Human-in-the-Loop AI: When Human Review Actually Helps

Your team is about to let an AI system draft decisions that people used to make from scratch. Someone proposes the safeguard everyone expects: “A human stays in the loop.” The room relaxes, and the meeting moves on.

That sentence ends the discussion without answering the question behind it. Does the reviewer make the decisions better than the AI would on its own? Can the reviewer spot the errors that matter, or does a person mostly add time and a signature? And when review does help, is the gain worth the hours it consumes?

This guide shows you how to answer them before you write the policy. You compare three conditions (human alone, AI alone, and human plus AI), record six measures, and decide with a worksheet.

What human-in-the-loop AI means in this guide

Human-in-the-loop AI here means a workflow in which a person reviews, edits, approves or overrides an AI system’s output before it takes effect. The output might be a draft decision, a classification, a summary or a recommended next step. The person might check every case or only the cases routed to them.

The phrase has a second meaning in machine learning: people labeling data or rating outputs to train and improve a model. That is a different job with different measures, and this guide does not cover it.

Why “better than before” is the wrong test

Most teams judge an AI-assisted process against the process it replaces. That comparison skips a third option: letting the AI’s output stand without review. A 2024 meta-analysis in Nature Human Behaviour by Michelle Vaccaro, Abdullah Almaatouq and Thomas Malone names the difference. Human augmentation means the combination beats the human working alone. Human-AI synergy means it beats both the human alone and the AI alone.

The researchers pooled 370 effect sizes from 106 experiments published between January 2020 and June 2023, each measuring all three conditions. Using a standardized difference (Hedges’ g), they found that combinations beat humans alone by a medium-to-large margin (g = 0.64). But combinations did worse than the better of the human or the AI alone (g = −0.23, a small difference).

Two moderators matter. Combinations lost ground in decision tasks, where people chose among fixed options, and fared better in creation tasks such as writing, though that estimate was less certain. When people alone outperformed the AI, combinations tended to beat both (g = 0.46); when the AI outperformed people, combinations tended to fall short of the AI (g = −0.54). The authors suggest that people who are less accurate than the AI are also poor at judging which of its answers are wrong.

An experiment by Ángel Alexander Cabrera and colleagues shows the pattern and its limits. Participants used a simulated AI fixed at about 73% accuracy, described to them as 90% accurate. In detecting fake hotel reviews, people alone scored about 55% and people shown the AI’s predictions about 62%, still below the AI. In classifying bird images, people alone scored about 81% and people shown the predictions about 83%; people who also saw where the AI tends to fail reached about 90%, beating both.

On satellite images, people alone were already about 90% accurate, and no AI condition reliably changed that. (These values are read from the paper’s chart.) Who is better alone predicts the direction on average, not in every case.

What older studies can and cannot tell you

The meta-analysis covers experiments published by mid-2023, many with crowdworkers and some with simulated AI. It supplies a method, not a verdict on your tool.

In a 2024 trial by Ethan Goh and colleagues, physicians given GPT-4 scored a median of 76% on diagnostic reasoning vignettes, against 74% for physicians without it. In an exploratory analysis, three runs of GPT-4 answering alone scored a median of 92%. Only measuring the AI alone exposed that gap. The trial’s formal test compared GPT-4 alone with physicians working without it, and GPT-4 scored 16 points higher.

A worked example: refund decisions

The rest of this guide follows one hypothetical case. The organization, numbers and results are invented to show the method.

An online retailer handles about 10,000 refund requests a month. Agents check the order history and shipping records, then approve, deny or escalate each request to a senior agent. The team wants an AI system to draft an approve-or-deny decision for each request, and proposes that agents review every draft before the customer sees it.

The operations lead wants three answers. Would reviewed drafts beat the current process? Would they beat unreviewed drafts? And what would review cost in agent time?

Step 1: Decide what counts as a correct decision

You cannot count errors without a standard. The team forms an audit panel of two senior agents, who decide each sampled case with the full record and no time pressure, without seeing the AI draft or knowing the condition. A third senior agent settles disagreements.

