国际清算银行-从基础开始_盘点监督中的人工智能应用（英）-2025.6_14页_625kb

Summary of FSI Report: Gen AI Applications in Supervision

Key Overview

This report, authored by the Financial Stability Institute (FSI), examines the current state of generative artificial intelligence (gen AI) applications in financial supervision based on a survey of 42 authorities. It highlights challenges and opportunities in leveraging gen AI tools for supervisory processes.

Key statistics:

• 32 authorities reported gen AI activities (including experimentation, development, or use), with only 12actively using such applications.
• A balanced representation across advanced economies (AEs) and emerging market economies (EMEs).

Principal Findings

1. Authority Activities

Financial authorities are actively engaged with gen AI, primarily to enhance operational efficiency through better information extraction. However:

• Challenges faced:
  • Outdated IT infrastructure, data security concerns, and a shortage of technical skills are major barriers.
  • Some authorities cite high costs of implementation and legal risks from data breaches.
  • Lack of clear strategies for integrating gen AI into core supervision processes.

2. Gen AI Use Cases

Most applications fall into three categories, based on survey responses:

• Basic document processing: Includes drafting inspection reports, summarizing, translating, or extracting key information from submissions. (Most “in use” applications).
• Knowledge management: Aids supervisors in navigating regulations, conducting research, and responding to queries through AI chatbots or tools like large language models (LLMs).
• Document review: Assists in compliance assessments, comparing entity submissions with regulations, and identifying trends in documents like board meeting minutes.
  Some developments involve complex use cases enhanced by methods like retrieval augmented generation (RAG) or Graph RAG for improved accuracy.

3. Integration and Challenges

Integration is limited, with:

• Few applications fully embedded into mandatory supervisory processes (only one authority required gen AI for completions).
• Main barriers to wider integration include user acceptance and inaccurate information provided by AI models, with hallucination risks being a concern despite some optimism about AI accuracy.
  User acceptance may stem from a lack of explainability in AI outputs, akin to issues seen in financial institutions.

4. Conclusions

Gen AI holds significant promise for improving efficiency in financial supervision, but adoption is hampered by infrastructural and technical gaps. Authorities need better integration strategies, user involvement in development, and the management of inaccuracies to leverage gen AI effectively. Forward-looking use cases, such as advanced knowledge management and document review, are emerging and require addressing challenges to scale up.

Supporting Data

• Response rate: 42 authorities from diverse regulatory areas.
• Focus areas: Banking, insurance, AML/CFT, and securities markets oversight.
• Technologies used: Cloud-based tools like ChatGPT, Microsoft Copilot, and custom-developed AI systems.