Human-in-the-Loop AI Is Not an Approval Queue: The Four Human Moments

Six · 9 min read · October 8, 2026

Human-in-the-Loop AI Is Not an Approval Queue: The Four Human Moments

Say your agency sends each client a status update every Friday. You've started having AI draft it from the week's meeting notes and the project tracker, and an account manager approves each one before it goes out.

This Friday's update reads well. It says the launch is on track for the 14th. The account manager reads it, clicks approve, and it goes to the client.

What did that approval check? Tuesday's meeting notes say legal signed off on the launch copy. The tracker still shows it waiting on legal. Did she see both? If she had, could she have changed the date the update promised, or was that someone else's call? The record shows an approval. It doesn't show whether the person approving had what she needed to change what happened next.

That gap shows up in a lot of human-in-the-loop AI. The loop gets built as a queue of approvals, and the queue fills up with clicks. A person in the loop only helps when they have a job to do there, the context to do it and the authority to change the outcome.

We find it more useful to look for four moments where a person changes the work: setting direction before it starts, rescuing missing context while it runs, making the call when the facts allow more than one honest answer, and checking the result before it's used.

What human in the loop means here

The phrase also covers people correcting a model in training, but here we mean the judgment people bring to AI work once it's running. NIST's AI Risk Management Framework, in its appendix on human-AI interaction, says human roles in decision making and oversight "need to be clearly defined and differentiated," and that human-AI configurations "can span from fully autonomous to fully manual." Some systems, like one used to improve video compression, may not need human oversight at all, while others may specifically require it. For the Friday update, that means knowing who checks the facts, who owns the date and who approves the send.

Approval earns its place right before a commitment that's hard to take back, and the Friday update is one. Once a client reads that the launch is on track for the 14th, they may tell their own team or stop asking, and a correction can't unsend the first message. So keep the approval. What failed in the opening is that nothing made sure the account manager would see the contradiction, or that the person who owns the launch date would decide how to describe it.

The four human moments

Siebert and colleagues propose four properties for keeping people in meaningful control of AI systems, and two of them describe the Friday update's problem. People and AI should have "appropriate and mutually compatible representations" of the situation, so the account manager should be looking at the same notes and tracker the AI drew from. And "responsibility attributed to a human should be commensurate with that human's ability and authority to control the system," so she shouldn't answer for a date she had no power to change.

We use the four moments to check whether those two things are true each time a person touches the work. They're questions to ask again whenever the work changes, not four more sign-offs.

Set direction before the work starts

Direction is what the AI and the people around it need to know before work begins, including who is allowed to widen the scope later. For the Friday update, it might say: the reader is the client's marketing lead; draw only from this week's meeting notes and the tracker; leave out the internal budget discussion; if the notes and the tracker disagree, flag it rather than pick one; the account lead decides anything about dates.

Direction comes back up when something important changes. If the client asks for their CEO to be copied, the audience has changed, and so has what belongs in the update.

Rescue missing context while the work runs

Halfway through the draft, the contradiction turns up. The notes say legal signed off. The tracker says it's still with legal.

Not every unknown needs a person, but this one changes what the update can honestly say. You can't count on any system to spot every gap like this by itself, which is why the direction asked for disagreements to be flagged.

The person to ask is whoever knows: the project manager who was in Tuesday's meeting. She might say legal approved the copy and nobody updated the tracker, or that legal approved the headline but not the body. Knowing the fact doesn't make her the one who decides what to tell the client. And the rest of the update, like design progress and next week's meetings, can keep moving while the one question is open.

Make the call before anyone commits

Say the project manager comes back with this: legal approved the headline, still has questions about two claims in the body, and expects to finish by Wednesday. Now the facts are clear and the decision isn't.

The update could say the launch is on track, that it's at risk, or that it's on track if legal finishes by Wednesday. Each is defensible, and each sets a different expectation with the client.

That's a trade-off, so it goes to someone with the authority to make it. If your agency already has a rule, like always flagging a date that depends on an open item, the rule decides. If not, the account lead gets the options, what each commits the agency to, and how easy each would be to walk back.

Verify before the update is used

Before the update goes out, someone or something checks it against what it's for. Some of that can be automatic: every date matches the tracker, and nothing from the budget discussion slipped in. The rest needs a person. The account manager reads the line about legal against what the project manager said, and can fix it, cut it or hold the update until Wednesday. Now her approval means something, because she knows what she's checking it against.

Checking doesn't stop once the update is sent. NIST's March 2026 report on monitoring deployed AI systems names "scaling human-driven monitoring alongside rapid rollouts" as a barrier and leaves "balancing automated vs. human-validated monitoring" as an open question. For one Friday update, a person reading it is plenty. For forty clients a week, you'd have to decide which updates a person reads and what gets checked another way.

A person in the loop isn't a quality guarantee

The research is mixed. De-Arteaga, Fogliato and Chouldechova studied the screening tool on Allegheny County's child welfare hotline during a period when a technical glitch made some risk scores wrong. Call workers were less likely to follow the tool when the displayed score was wrong, "even when overriding the recommendation requires supervisory approval." That's one county and one tool, looked at after the fact.

