APRA letter to industry on artificial intelligence

What it is

On 30 April 2026 APRA published a letter to all regulated entities on artificial intelligence, signed by APRA Member Therese McCarthy Hockey. It followed a targeted engagement in late 2025 with a group of large banks, insurers and superannuation trustees. APRA found that AI is being adopted actively, but that governance, risk management, assurance and operational resilience are not keeping pace with the scale, speed and complexity of AI.

The letter sets out APRA’s observations and its expectations. An attachment, written for accountable executives such as the CRO, CTO and CISO, gives the detailed findings and minimum expectations.

Who it applies to

All APRA-regulated entities. That covers authorised deposit-taking institutions, general and life insurers, private health insurers and registrable superannuation entity licensees. APRA says the lessons from its sample of large entities are meant to help those earlier in their AI adoption, and that it will judge risk management in proportion to each entity’s size, scale and complexity.

What it requires

APRA does not create a new standard. It says its principles-based framework is technology and vendor neutral and already requires entities to manage AI risk, set risk appetite, manage exposures and maintain oversight and accountability. CPS 230 (Operational Risk Management) and CPS 234 (Information Security) are the standards most directly engaged.

Boards. At a minimum, boards should have enough AI understanding and literacy to set direction and give effective challenge. They should oversee an AI strategy consistent with risk appetite, supported by monitoring and reporting, including on third-party dependencies, with clear triggers for action. APRA noted an overreliance on vendor presentations.

Governance. APRA saw a tendency to treat AI as “just another technology”, with weak controls after deployment over model behaviour monitoring, change management and decommissioning. These are its minimum expectations.

• Frameworks (policy, standard, guidance) and reporting lines for AI adoption.
• Ownership and accountability across the AI lifecycle, from design to decommissioning.
• An inventory of AI tooling and AI use cases.
• Human involvement for high-risk decisions and accountability.
• Staff training on AI use, misuse, limitations and secure practices.

Information security. Entities should assess how relying on AI affects resilience, with credible fallback where AI supports critical operations. They should have controls for AI-specific threats, including privileged access management, patching, hardened configuration, penetration testing and controls over agentic and autonomous workflows. Security testing should cover AI-generated code and components. APRA flagged identity and access management not yet adjusted to non-human actors such as AI agents, and staff using enterprise AI tools outside approved controls.

Suppliers. Entities should map the full AI supply chain, including third and fourth parties. Contracts should give transparency, auditability and assurance over model behaviour, material changes and performance. Concentration risk needs active management, with credible substitution and exit plans for critical AI providers.

Assurance. APRA found point-in-time, sample-based assurance ill suited to models that learn, adapt and degrade. It expects globally recognised control frameworks and change control for AI, integrated assurance across cyber, data, model, resilience, privacy and conduct risk, and second-line and internal audit functions with the skills and tools to assess AI. It also expects risk and security assessments before deployment and throughout the lifecycle, with continuous monitoring proportionate to the use case’s criticality.

Key dates

CPS 234 has applied since 1 July 2019 and CPS 230 since 1 July 2025. The letter was published on 30 April 2026. APRA said it was finalising its forward plan for AI supervision, including entity reviews, thematic work and engagement with AI suppliers, and encouraged entities to raise heightened AI risk concerns with its Non-Financial Risk team early.

How it relates to other frameworks

APRA does not name a framework to follow. It asks for globally recognised control frameworks, which in practice means instruments such as ISO/IEC 42001 and the NIST AI Risk Management Framework. The governance expectations line up closely with the six practices in the Guidance for AI Adoption. Decisions APRA names, such as loan application processing and claims triage, often fall within the Privacy Act ADM transparency rules from 10 December 2026.

The APRA AI expectations guide walks through how to do the work each expectation describes.