Human in the loop

If you want to use AI productively in your company, human in the loop is not a technical footnote but the answer to the question of how much autonomy you grant a machine before someone has looked at the result.

The basic idea is simple: automation is most valuable where it takes over frequent, uniform work, and most dangerous where a single wrong decision is expensive, irreversible or legally sensitive. Human in the loop resolves this conflict of goals by combining both: the machine handles the bulk, while people keep authority over the critical points.

The three levels of control: in, on and out of the loop

Expert discussion distinguishes three levels of human involvement, and the differences are highly relevant in practice:

• Human in the loop: The person is part of the decision path. Nothing takes effect without their active approval. Example: an AI system drafts a price change, an employee checks and confirms it, and only then does it go live.
• Human on the loop: The system acts on its own; the person monitors and can intervene. Example: an automation sends order confirmations independently, and a dashboard shows anomalies where someone can stop it.
• Human out of the loop: The system decides and acts fully autonomously, without human involvement in the individual case. In civilian business use this is the exception and should remain so for consequential decisions.

Which level is appropriate depends not on the technology but on the risk of the individual action. A product description that an agent drafts can tolerate a different depth of control than a pricing rule affecting 4,000 items or a dismissal decision in HR. Mature AI architectures therefore mix the levels: fully automatic where mistakes are cheap, mandatory approval where they become expensive.

Human in the loop in machine learning

In machine learning, HITL also describes a methodology in which people feed the learning loop itself. Three forms are common. In labelling, people annotate training data, for example by categorising product images or rating texts, and thereby lay the foundation for supervised learning. In active learning, the model turns the tables: it identifies the cases it is least certain about and presents exactly these to a person for a decision, so that human effort flows to where it has the greatest learning effect. And in reinforcement learning from human feedback (RLHF), people rate model outputs, and these preferences continue to train the model. RLHF is the mechanism by which modern language models such as GPT or Claude were aligned with human expectations, the most prominent use of HITL of all.

Companies can draw a practical insight from this: every correction an employee makes to an AI result is potentially training material. If you build approval processes so that corrections are captured systematically, you improve not just the individual result but the system. A classic field of application is content moderation on large platforms: algorithms filter the clear-cut cases, human review teams decide the grey areas, and their judgements flow back into the models as training data. This interplay of machine pre-sorting and human final decision is the HITL principle in its purest form.

HITL in agentic systems: from proposal to approval

With the rise of AI agents, systems that do not just generate text but carry out actions in real systems, human in the loop has acquired a new urgency. A chatbot that gives a wrong answer is embarrassing. An agent with write permissions that misunderstands an instruction is an operational risk. The standard safeguarding repertoire consists of four building blocks: approval gates (defined actions require explicit approval), role and permission restrictions (the agent may only do what its role allows), whitelisting (only selected objects or entities can be edited at all) and draft mode (the agent produces proposals that only take effect after review).

Real-world example: Shopware Copilot

What this looks like in a concrete product is shown by Shopware Copilot, which Shopware presented at Community Day 2026. The agent can run analyses in the shop back end, make recommendations and carry out actions, such as touching pricing rules or product data. Shopware has explicitly put the control model at the centre: human in the loop as the core of the safety mechanisms, plus role restrictions, entity whitelisting and a draft and review process before changes take effect. The logic behind it: autonomy where it saves time, control where a wrong decision would mean 4,000 wrongly priced items on Monday morning. The same architecture can be found in essence in other agentic products, from GitHub Copilot and Microsoft 365 Copilot to workflow platforms such as n8n, which offer approval nodes as a separate step.

Escalation by confidence: the most important design pattern

In practice one pattern has become established that dissolves the rigid either-or question: escalation by confidence and risk. The system rates every pending action on two axes, how certain it is and how consequential an error would be, and routes accordingly. High certainty at low risk runs through automatically. Low certainty or high risk lands in a review queue with a person. The borderline cases in between are collected and reviewed in batches, for example once a day. The result is a system that handles most of the work autonomously without a single critical decision slipping through unseen.

The pattern has a second advantage: it makes the degree of automation measurable and controllable. If you know that today 70 per cent of cases run automatically and 30 per cent escalate for review, you can work specifically on lowering the escalation rate, through better data, sharper rules or a retrained model. The rate becomes the metric that shows how mature your automation is, and the discussion with the business department and data protection gains a factual basis instead of a gut feeling.

The legal framework: human oversight as an obligation

Human in the loop has long ceased to be merely best practice; in part it is a legal obligation. Article 14 of the European AI Regulation (EU AI Act, Regulation (EU) 2024/1689) requires effective human oversight for high-risk AI systems: systems must be designed so that people can understand them, interpret their outputs correctly, intervene and stop them if in doubt. The text of the regulation is publicly available via EUR-Lex. The GDPR, too, knows a related principle in Article 22: nobody may be subjected to a decision based solely on automated processing that has legal effects, which in practice often comes down to a human review step.

For you this means: if your use of AI falls into regulated areas, such as applicant selection, credit decisions or certain safety functions, the question is not whether you bring a person into the loop but how you prove that this oversight is effective and does not exist only on paper.

The limits: when the human in the loop only nods things through

As important as the principle is, it has a well-known weak spot, and it sits in front of the screen. Automation bias describes the well-documented tendency of people to trust machine recommendations, especially when the system is right most of the time. Anyone who has approved a hundred correct proposals in a row no longer seriously checks the hundred-and-first. Human in the loop then turns into rubber-stamping: an approval that formally exists but no longer controls anything in substance.

What helps against this are design decisions, not appeals. Approvals should deliver the relevant information for the review directly instead of just showing a confirm button. Spot checks and the occasional deliberate test case keep attention sharp. The number of approvals per person must be set so that real review remains possible, because anyone expected to nod through 500 decisions a day reviews none of them. And responsibility must be clearly assigned: whoever approves answers for the decision, so they also need the time and information to be able to do so.

The second limit is economic. Every human review step costs lead time and staff, and a process in which every trifle has to be approved ends up slower than the manual process before it. The art lies in calibration: automate low-risk actions, bundle high-risk ones and review them specifically, and readjust the threshold regularly once the system has proven itself. Human in the loop is not an end state but a dial that you deliberately turn as trust and the data basis grow.

Frequently asked questions about human in the loop

Isn't human in the loop simply a brake on automation?

In the short term a review step costs time, that is true. But it is often the precondition for automation being approved at all, by IT security, by the data protection officer or by management. A well-calibrated HITL process speeds up the introduction of AI because it makes the risk manageable.

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

In the loop means: without human approval, the action does not take effect. On the loop means: the machine acts by itself, while a person monitors and can intervene. The first variant suits consequential individual decisions, the second high-frequency processes with low individual risk.

Does the law require human in the loop?

For high-risk AI systems, Article 14 of the EU AI Act requires effective human oversight, and Article 22 GDPR limits decisions based solely on automated processing that have legal effects. Outside these areas HITL is not mandatory, but it is usually the condition under which stakeholders agree to an AI deployment.

How do I prevent approvals from degenerating into mere nodding?

Through process design: show decision-relevant information directly in the approval, limit the number of approvals per person, build in spot checks and assign responsibilities clearly. Automation bias is human, so the process has to factor it in rather than ignore it.

Where is the best place to start in e-commerce?

With tasks that are high-frequency and have a clear tolerance for error: product texts and translations in draft mode, data enrichment with spot checks, analyses that only read anyway. Write actions on prices, stock or customer communication initially belong behind an approval gate that you only loosen once the system has proven itself over several weeks.