Actions

AI Risk Management

AI risk management is the continuous organizational process of identifying, assessing, treating, monitoring, and communicating risks arising from the development, acquisition, deployment, use, and retirement of artificial intelligence (AI) systems. It considers how an AI system may affect people, organizations, and society, and how those effects can be prevented, reduced, accepted within defined limits, or addressed when they occur.[1]

In an enterprise, AI risk management covers the whole deployed system: its data, models, interfaces, connected tools, operating environment, human decisions, third-party dependencies, and the consequences of its outputs. A technically accurate model can still create unacceptable risk if its recommendations are used in the wrong context, sensitive information is exposed, or automated actions exceed authorized limits. Risk therefore depends on both the AI technology and the purpose and conditions of its use.

AI risk management is an operational capability within AI governance and enterprise risk management. Governance determines who has authority and accountability; risk management supplies the assessments, controls, monitoring, and evidence needed to exercise that authority. AI security addresses important security threats but does not cover the full range of AI risks, including unfair treatment, reliability, safety, privacy, and harmful organizational or societal impacts.

Purpose and scope

The purpose of AI risk management is to support informed decisions about whether, where, and under what conditions AI systems may be used. It seeks to identify foreseeable harm early, make uncertainty visible, prioritize risk treatment, and establish a basis for continued operation or intervention. Risks may affect system users, people subject to AI-assisted decisions, employees, customers, affected communities, the deploying organization, and third parties.

Its scope ordinarily includes AI built internally, purchased from vendors, embedded in enterprise software, or accessed through external services. A single AI application may involve multiple parties with different responsibilities: model developers, application providers, enterprise deployers, users, and oversight functions. The extent of control and available evidence differs across those parties, so an organization should distinguish risks it can directly mitigate from risks it must manage through supplier assurances, contractual commitments, use restrictions, or alternative designs.

Risk management occurs before deployment and throughout operation. New data, prompts, system integrations, populations, model versions, user behavior, and external threats can change the risk of an otherwise unchanged application.

Main categories of AI risk

AI risks overlap and may compound. Common categories include:

Risk category Typical manifestation Principal concern
Reliability and performance Incorrect predictions, fabricated outputs, performance degradation, or failure outside tested conditions Decisions or services become inaccurate or inconsistent
Fairness and fundamental rights Unequal error rates, discriminatory outcomes, inaccessible services, or harmful treatment of affected groups Individuals or groups experience unjustified adverse effects
Privacy and data protection Exposure of personal information, inappropriate processing, memorization, or leakage through outputs Personal data and privacy interests are compromised
Security and misuse prompt injection, data poisoning, model theft, adversarial manipulation, or abuse of connected tools Attackers or unauthorized users change behavior or obtain access
Safety and human oversight Unsafe recommendations, autonomous actions, automation bias, or ineffective intervention mechanisms People or physical environments are harmed
Transparency and accountability Unclear system limitations, missing provenance, absent decision records, or no identified owner Errors cannot be explained, challenged, or corrected effectively
Legal, contractual, and third-party exposure Unlawful uses, intellectual-property disputes, supplier changes, or insufficient vendor evidence Compliance, rights, and operational dependencies are inadequately controlled
Operational and strategic exposure Excessive dependence, escalating cost, unavailable services, or weak contingency arrangements Business processes become fragile or fail to deliver intended value

These categories are an explanatory grouping, not an exhaustive or universally mandated taxonomy. Some harms have uncertain likelihood or cannot be adequately represented by a single numerical risk score. Risk judgments must therefore account for affected populations, severity, exposure, reversibility, and the quality of available evidence.[1]

How AI risk management works

Although organizations use different methods, an effective process has several recurring functions.

Establish context and accountability

The organization defines the system's intended purpose, users, decision authority, operating boundaries, relevant legal obligations, and acceptable risk conditions. This requires a meaningful inventory of AI systems and clear ownership across business, technology, risk, legal, security, privacy, and other affected functions. Higher-impact applications ordinarily warrant more independent scrutiny and stronger evidence.

Identify and assess risks

Risk identification examines the AI's intended use, reasonably foreseeable misuse, data, people affected, failure modes, dependencies, and downstream actions. Assessment combines technical testing with contextual review. Depending on the use, evidence may include representative performance tests, subgroup analysis, adversarial testing, safety evaluation, privacy assessment, supplier documentation, and the consequences of incorrect decisions. An AI impact assessment can help examine effects on people and fundamental rights; it is related to, but not identical to, the organization's broader risk assessment.

Select and verify risk treatments

Possible treatments include changing or limiting a use case, redesigning data or workflows, adding access controls and monitoring, requiring human authorization for consequential actions, providing appeal or correction mechanisms, reducing system autonomy, or deciding not to deploy. Controls must be tested for the actual operating context. A nominal human reviewer does not by itself provide effective oversight if the reviewer lacks time, information, competence, or authority to intervene.

Monitor, respond, and reassess

After deployment, organizations track performance, incidents, unexpected uses, changes in affected populations, supplier updates, and control effectiveness. Material changes may require renewed evaluation or suspension. AI incident management handles observed harmful events and near misses, while ongoing risk management also addresses hazards that have not yet produced an incident. Monitoring should include the system's effects on real-world decisions and outcomes, not merely model-level accuracy.

