerable for what occurs. Someone must decide what risks are acceptable. Someone must determine when exceptions matter. Someone must be able to say: Stop.

AI may help that person make a better decision. It may provide information. It may detect problems. It may recommend an action. It may execute approved tasks.

But the presence of an intelligent tool does not eliminate the need for accountable human governance.

The Goal Is Governed Autonomy

The choice is not between two extremes:

AI does nothing without constant human permission

or

AI does everything independently.

There is a much more useful middle ground: governed autonomy.

Give AI enough freedom to perform useful work efficiently.

But define what it can access. Define what it can change. Define where approval is mandatory. Define when it must escalate. Define how it can be stopped. Define who owns the outcome.

That is the model I advocate in The Necessary Evil: AI Orchestration for Real Work.

The goal is not to keep humans clicking buttons forever.

The goal is to keep human judgment and human authority governing the system even as AI executes more of the work.