FSB AI Sound Practices Meet Agentic Risk in Crypto

On 15 October 2026, the Financial Stability Board published its final report on sound practices for responsible AI adoption in financial services. The 12 practices are non-binding, but the FSB explicitly encourages boards and senior management to reference them when building AI governance, risk management, and control frameworks. For crypto compliance and risk teams, the timing matters: agentic AI is already moving funds on-chain, and the FSB's own report names AML, KYC checks, and financial crime investigations as live deployment areas. The central question the report raises, but does not fully answer, is what a firm must be able to demonstrate when AI sits inside its controls, and what it must require of the vendors supplying that AI.

A published eight-principle standard for agentic on-chain risk solutions, developed by Elliptic, addresses precisely that gap. Reading the two documents together, the direction of travel is the same. This article sets out where they align, what each one requires in practice, and what compliance teams and CFOs should now be asking of every AI provider in their supply chain.

What the FSB Actually Requires

The FSB is the international body that coordinates financial regulators and standard-setters for G20 members. Its AI sound practices carry weight even without binding force, because national supervisors look to the FSB's output when designing their own rules. Firms that demonstrate alignment with the 12 practices are, in effect, anticipating the regulatory trajectory in every major jurisdiction.

Governance and accountability

The FSB opens with governance. Boards and senior management are expected to align AI adoption with overall strategy and risk appetite, establish clear roles and responsibilities, and maintain risk frameworks that document every AI use case. The core principle is that legal and regulatory accountability stays with the institution, not with the AI vendor. No model score or agent action transfers that responsibility to a third party.

Data quality and explainability

The FSB asks institutions to ensure that data feeding AI systems is accurate, complete, consistent, reliable, and secure. It also asks firms to understand the explainability of the models they deploy, and to adopt more transparent approaches or put compensating controls in place where a model cannot explain its outputs. In AML and KYC contexts, this is not a theoretical concern: an alert that cannot be explained cannot be defended to a regulator, and a false negative that cannot be traced cannot be corrected.

Performance monitoring and third-party risk

The FSB expects performance to be evaluated through testing and ongoing monitoring, not just at deployment but continuously. It also asks institutions to manage third-party AI risk, including concentration risk and business continuity. The report notes that practices specific to agentic AI remain nascent at the supervisory level. That gap is significant, because agentic systems make sequential, autonomous decisions, and standard model-validation approaches were not designed with that architecture in mind.

Where the Eight-Principle Standard Fills the Gap

The Elliptic Standard translates the FSB's institutional expectations into criteria a buyer can test against any provider. Its eight principles cover data coverage, transparency, validation, human override, foundation model agnosticism, training, and institutional memory. Each one maps onto one or more FSB practices, and together they provide the vendor-assessment checklist that the FSB stops short of specifying. For firms that rely on crypto compliance reporting infrastructure, this kind of structured assessment is increasingly a board-level obligation, not just an operational preference.

Data coverage and attribution (Principles 1 and 2)

Principle 1 of the Standard sets a coverage floor of at least 99 percent of value-moving transactions, with high-confidence attribution throughout the dataset and consistent outcomes for identical inputs. Where coverage is incomplete, the gap must be surfaced explicitly, not absorbed into a confident output. Principle 2 requires transparency: outputs must be available as a machine-readable verdict, a plain-language summary, and the underlying structured data, so that both users and regulators can reproduce the conclusions reached. This directly addresses the FSB's explainability requirement and gives compliance staff a basis for constructing audit trails that will satisfy supervisory review.

Post-deployment validation (Principle 2, continued)

The Standard requires validation to continue after deployment, to catch model drift and performance degradation as underlying models are retrained. This is the area where abstract AI governance principles most commonly fail in practice. A model that passes initial testing can deteriorate quietly as on-chain typologies evolve. Continuous monitoring is not optional in an environment where agentic AI handles real transactions in real time. Teams reviewing their digital asset accounting software stack should ask vendors directly how they detect and report post-deployment drift.

Human oversight and user-configurable controls (Principles 5 and 6)

Principle 5 keeps consequential decisions with the firm. Users set risk-appetite thresholds, define escalation logic, align outputs to their own internal procedures, override agent determinations, and specify what constitutes a match. Principle 6 treats human oversight as non-negotiable and sets out what the person answerable for an agent actually needs to exercise that oversight in practice, including access to reasoning trails, not just scores.

The FSB's governance principle carries the same message: accountability stays with the institution. The Standard makes it operational by defining the control architecture through which that accountability can be exercised and evidenced. Firms evaluating crypto bookkeeping software or broader AML tooling should treat vendor responses to these two principles as a threshold test, not a differentiator.

Institutional memory and uptime (Principle 7)

Principle 7 addresses two risks that governance frameworks often overlook. First, audit trails and records of change must survive staff turnover. Institutional memory about why a threshold was set, or why an agent determination was overridden, cannot exist only in someone's head. Second, uptime must be evidenced above 99.9 percent, measured and reported rather than asserted, with a defined fallback for transactions in flight. The FSB's business continuity expectation maps directly onto both requirements.

Foundation model agnosticism (Principle 4)

Principle 4 reduces dependency on any single model by requiring agentic systems to be foundation model agnostic. This lowers exposure to a single model's failure or deprecation, supports multi-jurisdictional regulatory obligations that may impose different requirements on model use, and allows private or customized models where sensitive data should not leave the firm's environment. The FSB asks institutions to manage concentration risk in third-party AI relationships. Model-level lock-in is one of the most underappreciated forms of that concentration.