Controls and management evidence

An AI risk-management capability typically produces and maintains evidence such as a system inventory, use-case description, risk assessment, documented control decisions, predeployment evaluation results, deployment authorization, monitoring records, incident records, change history, and residual-risk decisions. These records make the organization's reasoning and accountability reviewable.

Different controls address different causal paths. For example, retrieval restrictions can reduce exposure to unapproved source material, while independent output checks may address factual errors. Neither establishes that a decision is fair in its actual setting. For a system able to call external tools, permissions, execution limits, logs, and emergency interruption mechanisms may be necessary because harm can arise from actions as well as generated text.

Risk treatment also has limits. Testing can miss rare conditions or emerging behavior; a vendor may restrict access to model internals; and some social effects cannot be resolved solely through technical controls. An acceptable residual-risk decision must identify the accountable authority and the evidence and conditions supporting continued use, rather than presume that completion of a checklist establishes safety.

Roles and responsibilities

The business or use-case owner is generally accountable for the intended purpose, resulting decisions, and ongoing suitability of the AI-enabled process. Technical teams assess model and application behavior, data teams examine data quality and lineage, and information security and privacy functions assess their respective risks. Legal and compliance functions interpret applicable obligations, while independent risk or audit functions may challenge assessments and test whether controls work as represented.

Executive leadership and designated governance bodies set risk appetite, escalation thresholds, and exceptions. The CIO commonly has a coordinating role for enterprise technology dependencies and controls, but responsibility cannot be assumed to rest solely with IT: consequential AI decisions require ownership by the functions using them. Responsibilities should be allocated explicitly for internally developed systems and externally supplied systems alike.

Standards and regulatory context

Several prominent references address different layers of AI risk management:

  • NIST AI Risk Management Framework (AI RMF 1.0). The voluntary framework organizes activities under Govern, Map, Measure, and Manage. Governance is cross-cutting rather than a one-time first step. NIST published version 1.0 in January 2023 and stated in 2026 that it was under revision.[1][2]
  • ISO/IEC 23894:2023. This international standard provides guidance for integrating AI-related risk management into the work of organizations that develop, provide, deploy, or use AI systems. It is guidance rather than an AI management-system certification standard.[3]
  • ISO/IEC 42001:2023. This standard specifies requirements for establishing, maintaining, and continually improving an AI management system. It provides an organizational management-system structure within which AI risk-management activities can operate.[4]
  • NIST Generative AI Profile (AI 600-1). Published in July 2024 as a companion to AI RMF 1.0, it addresses risks and actions specific to generative AI.[5]
  • European Union AI Act. Article 9 requires a documented and maintained risk-management system for systems classified as high-risk under the Act, with lifecycle review, foreseeable-risk evaluation, risk treatment, and testing. Applicability depends on the system's classification, the actor's role, and the relevant implementation dates; it is not a universal obligation for every AI use.[6]

These sources serve different purposes. A risk framework describes outcomes and practices; a management-system standard specifies organizational requirements; a law imposes obligations within its jurisdiction and defined scope. Adopting a voluntary framework does not, by itself, establish legal compliance or prove the safety of a particular system.

How it differs from related disciplines

AI governance sets policies, decision rights, oversight structures, and accountability for AI. AI risk management evaluates exposures and implements or informs controls and risk decisions within that governance system.

AI security focuses on malicious and accidental compromise of AI systems, assets, and interactions. It is a component of AI risk management, but broader risks also include bias, misuse in consequential decisions, unsafe reliance, and adverse social effects.

Responsible AI describes principles and practices for developing and using AI in ways consistent with values such as fairness, transparency, accountability, and human well-being. AI risk management is one mechanism for turning relevant principles into assessed risks and operational controls.

Model risk management has established practices for validating models and controlling model-related errors, especially in financial and quantitative settings. AI risk management overlaps with these practices but also considers wider application workflows, autonomy, affected parties, and societal consequences.

Illustrative enterprise application

Consider an organization deploying an AI system to assist with employee recruitment. Performance testing might show useful matching accuracy, yet the organization still needs to examine whether training data and outputs disadvantage particular groups, whether applicants can contest decisions, whether sensitive data is protected, and whether human reviewers can meaningfully override recommendations. The same model used only to summarize public job descriptions would have a different exposure and control profile. The distinction is the use and consequence, not the model label alone.

This example illustrates why AI risk management follows the deployed system, the decision it influences, the people it affects, and the conditions under which it operates.

See also

References

  1. ↑ 1.0 1.1 1.2 National Institute of Standards and Technology, Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1, January 2023, official publication.
  2. ↑ National Institute of Standards and Technology, AI Risk Management Framework, official status page, accessed 18 November 2026.
  3. ↑ International Organization for Standardization, ISO/IEC 23894:2023 — Information technology — Artificial intelligence — Guidance on risk management, official standard page, accessed 18 November 2026.
  4. ↑ International Organization for Standardization, ISO/IEC 42001:2023 — Information technology — Artificial intelligence — Management system, official standard page, accessed 18 November 2026.
  5. ↑ National Institute of Standards and Technology, Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, NIST AI 600-1, July 2024, official publication.
  6. ↑ European Union, Regulation (EU) 2024/1689, consolidated version of 27 July 2026, Article 9, official consolidated text.