What Decisions Should an AI Agent Never Make Alone?

An AI agent can sort requests, pull the right details, and prepare the next step. It should not make the final call when that call closes a door, creates a commitment, changes access, or cannot be easily undone.

The practical rule is simple: let the agent prepare and recommend. Put a person at the point where the decision takes effect. The NIST AI Risk Management Framework treats governance, defined roles, and accountability as core parts of managing AI risk.

What is the line between an action and a decision?

An action follows a known instruction. A decision applies judgment to a situation, especially when the facts are incomplete, the rule has exceptions, or the outcome changes something for someone.

An agent can collect the request, identify its type, check it against a written rule, and place a recommendation in front of the right person. A person should decide whether the rule applies and whether the outcome is acceptable.

That distinction keeps routine work moving without turning a queue into an unattended authority.

Which decisions should an AI agent never make alone?

An AI agent should not make a final decision alone when the outcome changes a person's rights, money, commitments, access, or the status of a request. NIST's Generative AI Profile identifies risks including confabulation, information integrity, and data privacy that organizations need to manage when using generative AI. (NIST Generative AI Profile)

Put these decisions behind human approval:

• Closing, canceling, or declining a request. The agent can summarize the request and explain the recommended outcome. A person makes the final call.
• Sending an external commitment. The agent can draft an email, proposal, approval, or answer. A person should approve any message that commits the business or changes a relationship.
• Changing access or permissions. The agent can identify the requested change and the relevant policy. A person confirms who gets access and why.
• Approving a payment, refund, discount, or purchase. The agent can assemble the facts. A person makes the financial decision.
• Making an exception to a rule. The agent can show the rule and the exception request. A person decides whether to depart from the rule.
• Resolving incomplete or conflicting information. The agent should flag uncertainty rather than choose the most convenient answer.

A request can be old and still need judgment. Queue age is a reason to look, not a reason to close.

What can an AI agent do without approval?

An AI agent can do the preparation work that makes a human decision faster, as long as the task has a clear scope and a safe way to correct it.

Useful no-approval tasks often include:

• classifying an incoming request
• extracting details from documents
• checking a request against a stated rule
• drafting a response or a decision summary
• routing work to the named reviewer
• flagging missing details or conflicting records
• making reversible updates that do not change a person's access, commitment, or standing

The boundary is not about making the agent wait for every click. The boundary is about preventing the agent from making the consequential call.

If an agent drafts sales replies, the same pattern applies: draft first, then let a person decide what goes out. See Should AI Send Sales Replies Without Approval?.

How do you set up an AI agent approval process?

A workable approval process names the decision, the reviewer, and the stop condition before the agent runs. The NIST AI Risk Management Framework recommends assigning roles and responsibilities for AI risk management rather than leaving accountability implied. (NIST AI RMF)

Use this setup:

1. Name the decision. Write the exact outcome the agent is not allowed to finalize, such as “decline this request” or “grant this permission.”
2. Name the reviewer. Assign a role that owns the decision. “Someone checks the queue” is not an owner.
3. Give the reviewer the evidence. Show the original request, the facts the agent used, the relevant rule, and the agent's recommendation.
4. Offer clear choices. The reviewer should be able to approve, reject, request more information, or route the item elsewhere.
5. Set a safe default. When the agent lacks a required fact, sees conflicting information, or reaches a deadline, it should hold the item for review rather than finalize it.
6. Keep a decision record. Record what was approved, who approved it, and what information supported the decision. A record makes a later correction possible.

What does an approval boundary look like in practice?

An approval boundary works best when the agent's role and the person's role are visibly different.

The table does not need to be complex. It needs to be specific enough that a reviewer can see where the agent stops.

Why does an approval process need an owner?

An approval queue without an owner is only a slower form of automatic decision-making. A request can sit until someone clears it quickly, or it can be closed because no one feels responsible for the outcome.

Give each approval type one accountable role. Give that role a written standard for approving, rejecting, and asking for more information. When the role is unavailable, route the item to a named backup or leave it pending.

An agent also needs a way to stop. For controls that limit what an agent can do when something goes wrong, see What Should Every AI Agent Have Before It Runs?.

What should you decide first?

Start with the one decision your agent currently makes that would be hardest to undo. Write down the effect of that decision, the person who owns it, the information they need, and the condition that makes the agent stop.

If nobody can name the owner of a decision, the agent should not make it. That is the foundation of an AI agent approval process: routine work moves forward, while consequential decisions remain with the person accountable for them.