Human review can contribute judgment, an exception decision, or authorization that a workflow requires. Its value depends on the question being asked, the reviewer's authority, and the information available. A request to click “continue” does not supply missing technical evidence or grant the person decision rights they do not have.
The governance program must establish those conditions. AI agent governance defines the delegation and responsibilities around an operation; human approval can serve a particular role within that arrangement. It cannot establish the whole arrangement merely by appearing in the execution path.
Give the reviewer a question they can answer
Consider a hypothetical document-conversion workflow. The agent is preparing to use a service, but the workflow has not established the identity or configuration of the destination that will receive the documents. It asks the person who requested the conversion whether to continue.
In this example, the requester understands why the conversion is needed but lacks evidence about the receiving service. No documents have been sent, and no real integration, incident, or test is being described. The question is what a confirmation from this person could establish before the proposed transfer.
The requester could clarify the business need or identify which documents the task concerns. Those answers do not certify which service receives them. If the missing question concerns destination configuration, it needs evidence from the component or person able to establish that configuration. Routing it to the requester does not make it a business question.
Separate a request for information from a request to exercise authority. The same person may be able to provide both, but the workflow must establish that ability rather than infer it from the fact that they initiated the task. Someone who cannot answer the question should be able to leave it unresolved or route it to the appropriate owner without their uncertainty becoming an approval.
Distinguish an exception decision from a verified fact
An organization may define conditions under which an authorized owner can accept a particular risk. That is a different decision from verifying the fact that remains unknown. In the conversion example, an applicable exception might authorize proceeding under specified uncertainty if policy permits it; it would not establish the service's identity or prove its handling of the documents.
Keep that distinction in the decision and its record. State the known facts, remaining uncertainty, authorized scope, and conditions of the exception. Do not translate permission to proceed into a claim that the destination was verified. Conversely, identifying the service would not by itself establish that sending these documents to it is permitted.
NIST's AI RMF Playbook calls for differentiating responsibilities in human-AI configurations and oversight, including the distinction between people using systems and those overseeing them. That supports assigning a review question to a role with the relevant mandate and resources. It does not designate an approver for this hypothetical exception. NIST AI RMF Playbook, GOVERN 3.2.
Keep the governance program around the approval
The conversion workflow needs a way to identify its tools and destinations, a policy for which material may reach them, and an account of the identities and permissions involved. These functions establish the facts and authority the reviewer needs. A confirmation dialog cannot replace the work of making that information available.
Where policy already permits a bounded class of operations, implemented controls can apply those conditions without asking a person to recreate the policy for every request. Human review can focus on questions that call for judgment or an authorized exception. OWASP recommends approval for high-impact operations in LLM-based applications alongside limiting permissions and enforcing authorization at the destination. It presents approval as one control among several. OWASP Excessive Agency.
The workflow also needs observation, an effective connection between its decision and execution, and evidence for later review. In the example, the team must determine whether the transfer waits when an applicable approval is required and what happens if that approval cannot be obtained. These behaviors depend on the implementation; neither a human-facing prompt nor an approval event proves that the destination call was constrained.
Recurring requests to approve an unidentified service should lead to examination of why its configuration remains unknown. Another confirmation may add no new information. The program should address the missing ownership or evidence, rather than count repeated approvals as resolution of the underlying question.
Preserve the decision's meaning when work resumes
An approval concerns the proposal and conditions the reviewer considered. If the destination or documents change before the transfer, establish whether the decision still applies. The pre-execution governance guide develops that connection between a reviewable operation and the control able to withhold its effect.
Preserve what the human actually contributed. Confirmation that conversion is needed, verification of a service configuration, and acceptance of a defined exception are different records of reasoning or authority. Collapsing them into a single approved status makes later review less informative.
The AI agent audit trail guide explains how to connect decision evidence to the available attempt and result. An approval is not proof that a transfer occurred, that the destination accepted it, or that every stated condition was applied. Keep unknown outcomes separate from the permission granted.
Evaluate the decision the human contributed
Review whether the person had a meaningful question, the information needed to address it, and authority over the decision requested. Examine what happens when they cannot decide: who can resolve the uncertainty, whether an exception route exists, and how the operation remains within policy while the question is open.
Approval counts do not answer those questions. More confirmations do not, by themselves, establish more informed decisions or stronger controls. A useful review examines whether the required decision was made on an adequate basis and applied to the relevant operation, with evidence that preserves its limits.
Where human approval is required, an evaluation of Kona for Agents should make the reviewer role and the integration's handling of that decision concrete. Governance gives the person a decision they can legitimately make and connects it to the operation it is meant to govern.