FINANCIAL AI / GOVERNANCE & SYSTEMIC RISK
Financial AI Governance in 2026: Why Model Compliance Is Not Enough
What banks, risk leaders, and supervisors need to know about shared AI dependencies, correlated decisions, regulatory obligations, and evidence-based assurance.
Financial AI governance is the system of accountability, controls, evidence, and ongoing review used to manage AI in financial decisions and operations. In banking, it covers model quality and customer outcomes, but also supplier dependencies, human oversight, resilience, and the risk that similar AI-driven behavior may affect multiple institutions.
A bank can validate every AI model it deploys and still overlook dependencies that its individual model tests never examine. A common provider, data feed, or trading signal can affect several institutions at once—even when each firm has performed its own checks.
That distinction matters because financial stability depends on how institutions behave together—not only on whether individual models pass testing. AI can strengthen fraud prevention, credit analysis, operations, and market surveillance. It can also concentrate technology dependencies, accelerate decisions, and make errors or shocks harder to contain. The Financial Stability Board’s 2024 analysis identifies provider concentration, market correlations, cyber risk, and weak model and data governance as potential systemic vulnerabilities.
01 / THE DISTINCTION
Why a compliant AI model can still contribute to systemic risk
Consider three levels of examination. They overlap but should not be substituted for one another.
Does it work as intended?
Validation of data, performance, robustness, explainability, stability, limitations, and the intended decision context.
Can the firm use it safely?
Governance, accountability, effective human challenge, downstream controls, monitoring, business continuity, and vendor oversight.
What happens when others use it too?
Shared provider outages, correlated investment or credit responses, common data errors, and feedback through interconnected markets.
A favorable model validation does not establish that a bank can recover from an external provider failure. Nor does a bank’s successful recovery test demonstrate that the market can absorb many institutions reacting in the same direction. Conversely, shared use is a risk indicator, not proof of a systemic failure; exposure depends on actual interconnections, business importance, fallback capability, and market conditions.
What counts as effective governance? For a consumer-credit model, the record should connect the approved use to data quality, lending outcomes, valid explanations, independent challenge, and recourse when an applicant is affected. For an AI-supported operations assistant, it should identify what content the assistant may access, which actions require human authorization, how results are checked, and what happens when the service fails. A single enterprise policy is not enough to answer either question.
For regulators, the additional system-level question is whether important shared exposures or similar automated responses could amplify disruption. A single bank can identify and reduce its own exposure, but generally cannot measure the entire market’s concentration without broader information. FSB (2024).
02 / THE EXPOSURE
Five ways AI risk can spread across financial institutions
Concentration in shared AI providers
Banks may rely on the same cloud infrastructure, foundation-model provider, data intermediary, analytics platform, or identity service. A single disruption can therefore affect apparently independent control environments. Test: trace material AI services to fourth parties, hosting regions, models, and fallback options—not just contract names.
Illustrative banking case: A fraud-monitoring supplier and an investment-risk service depend on the same infrastructure region, but their procurement records list different vendors. Retain: dependency map, subcontractor or fourth-party disclosures where obtainable, service-criticality rating, restoration-test results, and supplier notice obligations.
Correlated trading and portfolio decisions
Models exposed to similar data or market signals may produce similar recommendations. During stress, simultaneous selling, risk reduction, or liquidity hoarding can worsen market conditions. This is a plausible amplification channel, not evidence that every AI-driven strategy causes herding. Test: review common-factor exposures, crowding indicators, execution constraints, and stress responses.
Illustrative banking case: Several portfolio controls reduce exposure after a market signal moves. No single model is necessarily defective, yet combined execution could increase price pressure. Retain: common factor sensitivities, order constraints, stress assumptions, liquidity overlays, and any documented limits on automated execution.
Common data and feedback errors
A mispriced input, unreliable alternative-data source, or widely reused model component may bias decisions at several institutions. If those decisions feed subsequent training or risk estimates, errors can persist. Test: record lineage, freshness, source concentration, independent cross-checks, and reversal procedures.
Illustrative banking case: An incorrect externally sourced income field affects more than one lender’s eligibility checks. Retain: data contracts, lineage and version records, quality exceptions, affected-decision sampling, and a traceable process for correction or customer remediation.
