AI concepts
Human-in-the-loop (a person who approves AI decisions)
An architectural pattern in which an AI system's decision requires human approval before its consequences take effect. It is not a safety net, it is part of the architecture, and without it a solution is not production-ready in high-stakes applications. The share of deployments using HITL in the financial and medical sectors exceeds 95 percent.
Primary source: Stanford HAI 2024, FDA Software as Medical Device Guidance 2025
The term comes from 1970s research on aviation control systems, where full automation was technically possible but operationally risky. In the AI era the pattern returned with new force as a response to LLM hallucinations, model bias and the unpredictability of agentic AI.
Three levels of intervention
Human-on-the-loop. A person monitors, the system acts. Intervention happens only when an alert fires. The fastest and cheapest option. Used in marketing, content moderation and predictive maintenance.
Human-in-the-loop. Every decision requires active human approval. The AI proposes, the person accepts or rejects. Used in medicine, law, financial decisions above a threshold and HR (hiring, promotion, dismissal).
Human-on-the-system. A person does not see individual decisions but controls the policy, the parameters and the thresholds. Used in very high-volume systems where HITL does not scale.
When HITL is required by law
The AI Act of February 2026, Annex III, introduces an obligation of human oversight for high-risk systems. For regulated industries in Poland: banking (Recommendation S of the Polish Financial Supervision Authority), medicine (the Medical Devices Act and the MDR regulation), law (the Code of Ethics for Legal Advisers). Breaching the oversight obligation triggers penalties that exceed the cost of implementing HITL by tens of times.
Effectiveness
The FDA Software as Medical Device Guidance of 2025 reports that systems with documented HITL have 47 percent fewer incidents than fully autonomous systems on tasks of comparable complexity. Stanford HAI 2024 adds that in creative tasks (writing, design) HITL raises output quality by 31 to 58 percent in blind tests.
The cost of HITL
There are three components. First, the UI design for the human reviewer (an interface that makes it easy to approve or reject). Second, reviewer training (so they understand what they are assessing, when to agree and when not to). Third, monitoring of reviewer quality (whether they systematically reject certain types of cases or rubber-stamp everything).
A typical HITL setup raises AI infrastructure CapEx by 15 to 30 percent and adds the operating cost of reviewer work. That cost is often understated in AI vendor proposals because it lowers the proposal's ROI.
Anti-pattern: the rubber stamp
A common failure mode in organisations that implement HITL without designing reviewer incentives. The reviewer approves everything, because approval creates no work while rejection does. After six months HITL becomes a formality, decisions are effectively autonomous, yet the company keeps paying the operating cost. The fix: random audits, where a supervisor verifies 1 to 5 percent of approved decisions.
We decide where in the process HITL makes sense for your company as part of the AI Readiness Audit.