Short answer: Human in the loop means a person must approve an AI system’s output before it takes effect. The design question is where. Put approval points on steps where a wrong output is costly, hard to undo or hard to verify. Elsewhere, let the system act with logs and sampled checks, and measure whether reviewers really catch errors.
Human in the loop means a person must approve an AI system’s output before it takes effect. The real design question is not whether to add a person but where.
Put the approval point on steps where a wrong output is costly, hard to undo or hard to verify, and let the system act on its own, with logs and sampled checks, where mistakes are cheap and reversible. Then make the review real: show the sources and what changed, make approve, edit and reject one click each, capture reasons, and measure whether reviewers actually catch errors.
This matters more each quarter. In McKinsey’s 2026 State of AI survey, 40 percent of respondents from organizations with more than $1 billion in revenue report scaling AI agents, up from 27 percent a year earlier. In the 2025 edition, 51 percent of respondents from organizations using AI reported at least one negative consequence, and nearly a third reported consequences from inaccuracy.
Yet oversight varies widely: in McKinsey’s March 2025 survey, 27 percent of gen AI users said employees review all generated content before use, and a similar share said 20 percent or less gets checked. These are self-reported surveys, but the gap is plain.
What does human in the loop mean?
The term has two meanings. In machine learning, it describes people labeling data or rating outputs to train a model. In operations, which is what this guide covers, it means a person sits inside the workflow and nothing takes effect until they act.
The vocabulary comes from the debate on autonomous weapons. Human Rights Watch’s 2012 report Losing Humanity sorted systems by whether they act only on a human command, under a human who can override them, or without human input. The same three modes fit a business workflow:
Mode | What it means in a workflow |
|---|---|
Human in the loop | The system drafts or proposes. A person approves, edits or rejects before anything is sent, paid or saved. |
Human on the loop | The system acts on its own. A person watches metrics, reviews a sample and can intervene or stop it. |
Human out of the loop | No routine human review. People read logs and fix problems after the fact. |
Human in the loop vs human on the loop is a trade between prevention and cost. In the loop spends reviewer time on every case and stops errors before anyone sees them. On the loop spends a fraction of that time and finds errors after they happen, at a sample rate you choose.
For AI agents, the approval point belongs in front of any action that changes something outside the system: sending a message, making a payment, deleting or updating a record. OpenAI’s guide to building agents says sensitive, irreversible or high-stakes actions should trigger human oversight until confidence in the agent’s reliability grows. So approval points are not permanent. A step can earn its way out of one. If you are still deciding what kind of system you need, start with agent vs chatbot vs workflow.
Where should the approval point go?
Place approval points per step, not per workflow. An invoice workflow might read the document, match it, draft a supplier note and release payment. Each step gets its own answer. Score each on three factors.
Consequence of a wrong output. Low: an internal note is wrong and someone fixes it in a minute. Medium: a customer or partner sees it, or it costs a bounded amount of money. High: it harms a person, breaches a law or contract, or moves serious money.
Reversibility. Easy: undone before anyone is affected. Hard: undone with effort, such as a correction email or a credit note. Irreversible: a payment sent, a filing submitted, a record deleted, a statement published.
Put consequence down the side of a 3×3 grid and reversibility across the top. The low-and-easy corner needs no routine review. The high-and-irreversible corner needs a person every time, and perhaps two. In between, the bands look like this:
Out of the loop: low consequence, easy to undo. Keep logs and metrics.
On the loop (act, log, review a sample): low consequence and hard to undo, or medium consequence and easy to undo.
In the loop until the numbers justify a change: low and irreversible, medium and hard to undo, or high and easy to undo.
In the loop, always: medium and irreversible, or high and hard to undo.
A second person, or keep it manual: high consequence and irreversible.
Then apply the third factor: verifiability. If a rule or a reviewer can check the output against a source in seconds, such as an invoice total against the order, move the step one band lighter. If nobody can check it quickly, move it one band heavier, or redesign the step so the evidence is on screen. Never move the high-and-irreversible band lighter.
