What AI use cases and developments does IOSCO identify in capital markets, and what benefits and risks accompany their adoption?

The report shows increasing AI use across capital markets. Observed applications include market analysis, forecasting and portfolio optimisation; support and risk-management functions; fraud detection and prevention; compliance; and client service and engagement. Previous IOSCO work also identified advisory and support services, client identification and monitoring, trading-algorithm selection, asset and portfolio management, robo-advising, algorithmic trading, investment research, sentiment analysis, surveillance and AML/CFT transaction monitoring. GenAI is increasingly moving from proofs of concept and pilots towards broader integration, while Agentic AI is still largely at the experimentation or proof-of-concept stage. Members also reported experimentation with AI used to assess other AI systems, synthetic data, and combinations of AI and blockchain. The principal drivers are operational efficiency and cost reduction, improved collection of high-quality data, client personalisation and experience, and risk management and fraud detection; AI is also reported to support compliance, reporting and detection of cyber threats. Barriers include governance gaps, reliance on third-party models/providers and shortages of skilled personnel. Risks remain highly contextual and include governance and oversight, model and algorithm development and testing, data quality and bias, transparency and explainability, outsourcing and concentration, malicious uses, ethical concerns and human-AI interactions. GenAI adds uncertainty and hallucination risks, while Agentic AI can increase complexity through planning, memory, external access and greater autonomy, with potential data, cybersecurity, operational, unexpected-behaviour and cascading-failure risks. IOSCO consequently supports risk-based and proportionate supervision, with greater supervisory attention for higher-risk or higher-impact uses.

What does the first layer of the toolkit – areas of supervisory consideration – cover?

The first layer gives supervisors a framework for identifying where closer attention may be warranted and where supervisory resources can be deployed most effectively. It builds on IOSCO’s earlier AI work and focuses on the characteristics of the AI system, its use and deployment, its lifecycle, the degree of human oversight and its potential effects on investors, firms and markets. Supervisors are encouraged to assess the nature and complexity of the system, potential vulnerabilities in models or data, inappropriate use or cyber incidents, the effectiveness of risk mitigation and the firm’s governance and risk-management arrangements. Human oversight can range from human-in-control, where the human may use or disregard the AI output; human-in-the-loop, where AI acts only after human approval; human-on-the-loop, where AI acts unless the human disapproves; and human-out-of-the-loop, where the system acts without human involvement. Impact considerations include investor protection, suitability, privacy and data protection, non-discrimination, the number and type of investors affected and the availability of redress, as well as potential disruption to firms’ core activities and operational, legal or reputational consequences. Comprehensive impact assessment or risk mapping should support proportionality: higher-risk or higher-impact systems should receive more stringent examination, while lower-risk systems may receive more streamlined treatment. IOSCO also notes that systemically important institutions may warrant heightened consideration given their potential market-wide impact and that common dependence on third-party AI providers can create systemic vulnerabilities and cascading effects.

What tools does the second layer provide for supervisory oversight of governance and risk management, third-party and outsourcing risk, disclosure, and recordkeeping and reporting?

The second layer contains detailed supervisory tools and questions covering four areas. Governance and risk management should operate across the AI lifecycle, supported by documented policies, clear accountability, appropriate human oversight, escalation arrangements and audit trails. IOSCO highlights Board and senior-management oversight, firm-wide AI policies, AI inventories, risk and impact assessments, model validation and testing, ongoing monitoring for drift, degradation and bias, data governance, staff capability and training and practices specific to GenAI and other advanced systems. For third parties and outsourcing, firms should conduct risk-proportionate due diligence, assess transparency and fairness, validate third-party and open-source components, manage supply-chain and concentration risks, maintain contingency and exit arrangements and include appropriate contractual provisions on performance, data protection, audit and notification. The supervised firm retains responsibility, legal liability and accountability for outsourced tasks. Disclosure should give investors and clients meaningful information about material AI use, performance and limitations, relevant risks, governance and, where appropriate, client consent or opt-out arrangements and third-party dependencies. IOSCO also highlights misleading or promotional AI claims, or ‘AI-washing’. Recordkeeping should preserve documentation and logs throughout the AI lifecycle, including system development and changes, model choices, AI-generated outcomes, human interventions, explanations and incidents. Reporting should provide supervisors with timely and sufficiently detailed information, including significant failures, biases, cybersecurity events, data breaches and other reportable AI incidents, while allowing existing reporting frameworks to be reviewed as AI use evolves.

What does the third layer – indicators and data sources for monitoring AI use – provide to supervisors?

The third layer supports ongoing supervisory monitoring by providing indicators of AI adoption, use, dependencies and performance and by identifying ways to collect relevant information. Indicators include the proportion of firms using AI; the proportion of use cases in development or production; AI research and development spending and patents; inventories of AI use cases and system types; information on underlying models and system attributes; dependence on and concentration among third-party providers; the frequency and severity of failures, errors, outages, cyber incidents, data breaches and misconduct; and model-performance measures such as prediction accuracy, error rates, drift and degradation. IOSCO also gives sector-specific examples, such as AI-driven mis-selling or unsuitable recommendations for broker-dealers, valuation and portfolio-drift measures for asset managers, and market-quality, surveillance, volatility and settlement indicators for exchanges and other market intermediaries. Supervisors may obtain information through supervisory reporting, examinations and inspections, structured engagement with firms, sector-wide surveys, publicly available information, real-time monitoring tools, AI-assisted analysis, information sharing among authorities, and sandboxes, innovation hubs and other engagement initiatives. The purpose is to identify adoption patterns and areas of heightened risk, prioritise supervisory engagement and detect emerging risks and vulnerabilities, while maintaining a risk-based and proportionate approach.