Human-in-the-loop

1. Introduction: when AI does the work, who is responsible?

AI can already draft replies to customers, screen job applications, summarize patient records, assess credit risk and draft legal texts. That's impressive. And that's exactly why it's also dangerous.

What happens if a customer service bot gives the wrong instructions? If a recruiting system discriminates against applicants? If, in healthcare, an AI suggestion sounds plausible but turns out to be wrong?

This is where human-in-the-loop comes in: an operating model in which a person reviews, approves or steers what the AI does.

AI can produce the first draft. A person makes the final decision. And above all, a person carries the responsibility.

EU AI regulation puts particular emphasis on human oversight of high-risk AI systems, with the aim of preventing or reducing risks to health, safety and fundamental rights.^[1]

2. What does human-in-the-loop mean?

Human-in-the-loop, often abbreviated as HITL, refers to a model in which AI does not operate fully on its own; instead, a person is involved at a critical point in the process.

In practice, this can mean, for example, that:

the AI drafts a proposed response
an expert reviews it
a person approves, edits or rejects the proposal
only then does the content go out to the customer, patient or job applicant, or into a decision

A simple example:

AI drafts a customer service reply.
A customer service agent reviews it.
Only the reviewed message is sent to the customer.

It sounds like a small thing, but it fundamentally changes the role of AI. AI is not the decision-maker. It is a preparer, a sparring partner or an accelerator.

3. Why do we need people alongside AI?

AI is strong at certain things. It can read huge volumes of text without tiring, find patterns, produce drafts quickly and suggest alternatives.

But it doesn't actually carry responsibility. It doesn't understand a person's situation the way an expert does. It doesn't grasp the organization's values, the customer's emotional state, legal liability or the consequences of a decision the way a human does.

That's why human-in-the-loop is, above all, a model of accountability. It answers three questions:

Where is AI allowed to help?
Where must it stop?
Which person is responsible for the outcome?

NIST's AI Risk Management Framework approaches AI precisely through risk management: the goal is to help organizations identify and manage the risks AI poses to individuals, organizations and society.^[2]

4. The basic human-in-the-loop model

A good HITL model doesn't mean a person glancing at the AI's output in a hurry and hitting the approve button. It means a clear division of labor.

1. AI does the preparation

For example, AI can summarize a document, suggest a reply, categorize feedback, spot anomalies, produce a draft or carry out a preliminary risk analysis.

2. A person evaluates

A person checks: Is the information correct? Is the tone appropriate? Is anything essential missing? Is the decision fair? Does the outcome comply with the law, our values and our guidelines?

3. A person decides

This is the most important part: final responsibility does not rest with the system. A person approves, edits or rejects the proposal.

4. Learn from the process

In a good model, mistakes don't remain one-off slip-ups. Lessons are drawn from them: where AI often gets things wrong, where it is useful, where a person needs to step in earlier, and which instructions or data need improving.

5. Where is human review essential?

Human-in-the-loop is especially important when an AI error can affect a person's rights, health, finances or everyday life. Such areas include customer service, recruiting, healthcare, finance and credit decisions, law, insurance, education, public services, HR and safety-related processes.

In these situations, an AI error isn't just a “bad answer.” It can mean that a person doesn't get a service, a job, care, a loan, compensation or fair treatment.

The OECD's AI Principles emphasize trustworthy AI that respects human rights and democratic values. The principles were adopted in 2019 and updated in 2024.^[3]

6. Examples: what does HITL look like in practice?

Customer service

Poor model: AI replies to customers automatically in every situation.

Better model: AI answers routine questions but routes unclear, emotionally charged or money-related cases to a person.

Excellent model: AI drafts a proposed reply and shows the source or reasoning, and a customer service agent approves the reply before it is sent whenever the matter involves contracts, complaints or the customer's rights.

Here AI speeds up the work but doesn't take judgment away from people.

Recruiting

Poor model: AI scores applicants and the system automatically rejects some of them.

