What does IOSCO FR/02/2026 say about AI in banking?

What are the key points of IOSCO FR/02/2026?

• Final report of the Board of IOSCO, FR/02/2026, May 2026; the media release IOSCO/MR/13/2026 is dated Madrid, 25 May 2026; stakeholder feedback was invited by 26 June 2026.
• Method: a survey of IOSCO Fintech Task Force members (21 responded), roundtables in Tokyo, Singapore and New York, and a literature review; builds on FR06/2021 and CR/01/2025.
• Three layers: (1) areas of supervisory consideration; (2) tools for governance and risk management, third-party and outsourcing risk management, disclosure, and recordkeeping and reporting, with example questions for supervisory dialogue and examinations; (3) indicators and data sources for monitoring AI use (Section 3.5).
• Covers all AI system types: traditional machine learning, generative AI and emerging agentic AI, which the report says can access sensitive data, communicate externally and create cascading failures if misconfigured.
• Risk-based and proportionate supervision (Section 2.4, Box 2): factors are nature and complexity, level of human oversight (human-in-control, in-the-loop, on-the-loop, out-of-the-loop) and impact on clients or the firm.
• Systemic point: larger, systemically important institutions 'may face heightened supervisory expectations even for medium or low-risk AI applications', and supervisors may assess concentration where multiple firms rely on the same third-party AI provider.
• Table 2 areas of supervisory consideration include AI governance and oversight, model risk management, investment advice and suitability (including 'AI-washing' and embedded sponsored content in AI advisory conversations), market risks (including kill-switch and circuit-breaker policies), and system reliability and business continuity.
• Box 1 on hallucination: grounding and guardrails (RAG, Chain-of-Verification, multi-agent debate) 'are unlikely to fully eliminate' hallucinations; human oversight is subject to automation bias; the firm deploying the system is primarily responsible.

What did IOSCO FR/02/2026 change for banks?

The toolkit converts IOSCO's earlier high-level measures (FR06/2021) and risk survey (CR/01/2025) into examination-ready supervisory tools and extends them to generative and agentic AI. For a bank with securities businesses, it signals the questions national securities regulators are likely to ask about AI governance, third-party concentration, disclosure and recordkeeping.

What does the IOSCO Supervisory Toolkit for AI Use in Capital Markets require of firms?

The IOSCO Supervisory Toolkit for AI Use in Capital Markets (FR/02/2026, May 25, 2026) requires nothing of firms directly; it gives IOSCO member supervisors 'non-binding, non-prescriptive' tools for overseeing firms' AI. The toolkit covers the full AI lifecycle for traditional machine learning, generative AI and agentic AI across three layers: areas of supervisory consideration, tools for governance and risk management, third-party and outsourcing risk management, disclosure, and recordkeeping and reporting, and indicators for monitoring AI use. It tells supervisors to apply risk-based and proportionate oversight, weighing complexity, human oversight and impact, with possible heightened expectations for systemically important institutions. For a bank, it previews the questions securities regulators may ask of its broker-dealer, trading and asset-management businesses.

FR/02/2026 is the last step in what IOSCO calls a multi-phased approach: FR06/2021 gave six measures for firms using AI and ML, CR/01/2025 mapped use cases, risks and challenges, and the 2026 toolkit gives supervisors examination tools. The report says it responds to adoption shifting 'from proofs of concept and pilot programs to broader integration' and to concerns about governance, accountability and control, especially for agentic AI used in retail financial services.

The proportionality section is the most useful for banks. The report asks supervisors to look at the type of system, how much human oversight it has and what the impact on clients or the firm would be, and it says systemically important institutions can face heightened expectations even for lower-risk applications because of their market-wide impact, and that supervisors can examine systemic vulnerabilities where several institutions depend on the same third-party AI provider. These are signals about where securities regulators will focus on bank-owned securities firms.

Compared with the EU AI Act and PRA SS1/23, IOSCO's toolkit is the most operational for examinations: it is built around tables of concerns, evidence to request and example questions, and a separate standalone extract (OR/07/2026) is meant for on-site use. It is not a compliance regime; whether a firm is asked these questions depends on whether its national securities regulator uses the toolkit. IOSCO's next phase, a review of emerging industry practices on disclosure, recordkeeping and reporting, and governance, will show where expectations settle.

WHAT THIS MEANS IN PRACTICE

• Prepare evidence for likely supervisory requests: an AI inventory, governance committee records, training records, validation and monitoring reports, and human-override documentation, which the toolkit lists as evidence to review.
• Classify AI use cases by human-oversight level and client impact so you can show proportionate controls to a securities regulator.
• Map third-party AI providers and concentration across the group, since the toolkit asks supervisors to look at shared dependencies.
• Review marketing and client disclosures for AI-washing and for any sponsored content embedded in AI advisory conversations.
• Treat generative and agentic AI as in scope: test for hallucination, and apply kill-switch, access and data-exfiltration controls to agents.