botminds.aiRequest a demo
Guide · Human in the loop

Human in the loop vs human on the loop: who acts, and who can stop it.

In the loop, nothing happens until a person says yes. On the loop, the system acts and a person watches with a hand on the stop. Out of the loop, it acts alone.

This guide traces the three positions to their source in weapons policy and human-factors research, lays out the ten levels of automation between “human does everything” and “machine ignores the human”, shows how one AI agent can sit at different levels for different functions, and gives a way to place people by risk, reversibility and measured accuracy.

Updated · 6 sources

Read the guideHow Botminds does it
Human decides and actsMachine decides and acts
12345678910
  1. No assistanceThe person takes every decision and every action.
  2. Offers every optionThe computer lists the complete set of alternatives.
  3. Narrows to a fewThe computer cuts the options down to a short list.
  4. Suggests oneThe person still chooses and acts.
  5. Executes if approvedHuman in the loop: nothing runs without a yes.
  6. Executes unless vetoedHuman on the loop: a set time to say no first.
  7. Executes, then informsThe person is always told, after the fact.
  8. Informs only if askedThe person learns what happened on request.
  9. Informs if it decides toThe machine chooses whether to report at all.
  10. Ignores the humanDecides and acts alone: human out of the loop.

What is the difference between human in the loop and human on the loop?

Human in the loop means the system cannot complete an action until a person approves it — the person is a required step. Human on the loop means the system acts on its own within set limits while a person supervises and can intervene or halt it. Human out of the loop means the system acts with no human involvement once started. In human in the loop vs human on the loop, the difference is whether a person must say yes before the action, or can say stop during or after it.

Three positions

Human in the loop vs human on the loop vs human out of the loop.

The three terms describe where a person stands relative to an automated decision: inside it, above it, or absent from it.

approve before

Human in the loop

The system proposes; a person decides. The loop cannot close without them.

  • A required yes per action
  • Safest for irreversible acts
  • Throughput set by people
  • Risk: rubber-stamping
supervise during

Human on the loop

The system decides and acts within limits; a person monitors and can intervene, override or stop.

  • Acts without waiting
  • Person sees outcomes and alarms
  • Scales past headcount
  • Risk: missing the moment to act
Before each actionWhen the person actsDuring or after, by exception
Nothing happensDefault if the person is absentThe system keeps acting
Evidence for this decisionWhat the person needsSignals across many decisions
Approval, four-eyes reviewTypical controlsThresholds, sampling, alarms, a stop
Level 5: executes if the person approvesOn the ten-level scaleLevel 6: a set time to veto, then executes

Human out of the loop. Once started, the system selects and acts with no person involved — level 10 on the scale. Oversight moves to design, testing, limits set in advance and after-the-fact audit. Tying in, on and out to levels 5, 6 and 10 is a common reading of the scale; the original papers do not use the loop terms.

Where the terms come from

Where the terms come from: human factors and weapons policy.

The vocabulary was not invented for AI. It grew out of decades of research on supervisory control, and was made concrete by policy on autonomy in weapon systems.

  1. 1978Levels of automationSheridan and Verplank’s MIT report on undersea teleoperators sets out a ten-level scale of human–computer control.
  2. 1983Ironies of automationBainbridge: automating the easy parts leaves people the hardest job — monitoring — while their manual skills fade.
  3. 2000Types and levelsParasuraman, Sheridan and Wickens: automation applies to four functions, each at its own level.
  4. 2012DoD Directive 3000.09US policy on autonomy in weapon systems requires appropriate levels of human judgment over the use of force.
  5. 2023Directive reissuedUpdated 25 January 2023, keeping the human-judgment standard and senior review before development and fielding.
  6. 2024EU AI ActIn force from August 2024. Article 14 requires effective human oversight for high-risk AI systems, including a way to interrupt them.

How the Congressional Research Service maps the directive

  • Human in the loopSemi-autonomous systems: engage only targets a human operator has selected.
  • Human on the loopHuman-supervised autonomous systems: operators can monitor and halt an engagement.
  • Human out of the loopAutonomous systems: once activated, select and engage targets without further intervention.
Types and levels of automation

One AI agent, four functions, four different levels of autonomy.

Parasuraman, Sheridan and Wickens’s key point: “how autonomous is it?” has no single answer. A system can gather information at a high level of automation and still leave the decision to a person.

An invoice agent, function by function
01
Information acquisitionreads every incoming invoice and email, unattended
out of the loop
02
Information analysisextracts fields, matches to the purchase order, scores confidence
on the loop
03
Decision selectionrecommends pay, hold or query; routine matches decided automatically within limits
on the loop
04
Action implementationreleases a payment only after a person approves it
in the loop

Design at the function, not the system. Labelling a whole agent “human in the loop” hides where the person actually is. State the level for each function, and you can see exactly which decisions people own.

Decision guide

Human in the loop vs human on the loop: placing a person by risk and evidence.

Two properties of an action decide most cases: how much harm a wrong action does, and whether it can be undone. Measured accuracy then decides how far along the spectrum you can move.