Accelerated cyber and fraud activity
AI can help defenders investigate threats but can also help adversaries generate convincing fraud or discover vulnerabilities faster. The European Banking Authority’s June 2026 assessment highlights frontier-model cyber risks and the importance of recovery capabilities. Test: evaluate attack detection, containment, service continuity, and response when AI-enabled workflows must be disabled.
Illustrative banking case: An AI-enabled investigation assistant summarizes manipulated external content and recommends an unjustified account action. Retain: access controls, tool permissions, test transcripts, human authorization steps, incident logs, and safe-disable procedures. This is a workflow-integrity concern, not necessarily a failure of a predictive credit model.
Human oversight that exists only on paper
Human review is ineffective if reviewers lack time, information, authority, or a usable alternative when an AI recommendation is questionable. Over-reliance can occur even where policies require approval. Test: reconstruct actual decisions, overrides, escalation, workload, and what evidence reviewers saw at the time.
Illustrative banking case: A reviewer signs off on a loan exception but cannot inspect the data source or disagreement behind the recommendation. Retain: decision-time inputs, reviewer guidance, escalation outcomes, override rationales, review capacity, and evidence of corrective action. A signature alone does not establish effective challenge.
The BIS’s 2024 annual analysis discusses the potential for AI-driven herding and procyclical market behavior, while the FSB identifies interdependencies as a possible threat to financial stability. Neither source establishes a measured prevalence of AI-caused systemic incidents across banks. [BIS, 2024] [FSB, 2024]
03 / EVIDENCE
What does the research actually establish?
The evidence supports concern and targeted testing. It does not justify claiming that interconnected AI has already caused a global banking crisis.
| Source | Supported finding | Important limitation |
|---|---|---|
| BIS Annual Economic Report, 2024 | AI can enhance financial services but may also encourage herding, common exposures, and rapid reactions during stress. | Discusses mechanisms and potential risks; it does not establish an observed AI-triggered banking crisis. |
| FSB, November 2024 | Highlights third-party concentration, market correlations, cyber risk, and model/data risk as financial-stability vulnerabilities. | Calls for better monitoring and information; estimates of economy-wide exposure remain incomplete. |
| BIS FSI Insights 63, December 2024 | Finds existing financial rules address many AI risks, while provider oversight, skills, data governance, and novel AI uses merit attention. | Its governance discussion is not a mandate to create a single global AI law. |
| EBA risk assessment, June 2026 | Identifies operational, AI, cyber, fraud, and ICT-dependency concerns in EU/EEA banking. | Sector-level operational-risk data are not equivalent to counts of AI-specific failures. |
| Yang, Liu, & Wang, 2026, peer-reviewed synthesis | Integrates research on market microstructure, algorithmic trading, and AI; identifies common models and infrastructure as plausible contributors to synchronized decisions. | Its review notes weak direct evidence for vendor commonality, autonomous agents, and equilibrium feedback. [Yang, Liu, & Wang, 2026] |
| FSB sound practices, October 2026 | Publishes 12 voluntary sound practices spanning enterprise governance and AI lifecycle management. | Sound practices are not legally binding obligations in every jurisdiction or a model certification. |
The FSB’s latest guidance is particularly significant. On October 15, 2026, it finalized Sound Practices for Responsible Adoption of Artificial Intelligence, giving financial institutions 12 practices that address governance, use-case risk, data, explainability, performance, oversight, cyber risk, and third-party dependencies. It encourages proportionate adoption and cross-border information-sharing; it does not replace domestic law. [FSB, 2026]
What is still uncertain? It is difficult to infer a common vendor, an identical decision rule, or a financial-stability feedback loop from the mere fact that several banks use AI. The 2026 review by Yang, Liu, and Wang finds stronger evidence for average market-quality effects than for vendor-commonality, agent autonomy, or equilibrium feedback. Those weaker areas justify targeted data collection and scenario analysis—not claims that the causal mechanisms are already established across the banking sector (Yang et al., 2026).
04 / IMPLEMENTATION BASELINE
How can banks apply the FSB’s 12 AI sound practices?
