Human in the Loop

Human in the loop describes a design in which a person must act before the system proceeds — approving, correcting or rejecting at a defined point. It is distinct from human on the loop, where a person monitors an autonomous system and can intervene, and from human oversight generally, which may mean nothing more than someone receiving a report.

The distinction regulators care about

In the loop means the action does not happen without the person. On the loop means the action happens and the person can stop or reverse it. The difference is whether the human is a gate or a monitor, and it determines what the organization can claim about control.

It matters because on-the-loop degrades quietly. A reviewer approving one item an hour reads each one. A reviewer approving three hundred approves three hundred. The control looks identical in a policy document and has stopped operating in practice — which is why an auditor asks for approval latency and override rates rather than for the policy.

The EU AI Act requires human oversight for high-risk systems, and the useful question for a deployer is not whether oversight exists but which of these two it is, and whether the evidence shows the human function operating.

Where to put the gate

Gating everything destroys the value of automation and produces rubber-stamping, which is worse than no gate because it manufactures evidence of review that did not occur. Gating nothing leaves consequential actions unsupervised.

The workable pattern is to gate by consequence and reversibility. An agent reading a record needs no gate. An agent writing to one that cannot be reverted, moving money, contacting a customer or granting access does. That decision belongs in the same versioned policy as everything else — see policy as code — rather than in an agent's own prompt, because an instruction to ask permission is a request the agent can be talked out of.

Making it evidentiary

A step-up prompt that nobody records proves nothing afterwards. Four things have to be captured at the moment of the decision: which human, which action, what they were shown, and what they decided. Without the third, an approval cannot be defended — a person cannot be said to have authorized what they were not shown.

That record is what turns human oversight from a stated control into a demonstrable one, and it belongs in the same tamper-evident log as the enforcement decisions around it. For a supervised firm it is also what FINRA Rule 3110 is asking for when it asks who supervised the conduct.

Smartflow enforces step-up as a policy outcome on the request path rather than as an application feature, so it applies to every caller. See runtime governance, agent governance and examination readiness.

Related terms

Agent authority scope · Agentic AI · Policy as code · FINRA Rule 3110 · Full glossary