Agentic AI Glossary

Core terms

Human in Control

An operating model for AI agents in which autonomous agents execute the work while a human retains final decision rights - the authority to approve, redirect, or stop any agent action at any time.

Human-in-the-Loop (HITL)

An oversight pattern where a person sits inside the decision cycle and must approve each AI output before it takes effect. The system cannot act without explicit human sign-off.

Human-on-the-Loop (HOTL)

An oversight pattern where the system acts autonomously while a person monitors from outside the cycle, with the ability to intervene or override when needed.

Human-out-of-the-Loop (HOOTL)

Full autonomy: the system perceives, decides, and acts with no real-time human involvement or supervision.

Decision Rights

The formal authority over what gets approved, changed, or cancelled. In a human-in-control model, decision rights stay with the person even when execution is delegated to agents.

Intervention Rights

The guaranteed ability of a human to step into an agent's workflow at any point - to redirect, pause, or stop - with mechanisms that work in real time, not just in policy documents.

Agent Autonomy

The degree to which a system can decide and act without human input. Higher autonomy means faster execution and less direct human involvement per action. In a human-in-control model, autonomy is scoped by risk tier.

Accountability Gap

The space between regulatory requirements for human oversight and the engineering reality of how AI systems are actually deployed. Closed by making oversight testable and auditable.

Approval Surface

The set of agent actions that require explicit human approval before execution. A well-designed approval surface matches the organisation's risk tolerance.

Autonomy Scope

The boundary within which an agent may act without human involvement. Defined by guardrails, budgets, forbidden actions, and escalation thresholds.

Revocability

The property that a human can withdraw an agent's autonomy at any time - pause its operations, change its mandate, or shut it down entirely. A core guarantee of human in control.

Agent Network

A collection of AI agents that coordinate to achieve goals, often across multiple systems or domains. Oversight of a network requires layered intervention architecture, not single-point monitoring.

Guardrails

Predefined limits on what an agent may do - budgets, scopes, forbidden actions - that enforce the human's intent even between check-ins.

Escalation

A built-in route by which an agent hands a decision back to a human when it hits a threshold: ambiguity, risk, cost, or anything outside its mandate.

Traceability / Audit Trail

Logged records of what an agent did, with what inputs, and why - the evidence base that makes accountability enforceable after the fact.

Automation Bias

The human tendency to over-trust automated outputs and under-scrutinize them. The reason a human 'in the loop' can quietly become a rubber stamp.

Rubber-Stamping

Formal approval without genuine review. It satisfies the letter of human-in-the-loop while providing none of its protection - the most common failure of HITL in practice.

Governance Debt

Control that exists on paper but not in practice - a veto no one can actually exercise in time, an owner who lacks visibility, a kill switch never tested.

Contestability

The right of a person significantly affected by an AI decision to challenge that decision through a timely, accessible process. Distinct from operator override: contestability belongs to the impacted party.

Meaningful Human Control

The standard, drawn from autonomous-weapons debates and echoed in the EU AI Act, that human oversight must be real: the overseer needs understanding, time, and exercisable authority - not just presence in the workflow.