Model confidence belongs here only if it means something. A confidence score counts once you have checked, on your own past cases, that high-scoring outputs really are right more often. A model’s own claim that it is sure does not count.
Why do human reviewers rubber-stamp AI output?
Because people defer to machines, and the evidence is consistent. The EU AI Act names the effect in Article 14(4)(b): the tendency of “automatically relying or over-relying” on a system’s output, which it calls automation bias.
A systematic review of 74 studies of clinical decision support found erroneous advice was more likely to be followed with the system than without it (risk ratio 1.26). In prospective studies, users switched from a correct answer to a wrong one after consulting the system in 6 to 11 percent of cases.
In a Radiology experiment, 27 radiologists read mammograms with a purported AI. When it suggested the wrong category, inexperienced readers’ accuracy fell from 79.7 to 19.8 percent. Very experienced readers fell from 82.3 to 45.5 percent.
A review in Human Factors concluded that automation bias affects experts as well as novices, and that training or instructions alone do not prevent it.
A 2024 meta-analysis of 106 experiments in Nature Human Behaviour found human-AI combinations performed worse on average than the best of humans or AI alone (Hedges’ g = −0.23), with losses in decision tasks. Where humans outperformed the AI alone, the combination gained.
Volume makes it worse. Anthropic reports that users of its coding tool approve 93 percent of permission prompts, and names the result: “Over time that leads to approval fatigue, where people stop paying close attention to what they’re approving.” Ben Green’s survey of 41 oversight policies warns that human oversight can give a false sense of security. Review without design is a ritual, not a control.
How do you design an approval screen people actually use?
Design the screen so that checking is easier than trusting. A review of verification complexity found automation bias in single tasks with high verification complexity and suggests reducing cognitive load. Every item below cuts the effort of a real check.
Show sources, not arguments. Put the source passage next to the claim it supports, highlighted. A model’s narrative explanation is not evidence. In a CHI 2021 study, explanations made people more likely to accept the AI’s recommendation whether or not it was right. In an ICIS 2025 field experiment with 228 evaluators and 3,002 screening decisions, screeners with AI were 19 percent more likely to align with its recommendations, and narrative persuasiveness drove that alignment.
Show what changed. Highlight the difference between the output and its input, template or previous version, so the reviewer checks the delta rather than rereading everything.
Make approve, edit and reject one click each, the same size. Edit inline, not in another tool. Microsoft’s guidelines for human-AI interaction include “support efficient correction” and “support efficient dismissal.”
Capture a reason on every edit and reject. Five to eight reason codes and an optional note. This is your best improvement data.
Flag, do not score. Highlight the fields that failed a rule check or that the system was unsure of. Studies in the 74-study review found that updated confidence levels shown with each piece of advice made reliance more appropriate, and that more prominent advice increased bias.
Ask for the reviewer’s view first on the hardest calls. In a CSCW 2021 study with 199 participants, such “cognitive forcing” designs significantly reduced overreliance, but users rated them least favorably. Use them on the few decisions that justify the friction.
Name the approver and cap the queue. The same review lists accountability as a mitigator, and workload and time pressure among the factors that shape bias.
How do you know the human check is working?
Measure the reviewer, not only the model. NIST’s AI Risk Management Framework suggests collecting data on how often, and why, people overrule AI output in deployed systems. Track five numbers per step:
Override rate: the share of items edited or rejected, split by reason code. Near zero means either an excellent model or an inattentive reviewer. The next measures tell you which.
Edit distance: how much of the output changed on edits, as a share of characters or fields. A falling trend means the model is improving or reviewers are giving up.
Time per review: the median and the fast tail. If a two-page draft is routinely approved in four seconds, nobody read it.
Sampled audits: a second qualified person rechecks a random sample of approved items each week. The approved-error rate is your headline number. Compare it with the error rate before AI, which is why you measure a baseline first.
Seeded errors: slip known-wrong outputs into the queue and count how many reviewers catch. Airport screening has done this for years: according to the US Government Accountability Office, threat image projection superimposes images of guns, knives and explosives onto real bags to test screeners.