High impact
On the loop, fast rollbackPrice changes, routing rules: act, watch the outcome, reverse quickly if it goes wrong.on the loop
In the loopPayments, filings, deletions, messages to customers: a person approves each one.in the loop
Low impact
Out of the loop, auditedTagging, sorting, drafting: let it run and sample the results.out of the loop
In the loop, then earn itStart with approval; move on the loop once audited accuracy proves the error rate.graduate
Easy to undoHard to undo

Has the accuracy for this decision been measured on your own data?

NoKeep a person in the loop. Every approval and correction is the evidence you are missing.
Yes, and it meets the targetMove on the loop: act automatically, send a blind sample to people, and alarm on drift.Read more →
Yes, by confidence bandSplit it: high calibrated confidence on the loop, the rest in the loop.Read more →
Pitfalls

Why human oversight fails, and how to design it so it works.

Putting a person somewhere in the loop is easy. Making their judgment count is not. Bainbridge’s ironies and the EU AI Act’s Article 14 point at the same failure modes.

Oversight that holds up
  • Counter automation biasArticle 14 asks that overseers stay aware of the tendency to over-rely on output. Hide the score on audit samples so reviewers judge the evidence.
  • Show the evidence, not just the answerA reviewer who cannot see the source page can only agree or guess.
  • Keep the volume humanRoute only uncertain or high-impact cases. A person approving hundreds of near-identical items stops reading them.
  • Give on-the-loop people timeA veto window shorter than the time it takes to understand the case is not oversight.
  • Provide a real stopArticle 14 names a “stop” button or similar procedure. Make halting safe and quick.
  • Keep skills aliveBainbridge’s irony: people who only monitor lose the skill to take over. Rotate them through real cases.
  • Record every judgmentWho approved, edited or stopped what, and why — for audit and for measuring the machine.
Human on the loop, run as a loop
  1. Act within limitsthe system decides cases inside its approved scope
  2. Samplea blind share of automatic decisions goes to people
  3. Monitorerror rates and drift are tracked against the target
  4. Intervenean alarm or a person pauses the automation
  5. Adjustthresholds, scope or the model change, then resume

limits are tightened or relaxed by the evidence

On Botminds

How Botminds puts people in, on and out of the loop.

On Botminds each position is a setting you choose per tool, per stage and per field — and the evidence to move a field along the spectrum is collected as people work.

Three positions, one platformchosen per tool, per stage, per field — and reversible
Tool approvalsin the loop: approve, reject or edit before an agent’s tool runs
Four-eyes reviewin the loop: two review stages, and a rule can require a different person at each
Per-field straight-throughon the loop: fields that clear their threshold flow on
Blind audit sampleon the loop: passed documents reviewed without scores
Drift alarmon the loop: switches a field back to review until a person restores it
Auto-decideper collection: high confidence and zero flags stand on their own

Read the full design in human in the loop, and the Academy lesson on human in the loop.

Questions people ask

Human in the loop vs on the loop: common questions.

The technology behind itHuman in the loopApprovals, review and calibrated confidence
What is the difference between human in the loop and human on the loop?

In human in the loop vs human on the loop, the first requires a person’s approval before the system acts; the second lets the system act within limits while a person monitors and can intervene or halt it.

What does human out of the loop mean?

The system acts with no human involvement once it is started. Oversight moves entirely to design, testing and after-the-fact audit.

What are the levels of automation?

The best-known scale, from Sheridan and Verplank (1978) as restated by Parasuraman, Sheridan and Wickens (2000), has ten levels: from the computer offering no assistance, through executing a suggestion only if the human approves (level 5) or unless the human vetoes in time (level 6), to the computer deciding everything and ignoring the human (level 10). The 2000 paper adds that each function — acquiring information, analysing it, deciding, acting — can sit at its own level.

Is human in the loop required by the EU AI Act?

Article 14 requires effective human oversight for high-risk AI systems: people must be able to understand the system, stay aware of automation bias, override or reverse its output and interrupt it. It does not mandate one position for every decision; how it applies depends on the system.

What is automation bias?

The tendency to over-rely on automated output — accepting it without checking. It is why oversight designs hide scores on audit samples and show reviewers the source evidence.

When should an AI agent run without a human?

When actions are low-impact and easy to undo, and accuracy has been measured on your own data. Even then, sample results and keep a way to stop it.

Sources
  1. US Department of Defense Directive 3000.09, “Autonomy in Weapon Systems” (25 January 2023)esd.whs.mil
  2. Congressional Research Service, “Defense Primer: U.S. Policy on Lethal Autonomous Weapon Systems” (IF11150)congress.gov
  3. EU AI Act, Article 14: Human oversightartificialintelligenceact.eu
  4. Parasuraman, Sheridan & Wickens, “A model for types and levels of human interaction with automation”, IEEE Trans. SMC-A 30(3), 2000doi.org
  5. Bainbridge, “Ironies of Automation”, Automatica 19(6), 1983doi.org
  6. Sheridan & Verplank, “Human and Computer Control of Undersea Teleoperators”, MIT, 1978 (NTRL record)ntrl.ntis.gov
Read next
Straight-through processing (STP)What STP means in banking, insurance and documents — and how to measure itConfidence calibrationReliability diagrams, ECE, and Platt, isotonic and temperature scalingStructured data extraction with LLMsSchemas, structured outputs, validation layers and citationsAll guidesEvery explainer, by topic

See it on
your own work.

Bring one process and its documents. We’ll show you the agents, the controls and the record, running on your data.

Book a demoHuman in the loop