A firm that lets an AI agent act inside its systems should be able to answer two questions on any day of the week: who can stop it, and who decides what it does next? Two UK publications now put those questions in front of financial firms and the teams that run autonomous systems.
UK financial regulators call human oversight essential, in practices that are not formal expectations
The Prudential Regulation Authority (PRA), the Financial Conduct Authority (FCA), and the Bank of England jointly issued Effective practices: Cyber risk and resilience management relating to artificial intelligence (AI). The page shows a publication date of 19 November 2026.
The status of the document matters. The effective practices are complementary to, but separate from, the regulators’ supervisory statement and final guidance on cyber risk and resilience management, and the page says they are “not part of the regulators’ formal expectations.” They describe practices firms may adopt. Firms can adopt the practices in a manner appropriate to their size, complexity and risk profile.
On human oversight, the introduction is short and direct. It says it is important not to over-rely on AI, and it states that “Appropriate human oversight remains essential, especially for key decisions.” It adds that the practices are “particularly relevant for AI systems whose autonomy and capabilities create high cybersecurity risks.” Firms may weigh how autonomous a system is, how much access it has to the firm’s assets, and whether it can interact with the external environment.
The same page cites the UK AI Security Institute, which noted that “some capabilities required for AI models to evade human control are emerging.”
The NCSC asks for named individuals where agent errors would matter
The UK National Cyber Security Centre (NCSC) blog Managing the cyber risk of agentic AI is interim advice. The NCSC says it is working with partners on formal guidance that will “build upon, and ultimately supersede, this blog.”
The blog names three oversight approaches. Human-in-the-loop means humans approve actions before they happen. Human-on-the-loop means humans monitor actions and can intervene if needed. Human-out-of-the-loop means the AI acts autonomously without human review. Where unintended activity would have significant consequences, the NCSC recommends making named individuals or a group responsible for agentic AI activities and maintaining human oversight of the agent’s actions.
It also tells operators to decide when an agent should stop and seek human approval, and to ensure those approvals are both guaranteed and gated. It says additional judge AI models and agents can help manage an agent’s activities, and that those judges should be independently evaluated for performance. For higher risk scenarios it recommends human oversight alongside technically enforced controls.
A kill switch is Responsible AI, and the person who can say no is AI Governance
A kill switch, a monitor, and a judge model are useful. Under this column’s definitions they are Responsible AI: machine checking machine. They belong in the stack, and they suit reversible work.
AI Governance asks for a named person with the authority, time, and information to accept, modify, or reject what the agent proposes, and a record of that decision. Checkpoint-Based Governance (CBG) is the method for placing that person at defined checkpoints, and the CBG agents page applies it to agents.
The NCSC’s three terms map onto that split. Human-in-the-loop approval before an action is a checkpoint when a named person holds the decision. Human-on-the-loop monitoring is a checkpoint only if the monitor can act on what it sees. A person who only watches is machine on machine with a person standing nearby. This is the column’s reading, and the pages read for this column do not use the terms Responsible AI or AI Governance.
What a firm should be able to show for each agent
The items below are this column’s practitioner suggestions. The PRA, FCA, Bank of England, and NCSC pages do not set them as requirements.
- Who signs. Name one owner for each agent and one owner for each approval trigger, in writing, before deployment. The NCSC recommendation on named individuals points the same way.
- What record. Keep an accept, modify, or reject log for every triggered approval. Record the reason for each modification or rejection, the time the request arrived, and the time the person decided.
- Which measures. Track three numbers and one test. The override rate is the share of triggered actions modified or rejected; its complement is the share approved without change. Time-to-decision shows how long each triggered approval waited for a person. The third number is the share of approvals decided faster than the owner could have read the request, which works as an automation bias signal. The test is the date and result of the last check that the agent stops or returns to a safe state.
Review the numbers each month. An override rate near zero, paired with approvals that take seconds, is a reason to investigate, because it can mean the checkpoint has become a formality, or it can mean the agent’s proposals were simply sound. Pull a sample of approved actions and have a second person decide them cold. If the second person would have modified or rejected some of them, the firm is running Responsible AI under an AI Governance label. The fix is a named owner who has the time and information to say no, and a log that shows when they did.
Sources
- Prudential Regulation Authority, Financial Conduct Authority, & Bank of England. (2026, November 19). Effective practices: Cyber risk and resilience management relating to artificial intelligence (AI). https://www.bankofengland.co.uk/prudential-regulation/publication/2026/november/effective-practices-cyber-risk-and-resilience-management-relating-to-artificial-intelligence
- National Cyber Security Centre. (n.d.). Managing the cyber risk of agentic AI. https://www.ncsc.gov.uk/blogs/managing-the-cyber-risk-of-agentic-ai
- Puglisi, B. C. (n.d.). Checkpoint-Based Governance (CBG). basilpuglisi.com. https://basilpuglisi.com/cbg/
- Puglisi, B. C. (n.d.). CBG for agents. basilpuglisi.com. https://basilpuglisi.com/cbg/agents/
FAQ
What did the PRA, FCA, and Bank of England publish?
The PRA, FCA, and Bank of England jointly issued effective practices on cyber risk and resilience management relating to artificial intelligence. The page shows a publication date of 19 November 2026.
Are the effective practices mandatory?
No. The page says the effective practices are complementary to, but separate from, the supervisory statement and final guidance on cyber risk and resilience management, and are “not part of the regulators’ formal expectations.” Firms can adopt the practices in a manner appropriate to their size, complexity and risk profile.
What do UK financial regulators say about human oversight of AI?
The introduction says it is important not to over-rely on AI and that appropriate human oversight remains essential, especially for key decisions. It says the practices are particularly relevant for AI systems whose autonomy and capabilities create high cybersecurity risks.
What oversight approaches does the NCSC describe for agentic AI?
The NCSC names human-in-the-loop, where humans approve actions before they happen; human-on-the-loop, where humans monitor actions and can intervene if needed; and human-out-of-the-loop, where the AI acts autonomously without human review. For significant consequences it recommends making named individuals or a group responsible.
What is the difference between Responsible AI and AI Governance for agents?
Under this column’s definitions, Responsible AI is machine checking machine, such as monitors, kill switches, and judge models. AI Governance adds a named person who accepts, modifies, or rejects what the agent proposes, with a record of the decision. This is the column’s framing, and the regulator and NCSC pages do not use those terms.
How can a firm tell whether its agent checkpoint is real?
This column suggests watching the override rate and time-to-decision together, then testing a sample of approvals with a second person. The open question for each firm is whether its checkpoints change agent decisions or only record them. These are practitioner suggestions and are not requirements from the regulators.
AI disclosure: This column was researched and drafted with AI assistance by HAIA Agents (RAI by GrokBot), using primary sources verified at time of publication. #AIgenerated
Leave a Reply