Better model: AI helps structure applications, but a recruiter reviews every decision that leads to a rejection.

Excellent model: AI may summarize skills but may not use sensitive data or make the final decision. The recruiter reviews both the recommendations and any potential biases.

In recruiting, human-in-the-loop isn't just quality assurance. It's a safeguard for fairness.

Healthcare

Poor model: AI gives a patient treatment instructions without review by a professional.

Better model: AI helps a professional summarize information or suggest follow-up questions.

Excellent model: AI serves as clinical decision support, but a doctor or nurse evaluates the suggestion and the patient's overall situation, and is accountable for the decision.

In healthcare, AI can be a useful assistant. But it must not become an authority that no one dares to question.

Finance

Poor model: AI rejects a loan application without an understandable explanation.

Better model: AI assesses risk, but an expert reviews unusual or negative decisions.

Excellent model: AI produces an analysis, explains the key variables and highlights uncertainties. A person makes the final decision and makes sure the customer receives an understandable explanation.

In finance, human-in-the-loop is about both accountability and transparency.

7. A good HITL process in practice

Human-in-the-loop should be designed before AI is deployed — not by automating first and only then figuring out who is responsible when something goes wrong.

1. Define the AI's role

Is the AI a drafter, a classifier, a recommender, an early-warning system, decision support or an autonomous actor? The more autonomous the role, the stronger the oversight it needs.

2. Define the human's role

Who reviews? When? What do they need to know? Do they have the authority to stop the process?

Under the EU AI Act's obligations for high-risk systems, the person providing human oversight must have the necessary competence, training, authority and support.^[4]

This is a key point. Oversight isn't real if the person lacks the time, the expertise or the right to say no.

3. Define the stopping points

For example: AI may draft but not send. AI may recommend but not decide. AI may score but not reject. AI may identify a risk, but a person decides what happens next.

4. Define risk levels

Not every use of AI needs the same level of oversight. An email draft for internal use is a different matter from patient instructions or a credit decision.

Low risk: a person reviews occasionally
Medium risk: a person reviews before external use
High risk: a person always approves before a decision
Very high risk: AI may only support the expert, not make the decision

8. Risks: human-in-the-loop can also be a mere formality

This is the hardest part of the whole model. Human-in-the-loop sounds safe. But it can easily turn into theater.

The rubber-stamp problem

If a person always approves the AI's suggestions, they aren't really overseeing anything. They're just clicking. This is especially dangerous in busy jobs where the AI looks confident and the person grows tired of checking.

Automation bias

People tend to trust a machine's suggestion too much. When a system displays a number, a score or a recommendation, it can feel objective — even if it was built on incomplete data or wrong assumptions.

Blurred accountability

When something goes wrong, an organization may start passing the buck: “The AI suggested it.” “The expert approved it.” “That was the process.” “The vendor is responsible for the model.” A good HITL model prevents this. In such a model, responsibility is assigned to a named person.

Human intervention that comes too late

If a person only gets involved at the very end, the damage may already be done. In recruiting, for example, AI may narrow down the applicant pool before any human sees it, in a way no one notices later.

9. Human-in-the-loop in companies and startups

In the startup world, the temptation is obvious: automate fast, save time and scale up. That can be smart.

But for a growth company, an AI error can be costly. A single wrong customer message, a discriminatory hiring process or an ambiguous contract term can erode trust quickly.

The right question isn't “How do we take people out of the process?” It's: “At which point do people add the most value?”

Often the answer is this:

AI handles the repetitive work
people handle the judgment
AI suggests
people decide
AI speeds things up
people ensure quality

That's a healthy division of labor. AI doesn't have to replace the expert to be useful. Often the greatest benefit comes from the expert getting better drafts, faster summaries and more time for the difficult decisions.

10. Regulation and responsible AI

Human-in-the-loop isn't just good practice. Increasingly, it's also part of regulation, risk management and the principles of responsible AI.

