Human in the Loop, on the Loop or in Command: Practitioner Q&A on AI Oversight

Human in the loop, on the loop or in command: which oversight model fits which AI decision, what UK and EU rules expect and how to evidence it.

The three models of human oversight differ by when the human acts: before each decision, while the system runs or when deciding whether to use it at all. Choosing between them is a risk decision for each use of AI, and a named reviewer is only a control if you can show the review changes outcomes. This page is general information as at its publication date, not legal advice.

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

The difference is when the person acts. With human in the loop, a person can intervene in every decision cycle. With human on the loop, people intervene during the design cycle and monitor the system in operation, stepping in when something looks wrong.

Both models come from the EU High-Level Expert Group (HLEG) Ethics Guidelines for Trustworthy AI, published in 2019, which list human agency and oversight as the first of seven requirements. The guidelines say intervention in every decision cycle ‘in many cases is neither possible nor desirable’. They also say less oversight calls for more extensive testing and stricter governance, so moving from in the loop to on the loop trades per-decision review for heavier testing and stricter AI governance.

Is human in command the same as human in control?

No. Human in command is a defined oversight model, while human in control is a looser phrase with no formal definition in the HLEG guidelines or the EU AI Act.

Human in command originates in opinion INT/806 of the European Economic and Social Committee, adopted on 31 May 2017, which called for machines to remain machines and for people to retain control over them at all times. The HLEG adopted it as its third model: oversight of a system’s overall activity and wider impact, including the decision on when and how to use it or whether to use it at all.

A finance director may see this as committee overhead. It has control value when it leaves a record: a named owner and a dated decision to keep or withdraw each system. Algorithmic accountability depends on that record.

When should we use human in the loop and when human on the loop?

Use human in the loop where a single wrong decision is serious and hard to reverse, and volumes allow real review. Use human on the loop where decisions are frequent and each one is low-impact and reversible, provided you can detect and correct errors quickly.

Cost is a fair objection. If a reviewer agrees with the model 99% of the time, you are paying for the 1%, so the question is what that 1% costs when it reaches the person affected. Where the answer is ‘very little’, per-decision review is poor value and monitoring with sampling is the better control. Where the answer is ‘a discrimination claim’ or ‘a safety incident’, the review is cheap.

One caveat: the ICO’s draft guidance, covered in the next answer, says human involvement must happen for every decision, so under that draft, sampling or monitoring alone would be unlikely to amount to meaningful human involvement in a significant decision.

Make the choice for each use of AI and write down the reasoning. One system can need both models for different decisions.

Does human review make an automated decision compliant?

No, review alone does not make a decision compliant. UK and EU rules both look at what the reviewer can do in practice. In the EU, Article 26(2) of the AI Act asks deployers of high-risk systems to assign oversight to people with the necessary competence, training and authority; for Annex III systems this applies from 2 December 2027.

In the UK, Articles 22A to 22D of the UK GDPR, inserted by section 80 of the Data (Use and Access) Act 2025 and in force since 5 February 2026, treat a decision as solely automated where there is no meaningful human involvement. The ICO’s draft guidance on automated decision-making, published on 31 March 2026 and not final as of November 2026, says involvement must be active, take place before the decision is applied, consider the relevant factors and happen for every decision. The ICO’s recruitment report of the same date found that most employers described their tools as decision support, yet many decisions were in practice solely automated.

Some organisations will consider dropping review and relying on the Article 22C safeguards. Those safeguards include human intervention, so a competent reviewer is likely to be needed either way. Take legal advice on which route fits your processing.

How does automation bias affect human oversight of AI?

Automation bias can leave you with a reviewer who is present but defers to the system’s output even when it is wrong.

The evidence depends on context. In a 2023 mammography study published in Radiology, very experienced radiologists’ accuracy fell from 82% to 45.5% when the purported AI suggested the wrong category. Public-sector experiments by Alon-Barkat and Busuioc, also from 2023, found no evidence of general automation bias, with mixed results on whether people follow algorithmic advice that matches stereotypes.

Article 14(4)(b) of the EU AI Act names automation bias and asks that people overseeing high-risk systems are enabled to remain aware of it; for Annex III systems this applies from 2 December 2027. The practical response is to keep reviewers aware of the risk and test in your own setting, with your own reviewers, to measure its extent.

What human oversight do AI agents need?

Agents usually need a mix of models: human approval before irreversible actions and monitoring for the rest, under a command-level decision on what the agent may do at all.

AI agency changes what is practical. An agent can take many actions quickly, so reviewing every action is unworkable. A monitor cannot always step in before an irreversible action completes. This is our inference from how agents work; no regulator has said it. It points to placing actions such as payments and deletions behind human approval or technical limits, with human on the loop for reversible steps and human in command setting the agent’s permissions.

Regulatory guidance is still forming. The ICO lists agentic AI guidance as in drafting on its guidance plans page, with a final version due in spring 2027.

How do you evidence effective human oversight of AI?

Show that review changes outcomes, and use more than one measure. Override counts alone prove little: a low rate may mean an accurate model or a passive reviewer, and a high rate may mean good oversight or a poorly calibrated model.

Stronger evidence comes from seeded test cases with known wrong outputs, time spent per review set against case complexity and samples of reviewer decisions checked against the right answer. Add records of how and when human involvement took place, which the ICO’s draft guidance asks for.

Meaningful human involvement has no statutory definition yet: as of late October 2026 no regulations had been made under Article 22D, and the ICO guidance is a draft. Sceptics call any evidence standard guesswork until that changes. Until the regulations or final guidance arrive, the ICO’s draft is the best available indication of what regulators expect, and evidence that a competent person with authority changed outcomes is consistent with it. Final wording may differ.

Who is responsible for human oversight of an AI system?

Where the EU AI Act’s high-risk rules apply, oversight in use sits with the deploying organisation, even when a vendor built the system. Article 14 asks providers of high-risk systems to design them so that people can oversee them effectively. Article 26(2) asks deployers to assign oversight to natural persons with the necessary competence, training, authority and support. Whether those articles apply to you depends on your systems and your EU exposure; for Annex III systems they apply from 2 December 2027, the date set by Regulation (EU) 2026/1744.

Name the individuals and confirm they have the authority and the time to override or stop the system. For the wider accountability question, see who is responsible for AI.

What should a human oversight process for AI include?

At minimum, the process should record which oversight model applies to each use of AI and who performs the oversight, with evidence that it works.

We would expect to see:

  • An inventory of AI uses, each with a chosen model and the risk reasoning behind it
  • Named reviewers with the competence, training and authority to override or stop the system
  • A review point before the decision is applied to the person
  • Human approval or technical limits on irreversible actions
  • Regular tests for automation bias using seeded cases
  • Records of each intervention and retained logs (Article 26(6) of the EU AI Act asks deployers of high-risk systems to keep logs for at least six months, subject to other law)
  • A command-level review of whether each system should stay in use

Why start before the EU dates arrive and the ICO guidance is final? The UK provisions are already in force for organisations making significant automated decisions about individuals using personal data, and oversight evidence can take months to accumulate.

AI Behaviour Verification

AI Behaviour Verification is our ongoing independent verification that AI systems behave as intended, with regular reports that serve as evidence for auditors.