The FSB’s October 15, 2026 final report organizes 12 sound practices around enterprise governance and the AI lifecycle. They are voluntary international guidance, not a worldwide banking certification or a substitute for local law. Their practical value is in prompting a firm to connect a stated control to a responsible owner and verifiable output.
| Sound practices | Question for the bank | Evidence to retain |
|---|---|---|
| 1–4 · Governance and adaptability | Who sets AI risk appetite, who approves consequential use, and how are new capabilities and risks reassessed? | Board or committee decisions, responsibility map, AI inventory, risk appetite criteria, periodic governance reviews. |
| 5–7 · Materiality, selection, data | Does the selected system fit the business need, decision context, data rights, and materiality of the use? | Use-case assessment, supplier selection rationale, data lineage, restrictions, provenance exceptions. |
| 8–10 · Explainability, performance, people | Can reviewers understand limitations, challenge consequential decisions, and intervene effectively? | Validation tests, explanation assessments, human-review samples, override analyses, monitoring thresholds. |
| 11–12 · Cyber, ICT and suppliers | Can the institution detect misuse, withstand dependency failure, and manage concentration exposure? | Threat models, incident exercises, vendor assessments, resilience tests, fourth-party visibility, tested alternatives. |
Governance test: choose a material AI use case and follow its path from business approval to data, system output, human decision, monitoring, and recovery. Record gaps in the chain. A strong policy or supplier certificate may support one part of that chain, but it cannot establish that every part was evaluated.
05 / REGULATORY LANDSCAPE
How financial AI governance differs by jurisdiction in 2026
Rules differ in legal force, scope, definitions, implementation dates, and supervisory authority. A multinational institution should distinguish binding requirements from supervisory guidance and voluntary international practices.
European Union AI Act + DORA
The AI Act identifies certain creditworthiness and personal credit-scoring systems as high-risk under Annex III; AI used to detect financial fraud is expressly excluded from that particular Annex III category. The 2026 Digital Omnibus amendment moved application of the main Annex III high-risk requirements to December 2, 2027. Other AI Act duties may already apply. Separately, DORA has applied since January 17, 2025 to in-scope financial entities and addresses ICT resilience, incidents, testing, and third-party risk. [EU law] [DORA]
United States Revised model-risk guidance
On April 17, 2026, the Fed, OCC, and FDIC issued revised interagency model-risk guidance replacing SR 11-7. It recommends proportionate development, validation, monitoring, and governance of covered quantitative models, including relevant non-generative AI. Generative and agentic AI are expressly outside its scope. The agencies say the guidance itself is not a set of enforceable standards. Existing banking, consumer-protection, data, and security duties may still apply to AI uses outside its scope. [Federal Reserve / OCC]
United Kingdom PRA SS1/23
The PRA’s model-risk management statement, updated in April 2026, sets out five principles for in-scope banks: model identification, governance, development and use, independent validation, and mitigants. It includes AI-related modeling risks while recognizing proportionality. It should not be described as applying directly to every UK financial company. [PRA]
China Banking and insurance AI guidance
China’s National Financial Regulatory Administration published guidance in June 2026 on safe AI development and application in banking and insurance. It emphasizes compliant deployment, risk controls, classification, transparency, and trustworthy use. China cannot be reduced to a simple description of rapid or centralized rulemaking; applicability requires examination of the specific financial-sector measure and relevant data and technology regulations. [NFRA]
International frameworks Coordination, not a global law
The FSB’s 2026 sound practices, NIST AI RMF, and ISO/IEC 42001 can help organize governance and evidence across different business units and jurisdictions. They do not automatically satisfy local statutes or establish that a deployed AI system is safe under every condition. [NIST] [ISO]
06 / EXECUTION
What evidence should banks retain to support AI assurance?
Strong AI governance requires more than a policies folder or model inventory. The useful unit of evidence is an AI use-to-decision record that demonstrates who may rely on an AI output, for what purpose, under which conditions, and with which safeguards. The following structure is an InfoSecured.ai recommended practice—not a prescribed international regulatory form.
Minimum evidence record
- Use and materiality
- Business purpose, product, legal entities, downstream decisions, affected customers, business criticality
- Model and provider
- Version, owner, operator, upstream services, shared infrastructure, subcontractors, exit route
- Data and assumptions
- Lineage, validation period, restricted inputs, limitations, domain shifts, concentration exposure
- Testing and challenge
- Independent validation, stress scenarios, operational testing, reviewer findings, unresolved conditions
- Human decisions
- Approval authority, review evidence, overrides, escalation criteria, documented reliance rationale
- Monitoring and recovery
- Performance triggers, third-party alerts, incident procedures, fallback, rollback, retest schedule
Indicators that connect model risk to business outcomes
- Vendor concentration: share of material AI use cases that depend on each provider or underlying service.
