AI risk assessment

A risk-management process for identifying, estimating, and prioritizing risks arising from the operation and use of an AI system, incorporating threat and vulnerability analyses and considering mitigations provided by controls planned or in place.

Source: AIEOG AI Lexicon (Feb 2026), adapted from CSRC risk assessment glossary and NIST AI 100-1

An AI risk assessment is a structured process for understanding what could go wrong with an AI system and how significant the consequences would be. It identifies the specific risks an AI system introduces, estimates how likely and how severe those risks are, and prioritizes them so the organization can allocate resources to the most important ones.

This process goes beyond standard technology risk assessment. AI systems introduce unique risks that traditional IT risk frameworks may not fully capture: bias and fairness concerns, drift and degradation, explainability challenges, data dependency risks, and adversarial vulnerability.

A thorough AI risk assessment examines:

• Threat analysis. Who or what could cause the AI system to fail or be misused? This includes external adversaries, data quality issues, operational errors, and the natural degradation of model performance over time.
• Vulnerability analysis. Where is the AI system most susceptible to failure? This could be in the training data, the model architecture, the deployment environment, the integration points, or the human oversight process.
• Impact analysis. What are the consequences if the risk materializes? This includes financial loss, regulatory exposure, customer harm, reputational damage, and operational disruption.
• Control assessment. What controls are already in place, and how effective are they? This includes both preventive controls (designed to stop the risk from occurring) and detective controls (designed to identify when the risk has materialized).

Risk assessment is foundational to financial services compliance. Regulators expect institutions to conduct regular, documented risk assessments across all material risk categories. As AI becomes a material risk, AI-specific risk assessment becomes a regulatory expectation.

The NIST AI Risk Management Framework positions risk assessment as a core function within its “Map” and “Measure” components. The OCC, FDIC, and Federal Reserve have each signaled that AI risk should be incorporated into existing risk management frameworks — a theme explored in depth in the Treasury’s new AI guidance for banks. The AIEOG Lexicon codifying this term underscores that the Treasury views AI risk assessment as a standard practice, not an optional exercise.

Common challenges institutions face with AI risk assessment:

• Underestimating indirect risks. AI risk is often assessed narrowly (“will the model be accurate?”) rather than broadly (“what happens to downstream processes, customer outcomes, and regulatory standing if the model fails?”).
• Static assessments. AI risk is dynamic. A risk assessment conducted at deployment may not reflect the risk profile six months later. Institutions need to reassess periodically.
• Incomplete scope. Risk assessments may focus on the model itself while overlooking risks in the surrounding system: data pipelines, integration points, human override processes, and vendor dependencies.
• Lack of quantification. Many AI risk assessments are qualitative only. While qualitative assessment is better than none, regulators increasingly expect quantitative risk metrics where possible.

1. Conduct AI risk assessments for every AI use case. Every entry in your AI use case inventory should have a corresponding risk assessment, proportionate to the risk tier of the use case.
2. Use a structured methodology. Adopt a documented methodology that covers threat identification, vulnerability analysis, impact assessment, and control evaluation. The NIST AI RMF provides a strong starting framework.
3. Assess risk at multiple stages. Conduct initial risk assessments during the design phase, update them before deployment, and reassess periodically during operation.
4. Include stakeholders from across the organization. AI risk assessment should involve compliance, risk, technology, business line owners, and legal. No single function has the full picture.
5. Document everything. Risk assessments should be fully documented with methodology, findings, risk ratings, control assessments, and remediation plans. This documentation will be requested during examinations.
6. Connect findings to action. Risk assessment findings should drive tangible actions: new controls, enhanced monitoring, model modifications, or use case restrictions. Assessment without action is a governance gap. For a structured approach, see how the 12 pillars of a CMS connect risk assessment to broader compliance operations.
7. Report to senior management and the board. Aggregate AI risk assessment findings should be reported to risk committees and boards as part of the institution’s overall risk posture.

AI governance, Model risk, AI lifecycle, Documentation, Bias, Adversarial AI, Third-party AI risk, Performance monitoring