Score every condition on its final decision. For an escalated case, that is the senior agent’s decision, and the senior’s time counts toward that condition’s cost. Use later facts, such as a confirmed fraud, where they exist. The panel is still fallible, so the comparison is only as good as this standard.

Define a serious mistake before you see results: here, denying a valid refund worth more than $200, or approving a refund on an order with confirmed fraud. Track these separately, because a review that reduces minor slips while missing costly errors is not doing its main job.

Step 2: Set up three conditions

The team draws 600 requests from one month and randomly assigns each to one of two arms of 300. Random assignment makes the arms comparable, though chance still plays a part, which Step 4 deals with. The randomized controlled trial entry and the research methods guide explain the design.

• Human alone. An agent handles the request with the current process. The AI still drafts a decision, but the draft is hidden and only recorded.
• Human plus AI. An agent sees the AI draft and reviews it, exactly as the proposed policy would require.
• AI alone. No third set of cases and no live AI decisions are needed. Score the drafts in the human-plus-AI arm against the panel.

The hidden drafts in the human-alone arm check the randomization: the AI should score about the same in both arms. Test the workflow you intend to run, including whether agents see the draft before forming their own view.

Step 3: Record six measures

For each condition, record:

1. Quality: the share of decisions that match the audit panel.
2. Serious mistakes: decisions that meet your definition.
3. Caught errors and missed correct advice: in the human-plus-AI arm, wrong drafts corrected, correct drafts changed to a wrong decision, and drafts accepted unchanged.
4. Time: agent minutes per case, from the case log.
5. Review burden: total agent hours at your expected volume, including senior time on escalations.
6. Escalation: the share of cases sent to a senior agent and the senior minutes they used.

Read measure 3 two ways. The counts give the net effect: if overrides of correct drafts outnumber corrections, review is making decisions worse. The rates show whether reviewers can tell good drafts from bad: corrections divided by wrong drafts (the catch rate), and overrides divided by correct drafts (the false-alarm rate). These are the hit and false-alarm rates of signal detection theory; if they are close, reviewers are changing drafts almost at random.

The two readings can disagree. When the AI is usually right, correct drafts vastly outnumber wrong ones, so even a low false-alarm rate can break more answers than a good catch rate fixes.

Step 4: Read the results

Here are the invented results for the retailer.

Read the table in three passes.

First, compare with the best single condition. The AI alone (92.0%) beat agents working alone (89.0%). Reviewed drafts (96.0%) beat both. In Vaccaro’s terms, this is synergy, not just augmentation. My test for any intervention is the same one I use for nudges: show what changes, compared with what, and at what cost.

Second, find where the gain came from. Reviewers corrected 15 of 24 wrong drafts and broke 3 of 276 correct ones: 276 − 3 + 15 = 288. A 62.5% catch rate against a 1.1% false-alarm rate means they could tell bad drafts from good. Serious mistakes fell from 9 to 3: reviewers caught 7 of the 9 serious draft errors, and one override created a new one. Most wrong drafts involved parcels with an open carrier investigation; agents check those records routinely, and the AI’s inputs did not include them.

Third, ask whether the gain could be chance. The human-plus-AI arm gives a paired comparison: the same 300 cases before and after review, where only the changed cases carry information. For overall accuracy, 15 fixed against 3 broken makes chance an unlikely explanation by an exact sign test (McNemar’s exact test). For serious mistakes, 7 fixed against 1 broken points the same way, but the same test does not rule out chance at the 5% level. Confirm it on a larger sample of high-value requests.

Step 5: Price the review

At 10,000 requests a month, with 15 senior minutes per escalated case (also invented):

• Human plus AI: 2.5 minutes × 10,000 ≈ 417 hours, plus 600 escalations × 15 minutes = 150 hours. Total: about 567 hours.
• Human alone, the current process: 6.0 minutes × 10,000 = 1,000 hours, plus 1,000 escalations × 15 minutes = 250 hours. Total: 1,250 hours.
• AI alone: no agent time, because the drafts approve or deny and nothing is escalated.