- Fallback readiness: proportion of critical workflows with a tested alternate process and demonstrated recovery objective.
- Reliance exceptions: unresolved validation limitations or changed use conditions before deployment or reuse of a prior assessment.
- Customer outcomes: false declines, credit-decision errors, fairness indicators, fraud losses, and appeal or escalation performance.
- Decision correlation: common strategy, data, or factor exposure across material automated decisions where observable and relevant.
- Human oversight: review volume, override rates, turnaround, access to supporting evidence, and effectiveness of escalation.
Targets should be calibrated to the institution and use case; none of these indicators provides a universal pass/fail threshold. A low override rate alone does not show that reviewers are properly challenging the system. A high vendor-concentration score likewise establishes exposure, not inevitable failure.
Risk-to-evidence matrix: what should a reviewer actually inspect?
| Material risk | Control or test | Evidence and decision owner |
|---|---|---|
| Correlated model behavior | Test correlated decisions and shared factor shocks under relevant stress conditions. | Scenario assumptions, execution limits, results; treasury or market-risk owner. |
| Shared provider failure | Identify critical fourth parties, test outage and degraded operation, confirm fallback capacity. | Dependency map, test results, supplier notices; ICT resilience and business owner. |
| Unreliable data or labels | Check independent sources, provenance, validation windows and correction workflow. | Lineage, exception log, impacted-decision sample; data and model owners. |
| Human review failure | Sample contested decisions, validate authority, information, workload, escalation and overrides. | Review logs, sampled cases, challenge outcomes; operations and risk owner. |
| Unsupported reuse of a prior assessment | Compare original assessment scope, model version, data, use conditions, and known gaps with the current decision. | Assessment version, coverage comparison, unresolved limitations, reliance approval; designated reviewer. |
One control may serve several risks, but coverage must be demonstrated rather than assumed. A validated model may still depend on an untested supplier; a vendor’s certification may not cover the bank’s integration; and a prior assessment may become stale when the use case or data changes.
When can a bank rely on a prior AI assessment?
Reuse is defensible when the prior assessment addresses the same decision-relevant claim, its test conditions remain applicable, important dependencies have not changed, and open limitations have been evaluated. Before relying, a reviewer should identify what the earlier work actually tested, compare it with today’s model and workflow, verify any compensating controls, and decide whether additional testing or approval is necessary. Document the difference between evidence available and risk accepted: they are not equivalent.
The InfoSecured.ai evidence library provides related review resources, while the ISO/IEC 42001 implementation guide explains how AI management-system requirements relate to records and accountable decisions.
07 / THIRD-PARTY ASSURANCE
What should banks demand from AI providers and their supply chains?
Procurement evidence should describe the actual service and operating conditions, not simply the supplier’s organization. The bank remains responsible for evaluating how the purchased service is configured, combined with its own data, relied on by staff, and connected to consequential actions. The FSB’s Sound Practice 12 expressly emphasizes performance, transparency, data quality, supply chains, concentration, and business continuity (FSB, 2026).
Before purchase or material renewal, establish:
- Service identity
- AI capability, model or API version where available, hosting, subprocessors and material upstream dependencies.
- Use boundaries
- Authorized data, outputs, retention, human approval points, prohibited uses and customer impact.
- Assurance limits
- Independent tests, known failure conditions, security and data-control reports, inaccessible evidence, and gaps.
- Change rights
- Advance notification of material model or service changes, revalidation rights and incident cooperation.
- Recovery and exit
- Practical fallback, data portability, termination assistance, restoration objectives and documented test results.
When a supplier will not disclose training data, weights, or subcontractors: record the visibility limit and evaluate whether permitted black-box tests, contractual rights, monitoring, independent checks, or stricter use restrictions are sufficient for the proposed decision. Avoid treating withheld evidence as a positive assurance conclusion. For a deeper supplier-control analysis, see ISO/IEC 42001 and third-party AI accountability.
08 / WORKED SCENARIO
Illustrative scenario: when one AI dependency affects several banks
HYPOTHETICAL EXAMPLE — NOT A REPORTED INCIDENT
Three banks use different customer-facing lending applications but rely on the same external identity and risk-enrichment service. An upstream data error affects the provider’s output. Each bank’s credit model remains technically available, yet the shared feature is wrong. Loan decisions may become distorted before the issue is recognized.
