
Human-in-the-loop: where a person approves AI actions
In short: Human-in-the-loop puts a person at chosen decision points in an AI system, so costly or irreversible actions wait for approval before they run.
Human-in-the-loop (HITL) is a design where an AI system pauses at defined points so a person can review, approve, edit or reject its output before it takes effect. In agent systems it usually sits in front of irreversible or costly actions: payments, sending messages, deleting data or deploying code. Article 14 of the EU AI Act, Regulation 2024/1689, requires human oversight for high-risk systems.
Human-in-the-loop in plain words
The intern drafts the email; you press send. The intern does the work, you own the consequence.
How it works
The harness classifies actions by risk. Low-risk actions run automatically; high-risk ones produce a pending request showing what will happen and why. A person approves, edits or rejects, and the decision is logged. Good designs put the evidence needed to decide next to the yes or no button.
Published guidance points the same way. OWASP's prompt injection entry in its 2025 list recommends human approval for high-risk actions. The NIST AI Risk Management Framework from January 2023 asks organisations to define and document their human oversight processes, and notes that some AI systems may need no oversight at all.
Example
As an illustration, a finance agent reconciles 400 invoices automatically but routes any payment above a set threshold, or any new supplier, to an approver with the matched documents attached. The approval rate and the time-to-approve become metrics.
Where you'll meet it
- Permission prompts in coding agents before shell commands or edits
- Content moderation and claims-processing queues
- Agentic payments and procurement flows
- High-risk AI systems under the EU AI Act
What it changes for you
HITL moves work rather than removing it, so plan reviewer time like any other capacity. For developers, the job is designing the review screen: show the diff, the source and the confidence. For leaders, approval rates are one signal of when an action is ready to automate fully.
Human-in-the-loop vs human-on-the-loop
In the loop, a person must approve before the action happens. On the loop, the system acts and a person monitors and can intervene afterwards.
Common mistake
My view: putting a human on every step. Reviewers who approve hundreds of trivial actions start clicking yes without reading, and the one dangerous action slips through. Article 14 itself warns about over-relying on a system's output, which it calls automation bias. Reserve approvals for what is genuinely costly to undo.
Related terms
Guardrails · Permission modes · AI governance · EU AI Act · AI agent
From the archive: Agent approvals need independent policy checks · Price the time it takes to believe an agent is done
Reserve human approval for actions that are costly to undo, or reviewers will stop reading what they approve.


