AI risk management in practice: ISO/IEC 42001 and Oracle AI Agent Studio for Fusion Applications

Responsible AI principles may remain abstract unless they are translated into documented processes, defined controls, and ongoing evaluation. AI-related failures can create operational and reputational risks, including loss of trust. As with cybersecurity incidents, failures in AI systems or risk-management processes can damage trust and reputation, even when they do not result in regulatory consequences.

The parallel with security is instructive. The CEO Report on Cyber Resilience, a study by Saïd Business School at the University of Oxford and ISTARI, incorporates insights from interviews with 37 CEOs of large global enterprises, including executives who had experienced serious cyberattacks. The report examines how organizations can build cyber resilience through a practical playbook centered on anticipating, withstanding, responding to, and recovering from cyberattacks. By analogy, a similarly iterative operational approach can be useful in AI risk management, where deployment contexts can change, models may be updated, and controls may need to evolve over time to support resilience as conditions change and unexpected outcomes arise.

Another important parallel is that effective risk management does not depend solely on deploying the latest available AI models, just as cybersecurity is not achieved simply by using the strongest cryptographic algorithms. In both cases, resilience depends on a defense-in-depth approach: the operating environment around the technology, including people, processes, technical controls, governance practices, the capacity to adapt over time. In previous posts, we discussed Oracle’s ISO/IEC 42001:2023 certification and related operational practices. Building on that discussion, this post focuses on a repeatable approach to AI risk management.

AI risk management in context

An AI system’s intended purpose, configuration, users, and deployment context can affect its risk profile. Applicable requirements can also vary by use case and deployment context.

Several frameworks and standards provide concepts that organizations may consider as part of their own risk-management processes. The appropriate framework and implementation depend on each organization’s use case, risk profile, and applicable requirements. The NIST AI Risk Management Framework (RMF) 1.0 organizes AI risk-management activities into four functions: Govern, Map, Measure, and Manage. In the framework, these functions address topics including accountability, system context, evaluation, mitigation, and ongoing monitoring.

Oracle also describes relevant AI management practices in the Oracle Artificial Intelligence Terms (PDF), which reference ISO/IEC 42001 and the NIST AI Risk Management Framework. Oracle’s broader engagement with AI standards and industry frameworks is described on its AI Standards page.

From an initial evaluation to ongoing risk management

As with ISO/IEC 27001, risk assessment plays a central role in the Plan phase of ISO/IEC 42001 and the broader Plan-Do-Check-Act (PDCA) cycle. As risks and applicable requirements evolve, the cycle supports ongoing evaluation and continual improvement of the AI management system, including its suitability, adequacy and effectiveness.

ISO/IEC 42001 also calls for assessing the potential impacts of AI systems on individuals, groups, and society, complementing the organization’s AI risk assessment and risk-treatment processes.

Post-deployment monitoring is useful to provide information that informs reassessment and corrective action throughout the system lifecycle and consider residual risk and record relevant review and acceptance decisions. Appropriate governance and approval processes depend on each organization, its use case, and applicable requirements.

Oracle AI system development practices and customer responsibilities

Oracle has achieved ISO/IEC 42001:2023 certification for its AI services portfolio, within the scope identified in the applicable certification, including AI risk-management processes. Oracle’s development practices and customers’ responsibilities for their use of AI Functionality address different aspects of AI deployment and use.

Oracle AI system development practices. The publicly available Oracle Artificial Intelligence Terms describe practices adopted for reviewing AI Systems, addressing identified risks, reassessing systems after material changes, and tracking reviewed and deployed AI Systems. See the Oracle Artificial Intelligence Terms, §§3.1–3.4 (PDF). Separately, Oracle publishes information about its supply chain security and assurance practices, see Oracle Supply Chain Security and Assurance.

These measures do not, by themselves, mean that a customer’s use case complies with applicable laws or other requirements, or that it is responsible, safe, fair, or free of bias. The Oracle Artificial Intelligence Terms separately describe customer responsibilities for AI Functionality and Output. Among other things, customers are responsible for how they use AI Functionality and Output, review and independently verify Output before use, and comply with applicable laws. See the Oracle Artificial Intelligence Terms, §§4.1 and 5.1–5.2 (PDF).

Monitoring, evaluation, and control capabilities in Oracle AI Agent Studio for Fusion Applications: Oracle AI Agent Studio for Fusion Applications supports multiple LLMs and provides customers with product capabilities including guardrails, prompt management, policy models, human oversight, agent testing, monitoring and tracing, which help users investigate agent behavior, latency measurements and consumption metrics.

Customers remain responsible for determining what governance and risk-management processes and documentation are appropriate for their use cases, risk profiles, and applicable requirements. Information generated through monitoring, evaluation, and tracing in Oracle AI Agent Studio for Fusion Applications can provide inputs that customers may use in their own review, change-management, and risk-management processes.

As one example, and depending on an organization’s own use case, risk profile, and applicable requirements, an organization might choose to maintain an agents’ inventory containing information such as owner, purpose, data sources, actions, permissions, and status. The appropriate information and documentation will vary by organization and use case.

Conclusion

AI risk-management processes can help identify and address risks that may contribute to unfair or discriminatory outcomes. Such risks can arise from multiple factors, including data, model design, evaluation choices, deployment context, system configuration, and the way outputs are used. If not appropriately addressed, these risks may contribute to harmful outcomes and reduce trust in AI technology.

Oracle’s AI system development practices are complemented by capabilities in Oracle AI Agent Studio for Fusion Applications that can contribute to an enterprise AI harness and that customers may use as part of their own monitoring, governance, and risk-management processes. The appropriate combination of technical controls, organizational processes, and governance practices depends on each customer’s use case, risk profile, and applicable requirements.