Identify the shared signal
Compare outcome changes, data quality alerts, provider notices, and model-performance drift. Determine which affected decisions actually used the defective field.
Change the workflow safely
Pause or isolate affected automated decisions, activate alternative checks, preserve impacted records, and review customer remediation obligations.
Revalidate reliance
Confirm data correction, retest the complete decision pipeline, independently review resulting outputs, and document approval before restoring full automation.
One bank may contain the problem through its own controls. But whether the event creates sector-wide risk depends on how many institutions share the dependency, how material the affected decisions are, and whether losses or responses transmit through markets or customers. A regulator would need broader data to establish that aggregation. This separates organizational resilience from systemic-risk measurement.
What would an assurance decision look like?
Suppose Bank A finds that the faulty feature contributed to 120 applicant reviews and cannot yet identify which were wrongly declined. The quantity is hypothetical, not a study result. Rather than declaring that its credit model “passed validation,” the decision forum records: (1) affected use and time window; (2) uncertain data provenance; (3) stop or fallback rule; (4) sample and remediation plan; (5) named approver; and (6) conditions for renewed automated reliance. If Bank B has a different fallback, its business outcome may differ despite sharing the supplier.
This example shows why a firm-level finding is not proof of a financial-system outcome. Supervisors would need cross-firm exposure, severity, interconnection, and transmission evidence to assess broader stability implications.
09 / PRIORITIES
A 90-day AI governance plan for banking leaders
Identify material decisions and accountability
Owners: business executive, model risk management, CISO/ICT risk, procurement, compliance, and relevant data stewards.
Work: inventory material predictive, generative, and agentic uses separately; map each to the decision it changes and the legal entity that owns that decision. Identify provider and fourth-party dependencies, customer consequences, review authority, applicable rules, and whether a current validation record exists. Select two or three high-consequence workflows for challenge.
Gate: signed priority inventory, use-and-materiality rationale, jurisdiction map, evidence gaps, accountable risk owner, and an agreed fallback for each selected workflow.
Challenge decisions and dependencies
Owners: independent validation, business operations, ICT resilience, vendor risk, and data engineering.
Work: review prior-assessment sufficiency; test common-data corruption, service interruption and correlated decisions where relevant; sample human approvals and customer outcomes; confirm that agreed supplier testing reflects the bank’s integration. Define test boundaries and avoid disruption of live markets or customer accounts.
Gate: reproducible results with versions, assumptions, affected decisions, performance and recovery impact, independent challenge, and open findings assigned to owners.
Decide, monitor and retest
Owners: business risk owner and authorized approval forum, with model risk, cybersecurity and compliance challenge.
Work: remediate or restrict serious weaknesses, rehearse rollback and manual alternatives, approve monitoring indicators, assign supplier-change triggers, and assess concentration at the portfolio level. Escalate unverified dependencies instead of marking them complete.
Gate: signed reliance/acceptance decision, testable safeguards, monitored indicators, recovery record, unresolved exceptions, and retest dates.
Use the FSB’s 12 sound practices to challenge coverage across governance and the lifecycle. Use NIST AI RMF or ISO/IEC 42001 for management-system organization where appropriate. Then map controls to the bank’s actual supervisory, contractual, and legal requirements rather than treating a voluntary framework as compliance by itself.
The day-90 result is an assurance dossier for the scoped uses—not a statement that the bank or financial system is universally safe. Maintain a record of what was assessed, what remains unverified, who approved continued reliance, and which changes require renewed review.
10 / QUESTIONS
Frequently asked questions
What is systemic AI risk in banking?
It is the possibility that AI-related failures or coordinated behavior spread beyond one institution through common providers, data, market actions, or other dependencies, potentially impairing financial stability. The existence of shared systems alone does not establish an actual systemic event.
Can a bank be compliant and still have AI risk?
Yes. Compliance demonstrates conformity with identified requirements; it does not guarantee that a system has been tested against every plausible failure or that shared providers and market dependencies have been evaluated adequately.
Does the EU AI Act classify all banking AI as high risk?
No. Certain personal creditworthiness and credit-scoring uses are listed in Annex III. AI systems used to detect financial fraud are expressly excluded from that particular category. Classification requires evaluating the system’s purpose and any other applicable provisions.
