"Human in the loop" is one of those phrases that sounds responsible until you ask where the human sits, what they review, how much time they have, and what authority they hold.

Most teams cannot answer cleanly.

They add review because it feels safer. The AI drafts, a human checks. The AI recommends, a human approves. The AI flags risk, a human decides. Reasonable. But if review is not designed as real work, it becomes a bottleneck, a rubber stamp, or a liability shield nobody should trust.

Human review needs operating design.

Start with the trigger. What sends work to a human? Low confidence, high customer value, policy ambiguity, financial impact, legal risk, security concern, unusual tool action, sentiment, regulated data, model disagreement, eval failure, customer tier, or a novel edge case. If the trigger is "whenever the AI is unsure," define how the system knows that. If the trigger is "high risk," define high risk in workflow terms.

Then define the reviewer. The right human is not always the person closest to the tool. A support rep can review tone and policy fit. Legal reviews legal risk. Finance reviews credits. Engineering reviews production impact. A manager reviews exceptions that imply a decision, not every draft in the queue.

Review queues should respect capacity. If AI doubles the number of artifacts that need human checking, the review function becomes the constraint. That is not a safety success. It is a hidden operating cost.

A control plane should show review load: queue size, aging, reviewer utilization, approval rate, rejection reasons, edit distance, and downstream failures after approval. Without those numbers, teams will celebrate automation while humans quietly absorb the mess.

Escalation is different from review. Review checks quality before an action. Escalation moves a decision to someone with more context or authority. Mixing them creates confusion. A reviewer should not be asked to approve a decision they do not own. An escalation should not be buried inside a quality queue.

Good escalation rules say: when this condition appears, route to this owner, with this evidence, by this deadline, and pause or continue the workflow according to this rule.

For example, a customer support agent may draft replies automatically for routine cases. If the customer threatens churn, mentions legal action, reports a security issue, requests a large refund, or contradicts known account history, the workflow escalates with the ticket, customer context, draft, reason, and recommended next step. The reviewer should be the account owner, support lead, security team, or finance approver depending on the trigger.

Human review should also create learning. If reviewers repeatedly fix the same issue, that should update prompts, evals, routing, documentation, or tool boundaries. Otherwise review becomes permanent tax. The control plane should capture why humans changed outputs, because approval status alone is too thin.

There is a judgment trap here. Companies often keep humans in the loop because they do not trust the AI. Then they overload humans until the humans stop paying close attention. A rushed reviewer approving thirty AI actions an hour may be less safe than a narrower automated rule with good evals and sampling.

The goal is not maximum human involvement. The goal is the right human involvement.

Some actions should be fully automated. Some should be sampled. Some should be reviewed before execution. Some should escalate to accountable owners. Some should be blocked entirely. The control plane should let operators express those differences as runtime rules.

When human review is designed well, it protects quality without freezing the system. It also makes accountability clearer. The AI can propose, the workflow can route, the reviewer can approve, and the owner can be accountable for the decision rights they actually hold.

"Human in the loop" is a slogan. The real work is queue design, triggers, authority, evidence, capacity, and learning.

If those are missing, the loop is mostly theater.

This is part 9 of 10 in The AI Control Plane.