A Note for the International Conference of Banking Supervisors Says Human Oversight Could Be Weakened by an Overreliance on AI

Among the papers of the 24th International Conference of Banking Supervisors, the Bank for International Settlements published a six-page note for the conference's fourth workshop, on AI and its implications for banks and supervision. It is dated September 2026. The workshop, the note says, will focus on how AI technologies might shape the future of the banking sector, with an emphasis on supervisory considerations. Parts of it follow, not the whole.

How Banks Use AI

Banks are using or assessing AI tools across front-, middle- and back-office applications, the note says, and some supervisors report AI being deployed "in more material applications that are decision-critical and customer-facing." The uses it lists "include, but are not limited to" compliance work on knowing customers and money laundering, transaction monitoring and fraud detection, credit scoring and underwriting, software and systems modernization and development, trading, and direct customer interaction.

"However, AI tools do not always bring the hoped-for benefits," the note adds, citing a report that describes some institutions that have seen fewer or no improvements from them. Since 2024, it adds, banks have explored agentic AI tools, and some are implementing them.

The Risks It Lists

Reports from supervisors and international organizations highlight a range of risks, the note says, and it discusses some of them. Weaknesses in the quality, integrity or suitability of data, and deficiencies in model design or calibration, "could result in inaccurate, biased or unreliable outcomes." The reliance of banks around the world on a few large AI companies could pose systemic and cross-border vulnerabilities. AI tools can be vulnerable to data poisoning and prompt injection. Explainability, the note says, "can be defined as the extent to which humans understand how an AI model or system generates outputs from inputs," and limited explainability "can present problems in assessing the conceptual soundness of AI outputs and the compliance of AI tools."

"These risks can be amplified by improper governance and human oversight procedures. Human oversight could be weakened by an overreliance on AI or an incomplete understanding of the ways in which AI tools operate." The note adds that the unauthorized use of AI tools, which it calls "shadow AI," could make these risks worse and complicate risk management. As systems become more autonomous, there is also a risk "that they optimise for assigned goals in unexpected ways, leading to unintended consequences or control failures." A sign-off reads the same whether the reviewer checked the output or trusted it.

What Supervisors Are Weighing

Many supervisors are encouraging banks to adopt AI, the note says, and banks and supervisors share a view that "a failure to adapt to an AI-enabled world is a risk unto itself." Some supervisors, it says, "have expressed concerns that governance and risk management processes are not keeping pace while banks rush to adopt AI technology." Their concerns include concentration risks, a lack of AI expertise at board level, and a lack of transparency over the use of AI within organizations. Supervisors are working with banks to adopt AI responsibly, the note says, focusing on proportional responses that weigh how material the use case is.

Of the note's six questions, the fourth asks: "How are banks and supervisors striking a balance between rapid adoption of AI and ensuring that governance and risk management policies and processes are mature?"