A global financial watchdog has urged banks and other financial institutions to strengthen human oversight as artificial intelligence systems become more capable of making decisions and carrying out tasks independently. The Financial Stability Board (FSB) published its final guidance on responsible AI adoption on October 15, 2026, highlighting the need for effective governance, risk management and accountability.
The recommendations come as financial institutions explore ways to use advanced AI to improve efficiency, analyse information and automate routine work. However, the growing independence of these systems has raised concerns about cybersecurity, operational failures and the potential consequences of decisions made without sufficient human supervision.
Why the Global Financial Watchdog Is Concerned About Agentic AI
The Financial Stability Board coordinates international work on financial regulation and stability. Its latest guidance outlines 12 sound practices intended to help financial institutions adopt AI responsibly while managing the risks associated with its use.
According to the FSB, these practices cover organisation-wide governance as well as the different stages of AI development and deployment. They emphasise the responsibilities of senior management, appropriate risk assessments, data governance, system evaluation and effective human oversight.
The recommendations are particularly relevant as agentic AI systems become capable of completing multi-step tasks with limited human intervention. Unlike conventional software that follows fixed instructions, these systems can plan actions, use digital tools and adjust their approach as a task develops.
This flexibility may help financial businesses work faster, but it also creates new challenges. When an AI system makes several connected decisions, an error at one stage could affect subsequent actions before a person has an opportunity to intervene.
What Is Agentic AI and Why Does Human Oversight Matter?
Understanding Autonomous AI Systems
Agentic AI refers to artificial intelligence designed to pursue goals by planning and carrying out tasks with varying degrees of independence. Depending on the system, it may retrieve information, analyse data, interact with software and take actions based on its findings.
In banking, these capabilities could support fraud detection, customer service, compliance checks and internal operations. An AI agent might review a transaction, flag suspicious activity and prepare a report for a human employee to assess.
The benefits depend on how reliably the system operates and how well its responsibilities are defined. AI-generated results can be inaccurate, incomplete or influenced by poor-quality data, making careful evaluation essential.
The Importance of Human Decision-Making
Agentic AI human oversight means ensuring that people remain accountable for the way AI systems are designed, deployed and used. It does not necessarily require a person to approve every routine action, but it does require appropriate supervision based on the level of risk.
For example, an AI tool might be allowed to organise documents automatically while a decision involving a major financial transaction requires additional approval. Institutions can also establish limits on what an AI agent may access, change or authorise.
Clear escalation procedures are equally important. When an AI system produces an unexpected result, encounters a high-risk situation or exceeds its permitted authority, it should be possible to alert a responsible employee and restrict further action.
Key Details of the Financial Stability Board’s Guidance
1. Stronger AI Governance
The FSB recommends that financial institutions establish clear responsibilities for managing AI-related risks. Boards and senior management should understand how AI adoption fits into the organisation’s strategy, risk appetite and operating model.
This approach can help prevent responsibility from becoming unclear when an automated system makes a mistake. Employees and management teams should know who monitors the technology, who approves its use and who is responsible for responding to problems.
2. Risk Assessments Throughout the AI Lifecycle
AI risks can emerge during development, testing, deployment and everyday operation. A system that performs well in a controlled environment may behave differently when it encounters unfamiliar data or changing market conditions.
Regular assessments can help institutions identify weaknesses before they lead to serious problems. They should also evaluate whether an AI application is appropriate for its intended purpose and whether additional controls are needed.
3. Better Cybersecurity and Data Protection
Financial institutions manage sensitive customer information and operate systems that support payments, lending and other essential services. AI tools connected to these environments may create additional points of exposure if access permissions and security controls are inadequate.
The FSB’s guidance highlights the importance of managing AI-related cyber risks and dependencies on external technology providers. Institutions should understand which systems their AI tools rely on and how they would respond if a service became unavailable or compromised.
4. Effective Human Intervention
Human oversight must be practical rather than merely a written policy. Employees need sufficient information, authority and training to identify problems and intervene when necessary.
Institutions can introduce approval requirements for sensitive decisions, continuous monitoring for high-risk applications and emergency procedures for stopping or restricting automated operations. These measures help keep human accountability in place even when an AI system performs much of the work independently.
How AI Could Affect Global Financial Stability
Artificial intelligence offers financial institutions opportunities to reduce repetitive work, improve information analysis and strengthen certain compliance processes. When properly managed, these capabilities may support more efficient services and help employees focus on complex decisions.
However, widespread reliance on similar AI systems could create risks that extend beyond an individual institution. If several banks use comparable models, data sources or technology providers, their systems may respond to market events in similar ways.
Cybersecurity is another important concern. An attack affecting a widely used provider or connected financial systems could disrupt multiple institutions. Inadequate model testing, unreliable data and excessive dependence on automated decisions could also increase operational vulnerabilities.
The FSB previously highlighted these broader concerns in its November 2024 report on the financial stability implications of AI. Its analysis identified cyber risks, model risk, data governance and concentration among third-party providers as areas requiring attention.
The latest guidance builds on this wider regulatory effort by setting out practical approaches for managing AI throughout its lifecycle. It encourages financial institutions to pursue innovation while maintaining safeguards that support the resilience of the financial system.
