Definition
A design principle where a human validates the critical decisions or actions of an AI system before they are executed.
Human-in-the-loop (HITL) places a human control point in an automated process: the AI prepares, the human decides. Applied to AI agents, this means an email drafted by the agent is only sent after review, and a proposed transfer is only executed after approval.
It is both a growing regulatory requirement (the AI Act mandates human oversight for high-risk systems) and a matter of operational common sense: the AI's autonomy is calibrated according to the reversibility and exposure of each action, not according to the trust placed in the model.
An agent prepares the reply to a customer complaint and suggests a goodwill gesture. Nothing is sent until an advisor has reviewed it, corrected it if needed and clicked "send".
For any action that commits the company or a person: sending something outside, paying, deciding on a case, deleting data. Reading and preparation tasks, on the other hand, can run on their own.
Doesn't human validation cancel out the time saved?
No. Most of the time is saved on research and drafting; reviewing and approving takes a few seconds. It is also what makes the tool acceptable to your teams and to your clients.
All AI models, deployed on your infrastructure, with your documents connected.