FSB AI practices: oversight, vendors and India context

The Financial Stability Board published 12 practices for responsible artificial-intelligence adoption on 15 October, putting human accountability and controls on autonomous AI into a framework for financial institutions. The final report covers oversight across an institution and risk management throughout an AI system’s life.

The practices are voluntary guidance, rather than a binding international standard. Banks and other financial firms are encouraged to adapt them to their size, complexity and the risks of each use. Applying the toolkit does not replace compliance with domestic law.

For AI agents, the report describes limiting system access, setting triggers for human approval and testing ways to pause or remove permissions. It also calls for named accountability when machines monitor other machines. The monitoring AI should operate independently of systems it evaluates. These controls address what an agent can actually do, beyond checking the text it produces.

What consultation exposed

The separately published consultation overview records around 160 responses. Some respondents asked for unsuccessful cases and more examples relevant to smaller institutions and nonbanks. The FSB added insurer and other nonbank examples, but retained its caution that case studies illustrate possible outcomes rather than benchmarks.

The FSB says the final version adds detail on agentic AI and multi-agent systems, but acknowledges that these additions were drawn from consultation responses and were not fully consulted on. Further work may follow. It also retained a boundary around this report: wider systemic and macroprudential concerns, including the effects of shared providers across finance, were not incorporated into its scope.

The overview identifies expanded coverage of AI embedded in third-party products and subcontracted services. It points to more timely assurances from suppliers and collaboration on contractual terms, joint scenario tests and oversight of supply-chain dependencies. That matters because a financial firm’s direct supplier may itself depend on another AI or cloud provider, extending the chain of services it relies upon.

The practical difficulty is visible in the original submission from the Bank Policy Institute and Institute of International Bankers, dated 22 July. The banking groups argued that institutions cannot always inspect every link in a supplier chain or see proprietary model design and training. They asked supervisors to calibrate expectations to the information and control actually available, while recognising that financial institutions remain accountable for deployment.

Their submission also disputed the assumption that a firm can readily switch AI providers during a disruption. A replacement could behave differently, introduce different biases and affect essential operations. The groups proposed continuity arrangements, service-degradation procedures, fallback controls, contractual protections and incident response where substitution is impractical. These are arguments made during consultation, not evidence that every bank has such arrangements or that a supplier change is impossible.

On oversight, the groups distinguished agents carrying out tightly constrained processes from systems with greater autonomy. Their proposed approach combined action logs and clear intervention thresholds with limits on delegated authority and access. They cautioned against prescribing human review as the answer in every setting, arguing that the control should fit the particular task and its risk.

The Indian comparison

For India, a relevant comparison is the Reserve Bank of India’s FREE-AI committee report, published on 13 August 2025. It proposed seven guiding principles and 26 recommendations. Its consumer-protection section recommended telling customers when they are interacting with AI, offering access to a human representative and providing an accessible route to challenge problems. These are committee recommendations, not new Indian requirements announced by the FSB.

The Indian committee also separated customer compensation from supervisory treatment. It recommended that regulated firms remain responsible for customer losses, while allowing a more accommodating supervisory response to a first isolated failure where safeguards had been followed. That proposed tolerance would not extend to repeated breaches, gross negligence or failure to correct known problems. The distinction preserves redress while leaving room for supervised experimentation.

Its proposed board-approved AI policy covered accountability, risk appetite, operational safeguards, auditability and liability. A suggested policy outline also asked firms to define vendors’ responsibilities.