Skip to content
Request a consultation

Studio note

Human in the loop is a design choice, not a disclaimer

Decide where an assistant may act, where it should only draft, and where a person must approve before anything is sent, paid, or changed.

“Human in the loop” is often a sentence at the bottom of a proposal. It should be a box in the workflow. Someone needs to know which actions are automatic, which are drafts, and which are blocked until a person says yes. If the sentence is not attached to a step, it does not protect the customer or the business. It only protects the slide.

A person in the loop is not the same as a person who watches a dashboard. Watching is optional. A loop means the work cannot continue until that person acts, or the work is allowed to continue only inside a rule they already approved. Those are different designs. The first is a gate. The second is a boundary. Both are legitimate. Neither is a slogan.

Three kinds of steps

  • Automatic: safe, reversible, and already covered by a rule. Logging a note. Tagging a known question. Drafting a reminder that stays internal.
  • Draft: useful, but a person publishes or sends it. Marketing copy. A customer reply on a sensitive thread. A discount.
  • Blocked: the system may prepare the file, but it may not act. Refunds, access changes, legal wording, anything involving a promise you cannot undo.

If every step is automatic, you do not have a teammate. You have an unattended script. If every step waits for approval, you have added a tool without removing the work. The design is the line between those two. Draw it on one queue before you draw it on the whole company. A shop can auto-tag “where is my order?” and still block every refund. A clinic can draft a reminder and still block any message that mentions a symptom.

Write the failure down

Ask what should happen when the document is missing, the CRM field is empty, or the customer is angry. A serious system has an else. Usually the else is a person, with the context already attached so they do not start from zero. The handoff should include the question, the lookup that was attempted, and the rule that stopped the system. A handoff that says only “needs human” wastes the person you were trying to protect.

An illustration, not a result we measured: an agency inbox receives briefs, invoices questions, and the occasional angry client. The brief can be filed and summarized for the producer. The invoice question can be drafted from the last statement if the amount matches. The angry thread is blocked. Nobody needs a percentage to see that the third case is the one that can damage the relationship. The loop is the decision to stop.

Who is allowed to approve

Name the role, not “the team.” A refund over a threshold might belong to the owner. A tone edit on a help reply might belong to whoever is on the support shift. A change to the knowledge article belongs to the person who owns the policy, because the next hundred replies will copy it. If the approver is “whoever sees it,” the loop will be skipped on a busy afternoon. Put the approver’s name or role on the step.

Also name what the approver is looking at. A yes-button with no source is theater. Show the customer message, the record that was read, the draft, and the rule. If the draft cites a help article, link the article. If the order status came from the shop admin, show the status and the time it was read. People approve faster when the evidence is in the same screen, and they catch errors the model smoothed over.

When the loop should get tighter

Tighten the loop when the cost of a mistake is higher than the cost of the delay. Money, medical information, legal commitments, and public claims sit in that group. Loosen it only after you have seen the same step succeed on real items, not on a demo. A week of correct order-status drafts is evidence about that step. It is not evidence that the same system should issue refunds. Keep the evidence attached to the step.

Review the stops on a schedule you can actually keep. Once a month is enough for a single queue. Read the items that were blocked and the items a person rewrote. The rewrites are the real policy. If people keep deleting the same sentence, change the rule or change the source. If people keep approving without reading, the loop is fake and you should either remove the step or make the evidence harder to ignore. Do not add a second model to supervise the first one until a person has said what “wrong” means.

What to ask a vendor

Ask them to point at one screen and say which button a person must press. Ask what happens if that person is offline. Ask whether the system sends, posts, or pays while it waits. Ask for the log of the lookup, not a paragraph about safety. If the answer is a policy PDF and not a step in the workflow, you are still at the disclaimer. Nexen treats the loop as part of the build: the agent can read, decide within a rule, use the tools you name, and stop. Autonomy is not the goal. A shorter path for ordinary work is the goal.

The sample plan on this site includes approval points for that reason. It is not a quote and it does not book a call. Use it to see whether your queue has a real stop, or only a sentence you hope someone will remember.

Written by Nexen AI Solution. Published 14 October 2026.