Definition
Human in the loop, often abbreviated HITL, describes any process in which a person stays actively involved in an AI-supported workflow — reviewing, correcting, approving or overriding the AI's output before it has real-world effect. The AI does the heavy lifting of drafting, sorting, predicting or recommending; the human applies judgement, context and accountability, and makes the final call.
The term originally came from engineering and machine learning, where humans labelled data and corrected model predictions to improve training. In the age of generative AI and AI agents, it has taken on a broader workplace meaning: no important decision should be fully handed to a system that can be confidently wrong.
Levels of human oversight
Oversight is not all-or-nothing. A useful way to think about it is a spectrum:
- Human in the loop. A person must review and approve every output before it is used. Example: an AI drafts a performance review; the manager edits and signs it.
- Human on the loop. The AI acts on its own, but a person monitors it and can intervene or stop it. Example: an AI routes support tickets, and a supervisor watches a dashboard for misroutes.
- Human out of the loop. The AI acts with no real-time human involvement. Suitable only for low-risk, well-tested, easily reversible tasks, such as auto-tagging internal documents.
The right level depends on risk: the more an output affects people's rights, money, safety or reputation, the closer the human should be.
Why human in the loop matters
AI models hallucinate, reflect bias in their training data and lack the context a person has about a specific situation. A human reviewer catches factual errors, notices when a recommendation is unfair, considers circumstances the model cannot see and takes responsibility for the outcome. Courts and regulators increasingly expect this. In Moffatt v. Air Canada (2024 BCCRT 149), a Canadian tribunal rejected the airline's claim that it was not liable for what its chatbot told a customer. Organisations cannot outsource accountability to software.
Regulation codifies the principle. Article 14 of the EU AI Act requires high-risk AI systems — which, under Annex III, include systems used for recruitment, promotion, task allocation and performance evaluation — to be designed so that people can effectively oversee them, understand their limits, avoid over-reliance (so-called automation bias) and override or stop them. Article 22 of the GDPR gives individuals rights regarding decisions based solely on automated processing that significantly affect them. The NIST AI Risk Management Framework similarly treats human oversight as a core control.
Example
A recruiting team uses AI to help screen 600 applications for a customer-success role. A poorly designed process would let the AI rank candidates and automatically reject the bottom 80%. That approach risks unlawful bias, misses unconventional but strong candidates and may breach the EU AI Act and GDPR.
The team designs a human-in-the-loop workflow instead:
- The recruiter defines the job's must-have criteria with the hiring manager.
- The AI summarises each application against those criteria and flags evidence for and against — it does not reject anyone.
- A recruiter reviews every summary, opens the full CV for borderline cases and makes the shortlisting decision.
- Each week the team checks a sample of decisions for patterns, such as whether candidates from certain backgrounds are disproportionately screened out.
- Candidates are told that AI assists the process and can request human review.
Screening time drops from about two weeks to four days, every decision has a named human owner, and the company can explain and defend its process. The AI makes the recruiter faster; it does not replace the recruiter's judgement.
Designing effective human oversight
- Make review real. Give reviewers the time, information and authority to disagree with the AI. A rubber stamp is not oversight.
- Fight automation bias. People tend to accept confident AI output. Ask reviewers to check sources and justify approvals on high-stakes items.
- Match oversight to risk. Light checks for brainstorming; full review for hiring, pay, legal, medical or customer-facing decisions.
- Log decisions. Record what the AI suggested and what the human decided, and why.
- Close the loop. Use corrections to improve prompts, data and processes.
Key takeaways
- Human in the loop keeps a person reviewing and approving AI output before it matters.
- Oversight should scale with risk: in, on or out of the loop.
- The EU AI Act and GDPR make meaningful human oversight a legal expectation in many HR and high-impact uses.
- Real oversight requires time, information and the authority to override.