On October 15, the Financial Stability Board (FSB), the international body that coordinates financial regulators and standard setters for the G20, published its final Sound Practices for Responsible Adoption of Artificial Intelligence. The 12 practices are non-binding, but the FSB strongly encourages boards and senior management to reference them when they build AI governance, risk management and control frameworks. The report names anti-money laundering (AML), know-your-customer (KYC) checks and financial crime investigations among the areas where institutions already use AI.
For compliance and risk leaders, a key question is what you must be able to demonstrate when AI, and increasingly agentic AI, sits inside your controls, and what you must require of the providers supplying it. That question is arriving first in cryptoassets, where AI agents already move funds on-chain and agentic cryptoasset payments rose 500% in the three months to September, according to Elliptic data.
The FSB expects firms to run due diligence on every third-party AI provider and keep oversight of what they buy, but it stops short of saying what a provider should be able to evidence. That is the question the Elliptic Standard answers. Its eight principles set out what agentic on-chain risk solutions must do and what a regulated firm must expect of any provider. Read against the FSB AI sound practices, the direction of travel is the same. Here is where they meet.
Accountability does not transfer to the agent
Accountability does not transfer to the agent
The FSB opens with governance: boards and senior management aligning AI adoption with strategy and risk appetite, clear roles and responsibilities, and risk frameworks that document AI use. The common thread is that responsibility stays with the institution.
The Elliptic Standard is already specific where the FSB is still provisional. Principle 6 treats human oversight as non-negotiable and sets out what a person answerable for an agent needs:
An explicit scope of authorityClear attribution of every action to an agent and an accountable ownerVisibility of activity in progressA decision trailEvidence that oversight was exercised, not merely available.
Principle 7 adds institutional memory, so audit trails and records of change survive staff turnover. The test is the same in both documents: Can you show what was decided, on what basis, and who was accountable?
Data quality and explainability are the foundation
Data quality and explainability are the foundation
The FSB asks institutions to ensure data is fit for purpose, meaning accurate, complete, consistent, reliable and secure, and to understand the explainability of the AI they deploy, adopting more explainable approaches or compensating controls where appropriate.
Principles 1 and 2 of the Standard turn these expectations into criteria a buyer can test. On data, we call for coverage of at least 99% of value-moving transactions, high-confidence attribution throughout the dataset, consistent outcomes for identical inputs and an evidence trail behind every claim.
Where coverage is incomplete, the gap must be surfaced, not absorbed into a confident output. On transparency, agentic systems must not be black boxes. Evidence should be available as a machine-readable verdict, a plain-language summary and the underlying structured data, so that users and regulators can reproduce the conclusions.
Performance and oversight are continuous, not a one-off
Performance and oversight are continuous, not a one-off
The FSB expects performance to be evaluated through testing and ongoing monitoring, and human oversight to be effective in practice. These are the areas where abstract principles most often fail in implementation.
The Standard addresses both directly. Principle 2 requires validation to continue after deployment, to catch drift and degradation as models are retrained. Principle 7 requires uptime evidenced above 99.9%, measured and reported rather than asserted, along with a defined fallback for transactions in flight. Principle 5 keeps the decisions that matter with the firm: the user sets risk appetite thresholds, defines escalation and human oversight logic, aligns outputs to their own procedures, overrides agent determinations and defines what counts as a match.
Outsourcing the AI does not outsource the risk
Outsourcing the AI does not outsource the risk
The FSB also asks institutions to adapt as AI evolves, to select models and systems deliberately, and to manage third-party AI risk, including concentration and business continuity.
Principle 8 of the Standard addresses the human side: training that covers AI safety, typology comprehension, and how to interrogate an agent’s reasoning rather than simply accept or override a score. The Elliptic Institute’s free certification in governing AI agents, launched on October 20, teaches those skills through the Standard’s principles. Principle 4 addresses dependency through foundation model agnosticism, which reduces exposure to a single model’s failure, supports multi-jurisdictional obligations, and allows private or customized models where sensitive data should not leave the firm. The Standard is also explicitly designed for vendor assessment. Firms should expect a substantive answer from any provider on every principle.
What this means in practice
What this means in practice
Firms do not need to choose between the FSB AI sound practices and a provider’s standard. The FSB sets the supervisory expectation of the institution and the Elliptic Standard translates it into questions you can put to a provider, with criteria you can test. In our experience, the strongest governance outcomes come when the two are used together.
Elliptic has adopted the Standard’s principles in building its own agentic capabilities, and we expect to be held to them. As supervisors converge on a common view of responsible AI, the firms that can evidence what their systems decided, why, and under whose authority will be best placed. For what that evidence looks like in agentic AI risk management, and what to demand of any provider, read the Elliptic Standard.