What the person is shown matters, too. Chen and colleagues had 26 people think aloud while making predictions with an AI's help. The authors note the group skewed highly educated and experienced with machine learning. When the AI scored how much each input counted toward its answer, people went along with it more often when it was wrong. When it showed similar past cases instead, they did better. Alufaisan and colleagues had people predict income and reoffending. Seeing the AI's prediction tended to make them more accurate, but the authors found "no conclusive evidence" that the explanations they tested made a meaningful difference.

None of this was about status updates, and it doesn't add up to a rule. But if the update reaches the account manager without the tracker entry and the project manager's answer beside it, she's mostly reviewing how the sentences read.

Write the moment down

Writing a moment down won't prove the control works, but it will show you where it can't change anything.

Field Question The Friday update
Purpose What is this control for? An accurate launch status for the client.
Consequence What changes, and can it be undone? The client plans around the date. A correction can follow; the first message stays sent.
Trigger What sets it off? The notes and the tracker disagree about legal.
Job Direction, context, a decision or verification? Context first, then a decision.
Evidence What has to be visible? Tuesday's notes, the tracker entry, the project manager's answer.
Capability What does the person need to know? The client, and what "at risk" will set off on their side.
Authority What can they change, hold or stop? The wording, the send, or a call to the client.
Accountability Who answers for the outcome? The account lead, who made the call.
Next state What happens after? It goes out with the conditional date, or waits until Wednesday.
Recourse Who can question or correct it afterward? The client can reply, and the agency corrects it.

Escalations people can act on

When the contradiction gets flagged, what reaches the account lead decides whether she can act on it. She needs the conflict, what the project manager found, the options and what's ready to go. And it should come to her, because the date is hers to call.

For the Friday update, that might read: "Tuesday's notes say legal approved the launch copy. The tracker says it's still with legal. The project manager confirms legal approved the headline and expects to finish the body by Wednesday. Options: say on track, say at risk, or say on track pending legal by Wednesday. The rest of the update is ready." Compare that with "Please review" next to a draft.

Accountability needs real control

Madeleine Clare Elish introduced the "moral crumple zone" to describe how responsibility "may be misattributed to a human actor who had limited control over the behavior of an automated or autonomous system." Her paper isn't about approval buttons, but the risk is easy to picture. If the launch slips after the client heard "on track," the account manager's approval is on the record. If she never saw the tracker entry and had no say over the date, the record points at the one person who couldn't have changed the outcome.

Where the four moments stop

The four moments don't replace testing the system, security, legal review or a way for the people affected to challenge a result. They can't fix bad inputs either. If the tracker is routinely weeks out of date, no review step will make the Friday update reliable. Sometimes the right answer is to fix the source, narrow the work, bring in a qualified reviewer, or not use AI for it at all.

Redesign one workflow

Pick one workflow you already run with AI, like the Friday update, and go through it:

  1. List every place a person, a rule or an automated check touches the work.
  2. Name what each one is for and, where it fits, which moment it serves.
  3. For each, write down what triggers it, what evidence it sees, what the person needs to know, what they're allowed to change and what happens next.
  4. Mark any control with no clear job, and any consequence nothing is in place to catch.
  5. Before you remove or change a control, check whether it's there for another reason, such as a legal requirement, a policy, separation of duties or the rights of the people affected.

Running the Friday update on Supanova

Say the Friday update ran on Supanova atoms. You choose how often your workforce checks in, from approving every step to working more independently, and finished work comes back for your review. That setting decides when the account manager gets asked, and the four moments are how you make sure she has what she needs when she is.

So write the direction down first: who reads the update, which sources count, and that when the notes and the tracker disagree, the draft says so instead of picking one. Then, when the update comes back for review, she can see the disagreement, ask the project manager, and leave the launch date to the account lead.

Before you add another approval step, look at the one you already have. What does the person clicking it get to see, and what are they allowed to change?

Explore Supanova

Sources

  1. NIST AI RMF 1.0, Appendix C: AI Risk Management and Human-AI Interaction
  2. NIST AI 800-4: Challenges to the Monitoring of Deployed AI Systems (March 2026)
  3. Siebert et al., Meaningful human control: actionable properties for AI system development, AI and Ethics (2022)
  4. De-Arteaga, Fogliato and Chouldechova, A Case for Humans-in-the-Loop: Decisions in the Presence of Erroneous Algorithmic Scores (CHI 2020)
  5. Chen, Liao, Wortman Vaughan and Bansal, Understanding the Role of Human Intuition on Reliance in Human-AI Decision-Making with Explanations (2023)
  6. Alufaisan et al., Does Explainable Artificial Intelligence Improve Human Decision-Making? (2020)
  7. Elish, Moral Crumple Zones: Cautionary Tales in Human-Robot Interaction, Engaging Science, Technology, and Society (2019)
Human-in-the-Loop AIAI OversightAI GovernanceAI Approval WorkflowHuman-AI collaboration