Human-in-the-loop is an AI operating model where a person reviews and approves an agent's proposed action before it executes — keeping accountability with a named human while the agent does the work.
Definition
In a human-in-the-loop system, the agent handles perception, reasoning, and execution, but a defined checkpoint routes each proposed action to a person for approval, rejection, or edit. It sits between full manual operation (human does everything) and full autonomy (agent acts unsupervised), and it's the model regulators, clients, and CFOs are most comfortable with.
Why it matters
For paid media, the core risk of agentic systems is accountability, not capability — when spend moves, someone has to answer for it. HITL makes that answer legible: a named person approved the change, and there's a record. It also caps blast radius: nothing ships that nobody signed off on.
In agentic execution
Adlyse runs human-in-the-loop on every change. The loop isn't overhead — each approve or reject is labeled feedback that trains the account's agents on how the team wants spend run.
Related terms
See these concepts running on your own accounts.
Start with a free audit and see what an agentic workforce would change first.
Start a free audit