Against the current process, reviewed drafts are more accurate and take less than half the agent time, which supports adopting AI drafts once you add the tool’s running cost. The review decision rests on the comparison with AI alone. If the sample rates hold, the extra 567 hours of review buy roughly 400 more correct decisions a month, about 200 of them serious mistakes avoided. That is about 2.8 extra hours per serious mistake avoided and 1.4 per extra correct decision.

Is that gain large enough to matter? Suppose, with invented figures, that a serious mistake costs the retailer about $250 in chargeback fees and lost goodwill, and an agent hour costs $35. A serious mistake is then worth about 7 agent hours, well above the 2.8 hours it takes to prevent one. The worksheet also counts ordinary corrections at a smaller value, but serious mistakes carry most of the total. Review earns its place here, provided the larger sample confirms the serious-mistake result.

When review adds little: low-value refunds

Now change one condition. The team runs the same evaluation on a different queue: requests under $25 for items that are cheap to replace. These are again invented results.

Review moved accuracy from 98.0% to 98.3%, one net decision in 300 resting on three changed cases. That is indistinguishable from no effect. Even at face value, reviewing every draft would cost 250 agent hours per 10,000 requests for about 33 extra correct decisions. That is roughly 7.5 hours per extra correct decision, on refunds worth under $25.

Here review adds little. Let the drafts stand, audit a weekly random sample against the panel standard to catch drift, and send agents only cases that meet a defined exception rule. This arrangement is often called human on the loop: people monitor the system and handle exceptions instead of approving every output.

What changed between the queues? Not the agents or the tool. In both, the AI beat agents working alone, and both are decision tasks: the two conditions under which Vaccaro’s meta-analysis found that combinations, on average, fall short of the AI. The first queue goes against those averages, and it has a specific cause. On the cases where the AI erred, agents held evidence it never saw, the carrier records, so they were more accurate than the AI on its mistakes while less accurate overall.

In the low-value queue, agents saw the same order data as the AI, errors were rare, and each was cheap. Do not assume your queue looks like the first one; the averages point the other way.

Can your reviewers detect the errors that matter?

The carrier records point to the first condition for useful review: reviewers need evidence the AI did not use. A reviewer who sees only the draft and its explanation is re-reading the AI’s own case. Asking the model whether its draft is right is not an independent check either, and the article on AI sycophancy explains why a model’s answer can bend toward what the person asking appears to want.

Telling reviewers to be careful is no substitute; the automation bias entry explains why vigilance instructions, explanation boxes and confidence displays are not dependable fixes. Fluent output adds a newer problem. In a 2026 study of Boston Consulting Group consultants, GPT-4 access lowered the share reaching the right answer on a case built so that GPT-4 would struggle. Graders also rated the AI-assisted answers as more persuasive, even the wrong ones; the AI adoption guide covers the study in detail.

Some designs reduce overreliance on wrong advice, at a cost. In a 2021 experiment, Zana Buçinca, Maja Barbara Malaya and Krzysztof Gajos gave 199 online participants a simulated AI that was right 75% of the time. The task was choosing which ingredient to replace in a meal to cut carbohydrates. Their cognitive forcing functions made people decide before seeing the AI’s suggestion, request it, or wait 30 seconds for it.

Across all questions, people picked the right ingredient about 57% of the time with these designs and 56% with simple explanation designs (with or without a confidence display), against 42% with no AI. So the forcing designs gave no overall gain over simple explanations. Their benefit came where the AI was wrong: about 27% correct, against 8% with simple explanations and 49% with no AI. Participants also rated the most effective designs least favorably, so test a decide-first step as a variant of your human-plus-AI condition rather than assuming it works.

The UK Cabinet Office’s Mitigating ‘Hidden’ AI Risks Toolkit, published in June 2025, warns that human oversight without supporting measures “won’t be sufficient.” Reviewers may lack the expertise, time or authority to challenge AI outputs. Its three conditions make a practical checklist:

• Expertise: reviewers know the subject well enough to judge whether an output is accurate.
• Time: the workflow allows a real check, not a glance.
• Authority: reviewers can challenge an output, or the use of the tool, and escalate when needed.

