Human-in-the-loop is a scheme in which a system built on artificial intelligence (AI) does not perform an action until a human approves it. For agents, it is the most common control method: in 37 of 70 public descriptions of agent deployments in Russia in 2025–2026 (52.9%), the main containment measure is human approval. Under a stream of actions, however, approval quickly turns into a rubber stamp, and the person clicks “Approve” without reading; this is called approval fatigue.
A person in the loop approves an AI action before it runs, and this works while there are few actions. When an agent proposes hundreds of actions an hour, the attention of the person runs out before the stream does. Approval without checking is worse than no approval, because it creates the appearance of control and shifts responsibility onto whoever clicked the button. The way out is for the human to approve rules and for a program to check every action against them, with approval kept for exceptions and high-risk actions.
Three levels of human involvement
It is a control method in which a decision or action of a system takes effect only after a human approves it, and three levels of such involvement are distinguished.
| Level | What the human does | Example |
|---|---|---|
| Human-in-the-loop | Approves every action before it runs | The agent prepares a payment, an accountant clicks “Post” |
| Human-on-the-loop | Watches and can intervene, but does not approve each action | An operator monitors a dashboard and stops the agent when something fails |
| Human-out-of-the-loop | Takes no part in individual actions | The agent acts on its own within set limits |
The AI Act of the European Union requires human oversight for high-risk systems (Article 14) but does not require every action to be approved. Oversight measures must be commensurate with the risks, the level of autonomy and the context of use. The people assigned to oversight must understand the capabilities and limitations of the system, remain aware of the tendency to rely automatically on its output (automation bias), and be able to intervene or stop the system.
Why approving every action does not scale
There are three reasons, and all of them can be seen in any service where a person approves a stream of similar requests.
- The stream: an agent acts faster than a person can read, and the more approvals per hour, the less time remains for each one.
- Almost everything is normal: if 99 requests out of 100 are harmless, a person learns to click “Approve” automatically, and the hundredth, dangerous one goes through with the same gesture. The tendency to rely automatically on the output of a system is called automation bias, and the AI Act explicitly requires the overseeing person to remain aware of it (Article 14(4)(b)).
- The person sees the wrong thing: what is shown is a short description of the action written by the agent itself, not the call parameters. If the agent has been misled by prompt injection, the description will look reasonable.
Something similar has long been known in security operations centers: when there are too many alerts, a significant share of them go unexamined. The essay “Approval Fatigue” by Tian Pan (June 2026) carries this experience over to approvals of agent actions and describes how an approval gate degrades into a formality over time.
What the human signs off on, if not every action
The human signs off on rules. Instead of a hundred identical approvals an hour, the person approves once what the agent may do and within what limits.
- Which tools are available, and for which task they were issued.
- Within what limits the agent acts: amounts, recipients, paths and export volume.
- What is always forbidden, for example deleting backups or paying to new bank details without verification.
- Who is responsible for the rule, how it changes and who approves a change.
From then on, a program checks every agent action against the rule before execution, deterministically, with the decision written to a log, and the human reviews refusals and incidents and changes the rules. This is the architecture of information security of the third generation (IS 3.0); how checking an action differs from filtering text is explained on the page about guardrails.
Where human approval is still needed
Approval does not disappear but becomes rare and therefore meaningful, and it remains in four cases.
- Irreversible high-risk actions, such as a large payment, the deletion of production data or a mass mailing to customers.
- Exceptions, that is, an action the rule does not cover or explicitly forbids but which the business needs.
- Changing the rules themselves, and moving a class of actions from warning mode to enforce mode.
- The first weeks of work of the agent, until statistics have built up for each class of actions.
A good approval request shows the person not the agent’s summary but the call parameters and the reason why the rule did not let the action through automatically.
What to do
- Count the stream: how many approval requests each employee receives per day and how many of them were rejected; if almost none were, approval has become a rubber stamp.
- Split actions by risk: routine actions go under a rule, while irreversible and large ones stay with human approval.
- Write down the rules for routine actions and have management approve them.
- Show the approver the call parameters, not the description written by the agent.
- Keep a log of both approvals and automatic decisions, which will be needed in an incident review and in a dispute over who is liable for an AI mistake.
Frequently asked questions
What is human-in-the-loop in simple terms?
It is a scheme in which a system does not perform an action until a human approves it; for example, an agent prepares a payment and an accountant posts it.
What is approval fatigue?
It is a state in which a person who receives many approval requests stops checking them and approves automatically. Approval formally remains, but control disappears in substance, while responsibility falls on whoever clicked the button.
Is human-in-the-loop enough for agent security?
It is enough only when there are few actions. Under a stream, people stop reading carefully, and what they see is often the summary of the agent rather than the parameters of the action, so it is more reliable for the human to approve rules and for every action to be checked against them automatically.
How does human-in-the-loop differ from human-on-the-loop?
In human-in-the-loop, a person approves every action before it runs. In human-on-the-loop, the person watches the system work and can intervene, but does not approve each action.
When is human approval still mandatory?
For irreversible high-risk actions, for exceptions to the rules, for changing the rules themselves and during the first weeks of work of the agent, until statistics on its actions have built up.
Sources
- Regulation (EU) 2024/1689 (AI Act), Article 14 “Human oversight”, official text: https://eur-lex.europa.eu/eli/reg/2024/1689/oj
- Tian Pan, “Approval Fatigue: How Human-in-the-Loop Gates Decay Into Rubber Stamps”, 25 June 2026 (author’s essay): https://tianpan.co/blog/2026-06-25-approval-fatigue-how-human-in-the-loop-gates-decay-into-rubber-stamps
- Review of 70 public descriptions of agent deployments in Russia in 2025–2026, carried out by the institute in October 2026.
Checked: 16 October 2026; the review of deployments was carried out in October 2026.