Add one downstream measure, such as complaints, corrections or reversals after approval. It catches what the queue missed.
When can a step move from in the loop to on the loop?
When the numbers say the person is no longer adding much, and the step sits in a band that allows it. Our checklist:
Several hundred reviewed cases over at least four weeks, on a stable mix of inputs.
A low, stable override rate whose reasons you understand.
A high catch rate on seeded errors, so a low override rate is not inattention.
An audited approved-error rate at or below the pre-AI baseline.
No high-consequence misses in the period.
Move in stages: review everything, then a 20 percent sample, then 5 percent, with flagged cases still routed to a person. Write down what sends a step back into the loop: a model or prompt change, a new input type, an audit spike or an incident.
Moving out can make oversight better, not worse. NIST’s generative AI profile warns that over time people may over-rely on these systems. A sharp weekly audit beats a tired reviewer clicking through hundreds of items.
Worked example: a customer-service inbox
This example is hypothetical. A consumer-health brand receives about 1,500 customer emails a week and adds AI to three steps. At launch, every step sits one band heavier than the grid says, until the numbers exist.
Order-status replies drafted from shop data. Medium consequence, easy to correct with a follow-up, and verifiable against the order record: on the loop by the grid, in the loop at launch.
Refunds up to €40 for damaged goods. Low consequence but irreversible once paid: in the loop.
Any email mentioning a side effect or health condition. High consequence, hard to reverse: in the loop always, routed to the quality team, with the AI limited to tagging and a holding reply.
After six weeks, the order-status step shows 2,400 drafts reviewed, a 2.1 percent override rate (mostly tone), a median review time of 11 seconds, 28 of 30 seeded wrong dates caught, and one error per 100 in the weekly audit against three per 100 before AI. It moves on the loop with a 10 percent sample, and any reply with missing shop data still goes to a person.
Refunds show a 9 percent override rate, mostly coded “policy exception.” They stay in the loop while the team fixes the refund rules. The health step never moves.
A template for documenting approval points
Keep one entry per step, reviewed quarterly. It doubles as evidence of oversight if a regulator, auditor or client asks.
Field | What to write |
|---|---|
Step and output | What the system produces at this step |
Grid placement | Consequence · reversibility · verifiability, with one line of reasoning |
Mode | In, on or out of the loop, and the sample rate |
Approver | Named role, backup, and authority to override or stop |
Review screen | Sources shown · changes highlighted · flags |
Actions and reasons | Approve · edit · reject · escalate, with the reason codes |
Time budget | Expected seconds per review and the queue cap |
Escalation | Cases that always go to a person or a specialist |
Metrics | Override rate · edit distance · review time · audit and seeded-error results |
Change rules | Criteria to move lighter, and triggers to move back |
Stop and logs | Who can halt the step, and how long logs are kept |
What does the law require?
This is general information, not legal advice.
Under the EU AI Act, Article 14 requires high-risk systems to be designed so people can effectively oversee them, stay aware of automation bias, and “disregard, override or reverse the output” or stop the system. Article 26(2) requires deployers to assign oversight to people with “the necessary competence, training and authority.” Since the Digital Omnibus, these duties apply from 2 December 2027 for Annex III uses such as recruitment and credit scoring, as of October 2026. Our EU AI Act guide covers which uses count.
Data protection applies now. Article 22 GDPR restricts decisions based solely on automated processing that significantly affect people. European regulators’ guidelines say the human involvement must be “meaningful, rather than just a token gesture,” carried out by someone with the authority and competence to change the decision. A rubber stamp does not count.
Where to start
Take one workflow you run or plan to automate. List its steps, place each on the grid, and write the template entry for the steps that land in the loop. Then decide what the reviewer will see on screen and which five numbers you will check every week.
Want to place the approval points in one of your workflows? Book a fit check. The system flags, you decide. To pick the workflow first, read how to choose your first AI workflow.
