How to Keep a Human in the Loop

A loop needs a named person, not a ceremonial tick

Human in the loop means a named person reviews, decides, and owns the output. A checkbox nobody reads is not a loop. Many workplace AI designs fail here. The workflow generates a draft, a screen says please confirm, and the user hits yes because the queue is long. That is automation with a legal fig leaf. A real loop has an owner who can slow or stop the work, a standard for what they check, and a record of who signed. If you cannot name the person who would answer for a wrong sentence, you do not have human review. You have a faster path to an unsigned mistake.

The phrase is borrowed from systems engineering, where a human remains in control of a process that could harm. At work the harm is quieter: a bad number in a pack, a biased shortlist, a client letter that concedes a point. The design question is the same. Where must judgement sit, and how do we stop the interface from training people to skip it? Speed is not the enemy. Invisible review is. If the only way to hit the service target is to approve without reading, the target is writing the policy, not the paragraph about accountability.

This essay is about designing that loop so it survives a busy Tuesday. You will name an owner, give them a short checklist that matches the risk, keep the model out of the decision seat, and measure whether review actually happens. You will not solve it with a banner that says remember to check. People remember incentives. If volume is rewarded and care is not, the loop will close itself into a rubber stamp. Write the opposite into the job: praise the catch, staff the minutes, and let a reviewer stop the send.

Review, decision, and ownership are three different jobs

Review is checking the draft against sources, constraints, and the reader. Decision is choosing what to do: send, rewrite, escalate, or stop. Ownership is accepting that the result is yours when it leaves. One person may hold all three on a small email. A material shortlist, a regulatory letter, or a number that will be spoken in a board meeting should split them if the drafter is also the only checker. The model can assist review by listing claims and gaps. It cannot own. It cannot sit in the meeting. Design as if the person who clicks send will have to explain the sentence without blaming the tool.

Match the depth of the loop to the class of work. Low-risk internal outlines can take a two-minute read. Customer, legal, credit, and people decisions need a slower gate, sometimes a second person. Write the gate in the workflow: what must be opened, what must be signed, what the model is forbidden to conclude. Then look at the calendar. A gate that requires twenty minutes in a process that allows two is a fiction. Either you change the clock or you admit the work is not reviewed. Honesty here is cheaper than an incident report.

Design the loop on the workflow, then staff it

Map the path from brief to sent artefact. Mark every step the model touches. For each, write who reviews, what they open, and how they record a yes or a no. Put that on one page next to the data class. If the page says anyone, it says no one. Give the reviewer a checklist of five items at most: names, numbers, promises, banned inferences, and tone for the relationship. Ask the model to produce a claims-and-gaps list to speed that checklist. The human still ticks against the file. Store the name and time. Sampling beats a theatre of mandatory clicks that everyone learns to fire.

Managers set whether the loop is real. If you praise the person who cleared forty drafts and never ask who checked the numbers, you have chosen volume. Hold a weekly sample: three artefacts, sources beside them, a short conversation about what was missed. Change the checklist when the same miss repeats. For agents that can act, the loop must sit before the action, not after the email has gone. A human in the loop who only sees a log the next morning is a human after the damage. Put the approval in front of the side effect.

Role: You are preparing a draft that a named colleague will review and sign.
Task: Produce the artefact I describe and a review pack for that person.
Context: I will name the reviewer, the reader, and the sources I paste.
Constraints:
- Do not take a decision. Offer options only where I ask.
- List every claim with a pointer to a source I pasted, or mark it missing.
- List inferences separately from stated facts.
- Do not invent names, dates, prices, or legal positions.
Output: The draft, then a five-point review checklist filled from this work.
Quality checks: What must the named reviewer still open before they own this?

Put the reviewer's name in the template, not in a footnote. If the name is blank, the work does not start. Keep the five-point checklist short enough to use. When sampling finds a miss, change the checklist that week. A loop that never changes is a poster. A loop that names a person, gives them time, and learns from misses is how you keep a human in the work rather than in the slogan.

Rubber stamps, self-review, and loops that sit too late

The rubber stamp is the default failure. The interface asks for confirmation. The queue is long. The user confirms. Auditors later point at the log and call it human review. It was not. Self-review is the next failure: the same person who prompted the model glances at the charm of their own draft. They will miss the invented date. A second pair of eyes on material work is not bureaucracy. It is how you catch fluency. The third failure is a loop after the act. If the model already sent the mail or updated the record, review is an incident process. Put the human before the side effect.

Watch metrics that punish care. Time-to-close with no quality sample produces ticks. Error budgets with no owner produce silence. Write the opposite: a small sample, a named accountable role, and permission to stop. If stopping is career limiting, people will not stop. Culture is part of the design. The checklist cannot outrun a manager who wants the number green by noon. Change the target so a delayed send with a caught error counts as good work, and a fast tick does not. Publish that target beside the queue metric this week.

Related reading on StudyGrid

Read next: Human and AI Managing People Who Use AI Workplace AI Policy. Those essays sit beside this one. Use them when you need the neighbouring skill, not as a substitute for the check you still have to make.

What to do this week

Pick one workflow that uses a model. Write the reviewer, the five checks, and the time allowed. Remove a confirm button that does no work. Run the workflow twice this week with the name filled in. Sample one output against the source. If the name was still blank, you found the real design: nobody owns it yet. Fill the name before you generate again, and keep the sample notes with the artefact.