The Australian Prudential Regulation Authority (APRA) has fundamentally reshaped how regulated entities must approach operational risk management through Prudential Standard CPS 230. As artificial intelligence becomes increasingly embedded in financial services operations, understanding how CPS 230 applies to AI systems has become critical for banks, insurers, and superannuation trustees.

Understanding CPS 230: The Foundation

CPS 230, which came into effect in July 2023 with full compliance required by July 2025, represents APRA's most comprehensive operational risk management standard to date. The standard requires APRA-regulated entities to maintain operational resilience by managing risks that could disrupt their business operations or cause harm to beneficiaries.

The standard is built on several key pillars: robust risk management frameworks, clear accountability through the Board and senior management, comprehensive business continuity planning, and third-party risk management. What makes CPS 230 particularly relevant to AI is its focus on material risks - and AI systems increasingly fall into this category.

Why AI Falls Under CPS 230's Scope

AI systems present unique operational risks that trigger CPS 230 requirements. These systems often process critical data, make or influence decisions affecting customers, and integrate deeply into core business processes. When AI is used for credit decisioning, fraud detection, customer service, claims processing, or investment management, it typically qualifies as a material business activity under the standard.

The characteristics that make AI powerful - its complexity, opacity, and autonomous decision-making capabilities-also make it a source of operational risk. Model drift, data quality issues, algorithmic bias, and unexpected behaviors in production environments can all lead to material disruptions or customer harm.

Key CPS 230 Requirements for AI Systems

Role of the Board and Senior Management

CPS 230 places ultimate accountability with the Board, which must maintain oversight of material risks including those arising from AI. Boards must understand where AI is deployed, what risks it creates, and how those risks are managed. This requires AI literacy at the governance level - not technical expertise, but sufficient understanding to ask the right questions and challenge management.

Senior management must implement the Board-approved operational risk management framework and ensure it encompasses AI-specific risks. This includes establishing clear ownership for AI systems, defining risk appetites for AI deployment, and ensuring adequate resources for AI risk management.

Risk Management Framework

The operational risk management framework must identify, assess, and manage material risks from AI systems throughout their lifecycle. This means moving beyond traditional IT risk frameworks to address AI-specific concerns.

Risk identification should cover the full spectrum of AI risks: model performance degradation, data quality and availability issues, algorithmic bias and fairness concerns, explainability limitations, cybersecurity vulnerabilities specific to AI systems, and dependencies on third-party AI providers. The framework must also address risks from shadow AI - unsanctioned use of AI tools by employees.

Risk assessment for AI requires specialized approaches. Traditional risk assessment methods may not capture the dynamic nature of AI systems that can change behavior as they process new data. Entities need processes to evaluate AI model performance continuously, not just at deployment.

Business Continuity Management

CPS 230 requires comprehensive business continuity planning for material business activities. For AI-dependent operations, this creates specific obligations. Entities must identify single points of failure in AI systems, maintain fallback procedures when AI systems fail, and ensure human oversight can intervene when necessary.

Consider a bank using AI for fraud detection. The business continuity plan must address scenarios where the AI system becomes unavailable, produces unreliable outputs, or fails to detect fraud effectively. Manual procedures, alternative detection systems, or degraded service modes must be documented and tested.

Control Environment

The control environment must be designed to prevent, detect, and respond to disruptions. For AI systems, this translates into several practical requirements.

Model governance frameworks should define standards for AI development, testing, validation, and deployment. Change management processes must account for the fact that AI models may need frequent retraining or updates. Monitoring systems should track AI performance in real-time, detecting anomalies, bias, or drift.

Documentation requirements are particularly important for AI. Given the "black box" nature of many AI models, comprehensive documentation becomes essential for meeting CPS 230's transparency expectations. This includes model cards describing AI system capabilities and limitations, data lineage and quality controls, validation and testing results, and decisions made by AI systems that affect customers.

Third-Party AI Risk Management

CPS 230 includes heightened requirements for managing material service providers. In the AI context, this captures cloud AI services, third-party AI models and platforms, data providers for AI training, and AI development vendors.

Due diligence must extend beyond traditional vendor assessment to address AI-specific concerns. Can the provider explain how their AI models work? What data was used to train them? How do they address bias? What happens if the service is withdrawn? How is model performance monitored and maintained?

Contractual arrangements should establish clear service levels for AI performance, rights to audit AI systems and their outputs, notification requirements for material changes to AI models, and exit provisions including data portability.

The challenge with many AI services is that providers may be reluctant to disclose model details for proprietary reasons. CPS 230 doesn't necessarily require full transparency, but entities must be able to demonstrate they understand the risks and have appropriate controls, even if they don't understand every technical detail of the AI.