The EU AI Act classifies AI systems by risk level, and high-risk systems are subject to stricter obligations. According to the European Commission, AI systems that pose a clear threat to people's safety, livelihoods or rights are prohibited.^[5]

This shows the direction of travel: AI may be used, but not irresponsibly. Human-in-the-loop is a practical way of answering this question: How do we make sure AI serves people rather than overriding them?

On its own, it isn't enough. You also need information security, documentation, testing, model evaluation, high-quality data and clear responsibilities. But without real human authority, AI oversight easily becomes empty talk.

11. The future: will the human role change?

AI is evolving fast. Agents can already use tools, retrieve information, carry out tasks and act partly on their own. That's why human-in-the-loop is becoming more important than ever.

In the future, we may not review every sentence an AI writes. Instead, the human role will increasingly shift in three directions.

1. People design the boundaries

What is AI allowed to do? What is it not allowed to do? When must it stop?

2. People monitor exceptions

Not everything is checked by hand, but risky cases are escalated to a person. For example: an uncertain answer, an unusual customer case, a high-risk decision, conflicting data or a negative impact on a person.

3. People are accountable for values

AI can optimize for speed, cost or probabilities. But people decide what is fair, reasonable and responsible. That's a big difference.

12. Conclusions: what should we learn from this?

Human-in-the-loop isn't a brake. It's a structure for trust. It helps companies use AI boldly but sensibly.

These are the key takeaways:

AI can prepare, but people are accountable. This is the core of the whole model.
People need to be involved at the right point. If the review happens too late, it no longer provides real protection.
Oversight requires expertise and authority. A person can't be held responsible if they lack the time, the information or the right to stop the AI's suggestion.
The level of risk should determine the level of review. Not everything needs to be reviewed the same way, but in high-risk decisions the human role is essential.
A good HITL model is a process, not a button. An “Approve” button alone doesn't make AI responsible.

Ultimately, human-in-the-loop is a reminder of one important thing: the value of AI doesn't come from removing people from the picture entirely. It comes from people and AI doing the right things at the right points.

13. Practical templates and checklists

13.1 General HITL template

AI's task: [What is the AI allowed to do?]
Human's task: [What does the person review or decide?]
Stopping point: [At what point may the AI not continue without a person?]
Risk level: [Low / medium / high / very high]
Approver: [Who is responsible for the outcome?]
Documentation: [What gets recorded?]
Exceptions: [When is the matter escalated to an expert or management?]

13.2 Customer service template

AI drafts a proposed reply based on the customer's message. A person checks: the facts, the tone, any promises, points related to contracts or money, and situations where the customer is frustrated or in a vulnerable position.

AI must not send a reply automatically if the matter involves: a complaint, a refund, contract terms, personal data or legal liability.

13.3 Recruiting template

AI may: summarize applications, group skills, suggest interview questions.

AI may not: automatically reject an applicant, use sensitive data, make the final selection, or assign scores without human review.

A person reviews: the applicant's actual skills, potential biases, the reasoning behind the decision, equal treatment.

13.4 High-risk checklist

Before approving an AI suggestion, ask:

Does this decision affect a person's rights, money, health or everyday life?
Is the information the AI used up to date?
Is it clear what the suggestion is based on?
Could the suggestion contain a discriminatory or biased assumption?
Does a person have the right to change or reject the suggestion?
Who carries the responsibility?
How is the decision documented?
How can the affected person request a correction if needed?

13.5 A good basic prompt for a human-in-the-loop process

Act as an expert's assistant, not as the final decision-maker.

Your task is to draft a preliminary proposal on [topic]. State clearly:

1) what you propose
2) what the proposal is based on
3) which points require human review
4) what uncertainties or risks the proposal involves

Do not make the final decision. At the end, add a section: “Items for human review.”

13.6 A one-sentence rule of thumb

Let AI speed up the work, but don't let it make a decision on its own that could affect a person's rights, health, money or future.

Human-in-the-loop is the seat belt of the AI era. It doesn't stop you from moving — it makes speed safer.