Training (Principle 8)

Principle 8 covers the human side of agentic AI governance: training that addresses AI safety, typology comprehension, and how to interrogate an agent's reasoning rather than simply accept or override a score. The FSB equally stresses effective human oversight, but effectiveness requires people who understand what they are overseeing. The Elliptic Institute launched a free Certified AI On-Chain Risk Practitioner course on 20 October 2026, covering accountability, oversight mechanics, and how to question an agent's reasoning. It is self-paced and takes 90 minutes, with a certificate on completion.

Practical Implications for Compliance Teams and CFOs

The FSB and the eight-principle Standard are not competing frameworks. The FSB sets the supervisory expectation at the institutional level; the Standard translates it into questions that can be put to a provider, with criteria that can be tested. Used together, they give compliance leads and CFOs a structured basis for vendor due diligence, internal audit, and board reporting on AI risk.

What to ask every AI provider now

Based on the alignment between the two frameworks, the following are the minimum questions every firm should put to any AI vendor operating in their AML, KYC, or financial crime controls:

• What is your transaction coverage rate, and how do you surface gaps rather than mask them in outputs?
• Can you provide outputs in machine-readable form alongside a plain-language explanation and the underlying structured data?
• How do you detect and report model drift after deployment?
• What controls allow our team to set thresholds, define escalation logic, and override agent determinations without requiring vendor involvement?
• What is your documented uptime SLA, how is it measured, and what is the fallback for in-flight transactions during an outage?
• Are your agentic systems tied to a single foundation model, or can they operate across models, including private deployments?
• What training do you provide for staff who will oversee agent outputs?

Firms that have already reviewed their exposure to tightened AI accountability rules from the FSB, HKMA, and FCA will recognize these questions as consistent with where international supervisors are heading. Teams that have also assessed how AI agents can reach crypto mixers directly will understand why the adequacy of agentic controls is no longer an abstract governance question.

Accounting and audit implications

For accounting firms and CFOs supporting digital asset clients, the FSB's emphasis on documentation and audit trails has a direct balance-sheet dimension. Where AI agents are involved in transaction monitoring, the records generated by those agents form part of the evidentiary basis for AML sign-off. If those records cannot be reproduced, explained, or attributed to a named accountable person, the accounting team faces a control deficiency that external auditors and regulators will both identify.

Firms using crypto accounting software to manage digital asset portfolios should now assess whether the AI components of that software, if any, can satisfy the FSB's data quality and explainability standards. If a tool produces a risk score or a compliance determination that the firm cannot explain to a regulator, that tool creates liability rather than reducing it. Vendor contracts should be reviewed for explicit commitments on coverage rates, explainability, uptime, and post-deployment monitoring, and the absence of such commitments should be treated as a red flag in any renewal or procurement decision.

The Broader Regulatory Context

The FSB's AI sound practices do not exist in isolation. The Basel Committee on Banking Supervision has also addressed AI in transaction monitoring. National regulators across G20 jurisdictions are issuing their own AI governance expectations for financial institutions, and several are already referencing the FSB output as a baseline. Firms that build their internal frameworks now, grounded in both the FSB practices and a testable vendor standard, will be ahead of the supervisory cycle rather than reacting to it.

The convergence between a global standard-setter and a practitioner-developed set of principles is, itself, a signal. When regulatory bodies and industry practitioners independently arrive at the same governance requirements, those requirements are close to becoming the minimum acceptable standard. For compliance leads, the practical conclusion is straightforward: document your AI use, evidence your oversight, test your vendors against published criteria, and make sure someone named and senior is accountable for every agent action your firm relies on.

Source: Elliptic

Frequently Asked Questions

Are the FSB's 12 AI sound practices legally binding?

No. The FSB published them on 15 October 2026 as voluntary sound practices. However, G20 national supervisors routinely use FSB output as a reference when designing binding rules, so firms that align with the 12 practices now are anticipating the regulatory trajectory in most major jurisdictions.

What does the FSB specifically say about agentic AI?

The FSB acknowledges that practices specific to agentic AI remain nascent at the supervisory level. This is the gap that practitioner-developed frameworks, such as the eight-principle standard discussed here, are designed to fill for firms that already deploy agentic systems in AML and KYC controls.

How should a compliance team use the eight-principle standard in vendor due diligence?

Treat each principle as a question to put to the provider, with a testable answer required. Ask for documented coverage rates, sample machine-readable outputs, evidence of post-deployment monitoring, uptime SLAs with measured data, and written confirmation of which model or models the system runs on. Require contract clauses that commit the vendor to these standards, not just marketing assertions.

What are the accounting implications if an AI vendor cannot evidence explainability?

If an AI tool produces risk scores or compliance determinations that cannot be explained or reproduced, the underlying records may not satisfy the evidentiary standards expected by external auditors or regulators. That creates a control deficiency. For digital asset portfolios in particular, where transaction volumes and complexity are high, unexplainable AI outputs can undermine the entire audit trail supporting AML sign-off.

Where can staff learn to oversee agentic AI systems responsibly?

The Elliptic Institute launched a free Certified AI On-Chain Risk Practitioner course on 20 October 2026. It is self-paced, takes around 90 minutes, and covers accountability structures, oversight mechanics, and how to interrogate agent reasoning. Learners who complete it receive a certificate. Given that the FSB explicitly requires effective human oversight, having trained staff is increasingly a governance necessity rather than an optional development activity.