Testing and Assurance

CPS 230 requires testing of critical operations, controls, and response plans. For AI systems, this creates specific testing obligations that go beyond traditional IT testing.

Model validation should occur before deployment and periodically thereafter, covering technical validation of model accuracy and performance, assessment of bias across different customer segments, stress testing under various scenarios, and validation that the AI operates within defined parameters.

Scenario testing should explore realistic failure modes. What happens if data quality degrades? What if the AI encounters input data significantly different from its training data? How does the system behave in adversarial scenarios?

Regular audits should assess AI governance frameworks, compliance with internal AI policies, effectiveness of AI risk controls, and accuracy of AI risk reporting to the Board.

Data Quality and Management

While not explicitly an AI requirement, CPS 230's focus on operational resilience makes data quality critical for AI systems. Poor data quality can lead to inaccurate AI outputs, biased decisions, and model failure - all operational risks that could cause material disruption or harm.

Entities should establish data governance frameworks covering AI training data and operational data. This includes data quality standards and monitoring, data lineage tracking, processes for identifying and correcting data issues, and controls over data access and security.

For AI models, data drift - where the statistical properties of input data change over time - is a particular concern. Monitoring systems should detect when operational data diverges significantly from training data, triggering review and potential model retraining.

Reporting and Notification

CPS 230 requires entities to notify APRA of material operational risk incidents and disruptions. AI-related incidents that could trigger notification obligations include significant AI errors affecting customers, AI bias leading to unfair outcomes at scale, AI system failures disrupting critical operations, and cybersecurity breaches involving AI systems or data.

Internal reporting to the Board should provide regular updates on AI deployment, associated risks, and control effectiveness. This shouldn't be purely technical reporting - it should enable Board-level understanding of how AI is changing the entity's risk profile.

Practical Implementation Challenges

Implementing CPS 230 for AI presents several challenges. Many entities lack comprehensive AI inventories—they don't know all the places AI is being used across their organization. Creating this inventory is often the first step toward compliance.

AI literacy gaps exist at all levels. Boards need sufficient understanding to provide oversight, risk managers need to understand AI-specific risks, and business units need to understand their accountability for AI systems they deploy.

The pace of AI innovation creates compliance challenges. CPS 230 requires stable, controlled environments, but AI capabilities and use cases evolve rapidly. Entities must balance innovation with prudent risk management.

Legacy governance structures may not fit AI well. Traditional three-lines-of-defense models must adapt to address AI's unique characteristics, including its cross-functional nature and the need for specialized technical expertise.

Regulatory Expectations and Guidance

While CPS 230 doesn't explicitly mention AI, APRA has indicated through speeches and guidance that it expects the standard to be applied to AI systems. APRA's Information Paper on Managing the Impact of Automation and AI provides additional context on regulatory expectations.

APRA expects entities to take a proportionate approach based on the materiality of AI use. A simple chatbot for basic customer queries requires less governance than an AI system making credit decisions or managing investment portfolios.

However, APRA has also indicated it will be skeptical of entities claiming AI risks are immaterial when AI is embedded in critical operations. The onus is on entities to demonstrate through their risk assessments why particular AI uses don't require full CPS 230 governance.

Building an AI Governance Framework Under CPS 230

A CPS 230-compliant AI governance framework should include several components working together.

An AI steering committee or equivalent governance body should provide oversight of AI deployment across the entity, approve material AI use cases, and escalate significant AI risks to the Board.

AI risk appetite statements should define acceptable levels and types of AI risk, establish boundaries for AI deployment, and guide decision-making about AI adoption.

AI lifecycle management processes should cover development or procurement, validation and testing, deployment and monitoring, ongoing maintenance, and decommissioning.

Specialized AI roles may be necessary, including AI ethics officers, model risk managers, AI auditors, and designated accountable individuals for critical AI systems.

The Path Forward

CPS 230 compliance for AI isn't a one-time exercise but an ongoing commitment. As AI capabilities expand and adoption increases, operational risk management frameworks must evolve accordingly.

Entities should approach CPS 230 AI compliance pragmatically. Start with a comprehensive AI inventory, assess which AI systems are material under CPS 230, implement governance frameworks proportionate to risk, build necessary capabilities and expertise, and establish monitoring and reporting processes.

The intersection of CPS 230 and AI represents a maturing of AI governance in Australian financial services. While compliance requires effort and investment, it also drives better risk management practices that ultimately protect both the entity and its customers. As AI becomes increasingly central to financial services delivery, CPS 230 provides a framework for deploying it responsibly and resilably.

Entities that get ahead of these requirements - building robust AI governance now rather than treating it as a compliance checkbox - will be better positioned not just to meet regulatory expectations but to realise AI's benefits while managing its risks effectively.