What it means
- For each AI system, decide in writing what the AI may do on its own and where a person must check, approve or be able to step in.
- Named, trained people can pause, override or switch off the system quickly, including a "kill switch" for serious harm.
- Oversight is real: reviewers have time, information and authority to disagree with the AI.
- ISO 27001 reuse: None. New for AI, not covered by ISO 27001. ISO 27001 documented operating procedures (A.5.37) can hold the SOPs, but the oversight design itself is new.
Why it matters
- RBI FREE-AI asks for safeguards and SOPs before autonomous AI, human oversight for medium and high risk tasks, and the ability to shut an AI system down instantly.
- The 2026 IT Rules require human oversight and periodic review of automated content detection tools at SSMIs.
- MeitY advisories expect human oversight as part of safeguards against unlawful AI output.
- Agentic AI that can send emails, move money or change records raises the stakes sharply.
Checklist
- For each inventory entry, record the oversight mode: human in the loop (approves each action), human on the loop (monitors and can intervene), or fully automated with sampling (SC-AI-004 AI risk classification and use-case tiering).
- Require human in the loop for high-risk decisions about people and for irreversible agent actions (payments, deletions, external communications).
- Write SOPs: what the reviewer checks, when to override, how to escalate, how to record the reason.
- Build and test a kill switch or feature flag that disables each AI system within minutes; name who may use it (SC-AI-033 Business continuity and fallback for AI).
- Set limits on agents: spend caps, rate limits, allowed tools, approval steps (SC-AI-030 Access control for AI systems and agents).
- Train reviewers on the model's known weaknesses and on automation bias (trusting the AI too much).
- Track override rates; very low rates may mean rubber-stamping, very high rates may mean the model is unfit.
- Review automated moderation or detection tools periodically for accuracy, fairness and bias.
Evidence examples
- Oversight matrix listing each AI system, its risk tier and oversight mode.
- Reviewer SOPs and training attendance records.
- Kill switch test record (date, who triggered it, time to disable).
- Override logs with reasons, and a monthly override rate report.
- Agent configuration showing approval steps and spend or action limits.
Common pitfalls
- "Human in the loop" that clicks approve on hundreds of cases an hour without reading.
- Kill switch that exists only in a document and has never been tested.
- Reviewers with no authority to overrule the model or no access to the inputs.
- Agents given broad tool permissions "for convenience".
Satisfies
- VoluntaryISO/IEC 42001
- BindingEU AI Act
- SupervisoryPRA SS1/23
- VoluntaryNIST AI RMF and GenAI Profile
- BindingNYC Local Law 144 (AEDT)
- SupervisoryNYDFS AI Cybersecurity Letter 2024
- BindingKorea AI Basic Act
- SupervisoryOSFI E-23
- SupervisoryMeitY AI and Deepfake Advisories
- VoluntaryRBI FREE-AI
Also relevant to
- ICMR AI in Healthcare Guidelinesguidance mapping (human in the loop; no obligation IDs)
Standards crosswalk
Clause references only. Check the standard itself for the exact wording.
ISO/IEC 42001
- A.6.2.6 operation and monitoring
- A.9.2 processes for responsible use
- A.9.4 intended use
ISO/IEC 27001
No direct equivalent
SOC 2
No direct equivalent
Key takeaways
- Decide and record, per system, where a human must approve and where they only monitor.
- A tested kill switch is part of oversight, not an optional extra.
- Measure override rates to prove oversight is real.