Human in the loop means a person reviews, approves or can override what an AI system does before the result takes effect. Require it when actions are hard to undo, affect people's rights or money, or run with broad permissions. It only works if the reviewer has context, time and real authority to say no.
Why not let the model act alone?
Models make confident mistakes. They can also be manipulated by text hidden in a document or web page. OWASP's list of risks for LLM applications describes excessive agency as the vulnerability that lets damaging actions happen in response to unexpected, ambiguous or manipulated model output, whatever the cause. It names triggers such as hallucination, prompt injection from a malicious user or an earlier compromised extension, and, in multi-agent systems, a compromised peer agent.
OWASP traces the root cause to one or more of three things: excessive functionality, excessive permissions and excessive autonomy. A person in the loop addresses the third. Reducing tools and permissions addresses the first two, and should come first.
What does the EU AI Act require?
Article 14 of the EU AI Act covers human oversight of high-risk AI systems. It says oversight aims to prevent or minimize risks to health, safety or fundamental rights. The text describes what people assigned to oversight must be able to do, including:
- Understand the system's capabilities and limits.
- Stay aware of the tendency to over-rely on its output, called automation bias.
- Correctly interpret its output.
- Decide not to use the system, or disregard, override or reverse its output.
- Intervene or stop it through a stop button or similar procedure that lets it come to a halt in a safe state.
Note the verbs. The law is not satisfied by a checkbox. The person must be able to understand, doubt and stop. Whether your system counts as high risk is a legal question for you and your counsel. The practical lesson applies far beyond it.
When does a human review actually work?
Review fails in a familiar way. A queue shows 200 approvals a day. The reviewer sees a title and a green Approve button. After a week they approve everything. The loop exists on paper only.
A constructed example: an AI assistant drafts refund approvals for a support team. At first, a person reviews each. Volume grows, and reviews take four seconds. A bad prompt injection in a customer message makes the assistant approve an inflated refund, and nobody notices. A better design: auto-approve small refunds within a limit, require a person for larger ones, show the reviewer the evidence the model used, and sample the auto-approved ones for audit.
Good review design has four traits: the reviewer sees the inputs and the reasoning evidence, not only the conclusion; the volume is low enough to think; the default on timeout is no action; and the reviewer is allowed, and expected, to reject.
Which actions need a person?
Use a simple test. Ask whether the action is reversible, how much it costs if wrong, who it affects and what permissions it uses. Deleting data, moving money, changing access, sending messages to customers and shipping code to production all deserve approval or a tight limit. Reading and summarizing usually do not.
Where Safeguard fits: its autonomous remediation can propose fixes as pull requests, with that capability off by default, so a person decides whether a change is opened or merged. Policies and guardrails let you define which agent actions are allowed at all.
How to check or apply this
- List every action your AI system can take. Mark each as read-only or side effect.
- Remove tools and permissions the task does not need before adding review steps.
- Set approval rules by risk: automatic below a limit, a person above it, never for irreversible actions.
- Show reviewers the inputs, sources and diff, not only the conclusion.
- Cap the daily volume a reviewer sees, and sample automatic decisions for audit.
- Make no action the default when nobody responds.
- Add a stop control and test it.
Quick check
Is human in the loop required by law for all AI?
No. The EU AI Act's Article 14 applies to high-risk AI systems. Other laws and sector rules may add duties. Check which apply to your use.
What is automation bias?
It is the tendency to trust an automated output too much. The EU AI Act asks that overseers stay aware of it, especially when a system recommends decisions to people.
Sources
Last reviewed: 18 October 2026.