What the Guidance Means for Banks and Financial Institutions
Banks may need to review existing AI policies, update risk assessments and establish more detailed procedures for supervising autonomous systems. The level of control should reflect the potential consequences of a system’s actions.
Low-risk applications may require routine monitoring, while tools involved in sensitive customer or financial decisions could need stricter approval processes. Institutions should also maintain records that help employees understand what a system did, which information it used and why an action was taken where that can be established.
Training will be important as well. Staff members responsible for AI oversight need to understand both the technology’s capabilities and its limitations. Without appropriate knowledge, employees may struggle to recognise errors or may place too much confidence in automated recommendations.
Smaller institutions could face additional implementation challenges because they may have fewer technical specialists and less funding for independent testing. Clear responsibilities, proportionate controls and cooperation with trusted technology providers can help them manage these demands.
The guidance is intended to support responsible adoption rather than prevent financial institutions from using AI. Its practical effect will depend on how organisations apply the recommendations and how national authorities incorporate AI-related risks into their supervisory work.
The Global Impact of Human Oversight in Artificial Intelligence
The debate surrounding agentic AI is not limited to banks. Governments, regulators and businesses across industries are considering how to maintain accountability as automated systems become more capable.
In finance, the consequences of poor controls can extend to customers, businesses and markets. An unreliable system could contribute to incorrect decisions, service disruptions or security incidents, depending on the task it is permitted to perform.
International coordination is therefore important. Financial institutions often operate across borders and depend on technology services used in multiple countries. Cooperation among regulators can help improve information-sharing and make it easier to identify risks that affect several markets.
For countries developing their digital financial sectors, including Bangladesh, the wider discussion is relevant to the adoption of AI in banking and financial services. Responsible implementation can help institutions explore new technology while protecting customer information and maintaining appropriate human accountability.
Bangladesh cricket is a separate subject and is not directly connected to the FSB’s financial AI guidance. It should be covered in a separate sports article rather than inserted into this report as an unrelated keyword.
What Comes Next for Agentic AI Regulation?
The FSB’s final guidance provides financial institutions with a framework for improving governance, risk management and oversight. The next challenge is turning these recommendations into everyday practices that work across different institutions and jurisdictions.
Banks will need to evaluate their existing systems, identify gaps and determine whether current controls are sufficient for increasingly autonomous AI applications. Regulators will also need to monitor technological developments and consider whether existing supervisory approaches remain effective.
The pace of innovation means that risk management cannot be treated as a one-time exercise. AI systems may change through updates, new integrations or shifts in the data they process, making ongoing evaluation important.
At the same time, excessive restrictions could make it harder for institutions to benefit from useful technology. A proportionate approach can help balance innovation with the need to protect customers, institutions and the wider financial system.
Conclusion
The Financial Stability Board’s latest guidance highlights why agentic AI human oversight must remain a central part of responsible financial technology adoption. Autonomous systems can improve efficiency and support complex work, but their use also creates risks that require clear accountability, effective controls and continuous monitoring.
For banks and regulators, the priority is to ensure that AI operates within defined boundaries and that people can intervene when necessary. As the technology develops, strong governance and practical human supervision will be important for maintaining trust and protecting global financial stability.
Frequently Asked Questions
1. What are the four types of agentic AI?
There is no universally accepted classification of agentic AI into exactly four types. One practical way to understand agentic systems is to group them by capability: simple reactive agents respond to current inputs, goal-based agents plan actions to achieve specific objectives, utility-based agents compare possible actions according to defined preferences, and learning agents improve their behaviour using experience or feedback. These categories are useful for understanding different approaches to autonomous decision-making, although modern AI systems can combine several of these characteristics.
2. What are seven human skills that AI cannot replace?
AI can assist with many professional tasks, but important human capabilities remain essential. These include emotional understanding, ethical judgement, interpersonal communication, leadership, creativity grounded in lived experience, complex contextual reasoning and relationship-building based on trust. AI can imitate aspects of some of these skills and support people in applying them, but its outputs do not automatically provide human empathy, moral responsibility or genuine personal experience. Their importance is especially clear when financial decisions affect customers or require accountability.
3. Does AI need human oversight?
Yes, human oversight is important when AI systems can influence sensitive decisions, handle personal information or take actions with significant consequences. The level of supervision should depend on the system’s purpose, capabilities and potential risks. Routine, low-risk tasks may be handled with automated monitoring, while high-impact decisions may require human review or approval. In financial services, clear intervention procedures, access controls, audit records and assigned responsibilities can help prevent mistakes from becoming larger operational or security problems.
4. What is the 30% rule for AI?
The 30% rule for AI is an informal expression used in some workplaces to describe the idea that AI can complete or accelerate a portion of a task while a human contributes the remaining judgement, review and expertise. It is not a universal scientific standard or an official rule established by the Financial Stability Board. The actual amount of work AI can perform varies considerably according to the task, available data, system quality and required accuracy. Organisations should measure real performance rather than assume that AI will always deliver a fixed percentage of productivity gains.