The toolkit is practice guidance; your comparison supplies the test. If the errors you fear most are rare, a 300-case sample may contain few of them. In a test environment, add cases with known errors of that type and count how many reviewers catch.

Accountability is a separate requirement

Sometimes a person must make or sign a decision whatever the accuracy results show. A statute, regulation, contract or internal policy may require a named official to decide a benefit claim, or a manager to approve every denial above a threshold. That requirement tells you a person must be in the loop. It does not tell you the person improves the decision.

Keep the two questions apart. Suppose a hypothetical benefits program requires a caseworker to make each eligibility decision, and the agency introduces AI drafts. If caseworkers reviewing drafts do worse than the drafts alone, the agency cannot remove them. It can change the conditions: source documents rather than only the draft’s summary, a decide-first step, or more time on complex cases. Then it runs the comparison again, without reporting the signature as evidence of accuracy.

Try it on a new case

A city housing office tests an AI tool that flags rental-assistance applications with missing documents. Caseworkers do this check by hand, using the same uploaded files the tool reads. The office runs the three-condition comparison with an audit panel and 300 applications per arm (hypothetical numbers):

• Caseworkers alone: 255 of 300 applications classified correctly (85%).
• AI alone, scored on the drafts in the review arm: 279 of 300 (93%).
• Caseworkers reviewing the AI’s flags: 270 of 300 (90%). In that arm, caseworkers corrected 6 of the 21 wrong flags and overrode 15 of the 279 correct ones.

What does this tell the office, and what should it do next?

Answer. Reviewed results beat caseworkers alone but fall below the AI alone: 279 − 15 + 6 = 270. The rates show why. Caseworkers changed 29% of wrong flags and 5.4% of correct ones, so they can partly tell good flags from bad. But correct flags outnumber wrong ones 279 to 21, so even a 5.4% false-alarm rate breaks more flags than the catch rate fixes. Unlike the refund agents, caseworkers had no evidence the tool lacked.

Next, the office should examine the 15 overridden flags, then redesign the review. It might limit review to the kinds of flags where the audit shows the tool errs most often, which raises the share of reviewed flags that are wrong. Or it might give caseworkers a source the tool cannot read, such as the agency’s record of documents already on file. If a caseworker must legally decide each application, the review stays either way, and the office runs the comparison again.

Decide where review belongs

Go back to that meeting. “Keep a human in the loop” is a hypothesis about a particular task, tool and team. In some queues, review is the most valuable part of the system. In others, it is a cost with a signature attached.

Pick one queue where the review policy is still undecided, draw a sample, and fill in the review-evaluation worksheet before the policy is written. If you have not yet settled which tasks should get AI drafts at all, start one step earlier with AI adoption at work: choose the task before the tool.

References

• Buçinca, Z., Malaya, M. B., & Gajos, K. Z. (2021). To trust or to think: Cognitive forcing functions can reduce overreliance on AI in AI-assisted decision-making. Proceedings of the ACM on Human-Computer Interaction, 5(CSCW1), Article 188.
• Cabinet Office, Government Communications (2025). The Mitigating ‘Hidden’ AI Risks Toolkit. GOV.UK, published June 4, 2025.
• Cabrera, Á. A., Perer, A., & Hong, J. I. (2023). Improving human-AI collaboration with descriptions of AI behavior. Proceedings of the ACM on Human-Computer Interaction, 7(CSCW1), Article 136.
• Dell’Acqua, F., McFowland, E., III, Mollick, E., Lifshitz, H., Kellogg, K. C., Rajendran, S., Krayer, L., Candelon, F., & Lakhani, K. R. (2026). Navigating the jagged technological frontier: Field experimental evidence of the effects of artificial intelligence on knowledge worker productivity and quality. Organization Science, 37(2), 403–423.
• Goh, E., Gallo, R., Hom, J., et al. (2024). Large language model influence on diagnostic reasoning: A randomized clinical trial. JAMA Network Open, 7(10), e2440969.
• Vaccaro, M., Almaatouq, A., & Malone, T. (2024). When combinations of humans and AI are useful: A systematic review and meta-analysis. Nature Human Behaviour, 8(12), 2293–2303.