Frequently asked questions
What is the difference between human in the loop and human on the loop?
In the loop, a person acts on every case before anything takes effect. That stops errors before anyone sees them, at the cost of reviewer time. On the loop, the system acts and logs what it did, and a person reviews a sample. That costs less, but finds errors after they happen.
Where should the approval point go in an AI workflow?
Score each step, not the whole workflow, on the consequence of a wrong output and how easily it can be reversed. Then adjust for how quickly someone can verify it. High-consequence, irreversible steps need a person every time, and perhaps two. Low-consequence steps that are easy to undo need logs and metrics.
Why do reviewers rubber-stamp AI output?
Because people defer to machines. The EU AI Act calls this automation bias. A review in Human Factors found that it affects experts as well as novices. Volume makes it worse: Anthropic reports that users of its coding tool approve 93 percent of permission prompts. Training or instructions alone do not prevent it.
How do you know the human check is working?
Measure the reviewer, not only the model. Track the override rate, edit distance, time per review, a weekly audit of approved items by a second qualified person, and how many seeded errors reviewers catch. Add one downstream measure, such as complaints or corrections after approval.
Does the law require human oversight of AI?
For high-risk systems under the EU AI Act, yes, from 2 December 2027 for Annex III uses such as recruitment and credit scoring. Article 22 GDPR already restricts decisions based solely on automated processing, and European regulators say the human involvement must be meaningful. A rubber stamp does not count.
Related reading
Sources
McKinsey & Company, The state of AI in 2026: On the road to ROI, August 25, 2026.
McKinsey & Company, The state of AI in 2025: Agents, innovation, and transformation, November 5, 2025.
McKinsey & Company, The state of AI: How organizations are rewiring to capture value, March 12, 2025. Online survey of 1,491 participants, July 2024.
Human Rights Watch, Losing Humanity: The Case against Killer Robots, November 19, 2012.
OpenAI, A practical guide to building agents, 2025.
Regulation (EU) 2024/1689 (Artificial Intelligence Act), Articles 14 and 26.
Regulation (EU) 2026/1744 (Digital Omnibus on AI), Official Journal, July 24, 2026.
Goddard, Roudsari and Wyatt, Automation bias: a systematic review of frequency, effect mediators, and mitigators, Journal of the American Medical Informatics Association, 2012.
Dratsch et al., Automation Bias in Mammography, Radiology, May 2023.
Parasuraman and Manzey, Complacency and bias in human use of automation, Human Factors, June 2010.
Vaccaro, Almaatouq and Malone, When combinations of humans and AI are useful, Nature Human Behaviour, October 28, 2024.
Anthropic, How we built Claude Code auto mode, March 25, 2026.
Green, The flaws of policies requiring human oversight of government algorithms, Computer Law & Security Review, vol. 45, July 2022.
Lyell and Coiera, Automation bias and verification complexity: a systematic review, Journal of the American Medical Informatics Association, March 2017.
Bansal et al., Does the Whole Exceed its Parts? The Effect of AI Explanations on Complementary Team Performance, CHI 2021.
Lane et al., Narrative AI and the Human-AI Oversight Paradox in Evaluating Early-Stage Innovations, ICIS 2025 Proceedings.
Microsoft, HAX Toolkit: Guidelines for Human-AI Interaction, based on Amershi et al., CHI 2019.
Buçinca, Malaya and Gajos, To Trust or to Think: Cognitive Forcing Functions Can Reduce Overreliance on AI, CSCW 2021.
NIST, AI Risk Management Framework (AI RMF 1.0), January 2023.
NIST, Generative Artificial Intelligence Profile (NIST AI 600-1), July 2024.
US Government Accountability Office, Aviation Security: TSA’s Change to Its Prohibited Items List, GAO-07-623R, April 25, 2007.
Regulation (EU) 2016/679 (General Data Protection Regulation), Article 22.
Article 29 Data Protection Working Party, Guidelines on Automated individual decision-making and Profiling, WP251rev.01, revised February 6, 2018, endorsed by the European Data Protection Board.