Does the U.S. April 2026 model-risk guidance cover generative AI?
No. The updated interagency guidance explicitly excludes generative and agentic AI models. Banks should nonetheless address their risks under applicable law and their wider governance, security, and risk-management practices.
Are the FSB’s October 2026 AI sound practices mandatory?
No. The 12 sound practices are international guidance to support proportionate governance, risk management, and cooperation. Whether an institution has binding obligations depends on its applicable jurisdiction and regulations.
What should a bank’s board ask about AI risk?
Ask which decisions materially rely on AI; which providers or data are shared; what was independently tested; where human oversight is consequential; what would happen if the system or supplier failed; and which unresolved limitations the institution is accepting.
What is the difference between model risk and systemic AI risk?
Model risk concerns errors, limits or inappropriate use of a particular model. Systemic AI risk concerns potential disruption that spreads through institutions or markets, for example through shared providers or correlated automated behavior. Neither a model failure nor shared use alone establishes an actual systemic event.
Can an ISO/IEC 42001 certificate prove that a bank’s AI decision is safe?
No. Certification can provide evidence about the defined scope of an AI management system. It does not by itself validate a particular model, supplier integration, credit outcome, human approval, or the bank’s exposure to common provider failures. Those need use-specific evidence.
What should a bank do if a vendor will not provide full AI model information?
Document the missing information and decide whether limited-scope testing, independent checks, contract protections, or restricted use can reasonably address the risk. If critical assumptions cannot be verified, escalate or defer a consequential use rather than treating absence of evidence as a successful assessment.
BOTTOM LINE
Assess the decision, the dependency, and the institution—not only the model.
Financial AI governance is maturing from isolated model review toward operationally grounded assurance and better monitoring of shared risks. The 2026 FSB sound practices provide a useful international reference, but regulators and financial firms still need dependable evidence about real deployments and connected exposures. The goal is not uniform paperwork; it is credible risk decisions and systems that remain accountable when assumptions fail.
References and further reading
This guide distinguishes published research, regulatory and supervisory materials, and InfoSecured.ai practitioner recommendations. Worked examples are hypothetical; suggested evidence records are not prescribed universal regulatory forms. Check legal scope and current requirements for each jurisdiction.
- Bank for International Settlements. (2024). Annual Economic Report 2024, discussion of artificial intelligence and financial stability. BIS report.
- Financial Stability Board. (2024, November 14). The financial stability implications of artificial intelligence. FSB report.
- Crisanto, J. C., Leuterio, C. B., Prenio, J., & Yong, J. (2024, December 12). Regulating AI in the financial sector: Recent developments and main challenges (FSI Insights 63). BIS FSI.
- Financial Stability Board. (2026, October 15). Sound practices for responsible adoption of artificial intelligence (AI): Final report. FSB guidance.
- Yang, W., Liu, H., & Wang, Y. (2026, September 24). From algorithmic trading to generative AI: Artificial intelligence, systemic risk, and the transformation of financial markets. Journal of Economic Surveys (online first). https://doi.org/10.1111/joes.70167.
- European Banking Authority. (2026, June). Risk assessment report, discussion of operational risks and AI adoption. EBA report.
- European Union. (2024, as amended in 2026). Regulation (EU) 2024/1689: Artificial Intelligence Act, especially Annex III and Article 113, as amended by Regulation (EU) 2026/1744. Consolidated text.
- European Union. (2022). Regulation (EU) 2022/2554: Digital Operational Resilience Act. Official text.
- Board of Governors of the Federal Reserve System, Office of the Comptroller of the Currency, & Federal Deposit Insurance Corporation. (2026, April 17). Revised guidance on model risk management (SR 26-2 / OCC Bulletin 2026-13). Interagency guidance.
- Prudential Regulation Authority. (2026, April 23). SS1/23: Model risk management principles for banks. Bank of England.
- National Financial Regulatory Administration of China. (2026, June 18). Guiding opinions on the safe development and application of artificial intelligence in banking and insurance [Chinese-language regulatory guidance; 金发〔2026〕8号]. Official document.
- National Institute of Standards and Technology. (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0). NIST resource.
- International Organization for Standardization. (2023). ISO/IEC 42001:2023—Artificial intelligence management systems. ISO overview.