APRA has been direct about its expectations in a way that few regulators are. AI can assist. It must never be an autopilot. Boards of APRA-regulated entities are required under CPS 230 to understand and oversee every critical AI decisioning system — which in practice means they must be able to describe what the system does, how they would know if it was producing unexpected outputs, and what the human oversight mechanism looks like in substance rather than on paper. These are not aspirational governance standards. They are the current requirements of a prudential standard that has been in force since July 2025, and APRA has explicitly stated that governance documentation is a supervisory focus.
The institutions most exposed to this environment are not the ones that built bad AI systems. They are the ones that built technically capable AI systems without the governance infrastructure to document, oversee, and defend them. The cost of that gap is not theoretical. Westpac’s AUD 1.3 billion AUSTRAC settlement in 2021 — the largest corporate penalty in Australian history at the time — was not for having no monitoring system. It was for having a monitoring system that was inadequately designed, inadequately documented, and inadequately calibrated to the institution’s actual risk profile. The governance gap is the expensive gap.
The ANZ regulatory stack creates a specific governance architecture requirement
The three frameworks now in force in ANZ each impose distinct but complementary governance requirements, and together they specify something close to a complete AI governance architecture for fraud and credit decisioning.
CPS 230 requires operational resilience documentation for critical AI systems: identification of the system as a critical operation, tolerance thresholds for disruption, business continuity planning, and board-level oversight evidence. CPS 234 adds information security requirements specifically for AI and technology systems. The Scams Prevention Framework requires evidence that real-time detection controls were adequate — which in a regulatory dispute means producing the documentation of what controls existed, how they worked, and why they were considered sufficient for the institution’s specific scam typology exposure. AUSTRAC’s reform law requires expanded transaction monitoring with documentation of alert calibration, investigation outcomes, and programme effectiveness. ASIC’s conduct and fairness obligations require that AI credit and decisioning models do not produce discriminatory outcomes and that the institution can demonstrate this.
A bank that has built AI systems with governance-first architecture — explainability built in, audit trails operational, human oversight mechanisms documented and tested, bias monitoring active — satisfies all five frameworks simultaneously from the same infrastructure. A bank that has built capable models without this infrastructure must now retrofit five different compliance requirements onto systems that were not designed to accommodate them, at a cost that is typically three to five times higher than building it in from the start.
The compliance view and the advantage view produce different institutions
The framing that produces poor outcomes is familiar: governance is box-ticking, documentation is a cost to minimise, explainability slows development, audit trails are a liability requirement, and fairness metrics constrain model accuracy. Institutions operating under this framing consistently find that models in production lack the documentation required for CPS 230 review, that explainability cannot be added without re-engineering, and that APRA supervisors are adversarial rather than cooperative because the evidence of sound process is unavailable.
The advantage framing produces materially different outcomes. Banks that treat explainability as a design requirement find that models receive APRA and AUSTRAC sign-off faster — examiners can engage with decision logic rather than being asked to take on trust that the model is sound. Banks that treat audit trails as a data asset find that every logged fraud or credit decision is a labelled training example, enabling continuous improvement in ways that a system without structured logging structurally cannot. Banks that treat operational resilience as an architecture requirement find that CPS 230 compliance is a natural output of good system design rather than a retrofit project. And banks that treat CDR data governance as an investment find that the consent infrastructure and data pipelines required for compliance are the same ones that enable the more accurate, more inclusive credit models that represent a genuine competitive advantage in the Australian mortgage market.
The financial difference is quantifiable. Retrofitting CPS 230-compliant governance documentation, SPF-defensible detection evidence, and AUSTRAC programme effectiveness infrastructure into production AI systems that were not designed to accommodate them typically costs three to five times what it would have cost to build them in from the start, because the work touches model architecture, data pipelines, monitoring systems, and regulatory documentation across the entire deployment. Institutions that invested in governance-first AI before July 2025 are now spending a fraction of what those currently remediating under examination conditions will spend to achieve the same outcome.
There are three paths — but in ANZ, one of them is already closed
In North America and EMEA, the three paths remain open to varying degrees: lead, follow, or defer. In ANZ, the third path is not available. CPS 230 is in force. The Scams Prevention Framework is law. AUSTRAC reform is enacted. An institution deferring AI governance is not making a strategic timing choice — it is accumulating current non-compliance that will present as either an examination finding, an SPF liability claim, or an AUSTRAC enforcement action, depending on which failure crystallises first.
The first path is to lead from the current position. For institutions that began CPS 230 preparation before the July 2025 commencement date, the priority is to move governance-grade AI into production on the highest-value decisions — NPP fraud detection, AML transaction monitoring, mortgage credit decisioning — and to establish the feedback loops and audit trail infrastructure that will produce compounding model quality. The regulatory investment is already made. The commercial advantage is in using it.
For institutions that are still closing their CPS 230 gaps, the sequence is critical. Governance infrastructure and the first production AI deployment should happen simultaneously, not sequentially. Remediating CPS 230 compliance as a separate project and then beginning AI deployment is the most expensive path — it builds the governance infrastructure twice and extends the timeline for capturing the commercial value that justifies the investment in the first place. The Westpac precedent makes clear that AUSTRAC is prepared to impose consequences commensurate with the scale of the failure. APRA’s supervisory posture under CPS 230 is one of active engagement. Neither regulator is waiting.
The question facing ANZ banks is not whether governance-first AI matters. Every regulatory framework currently in force in the market has answered that question. The question is how much of the remediation cost will be paid under examination pressure versus how much will be invested as deliberate commercial strategy. The institutions that treat compliance infrastructure as the foundation for competitive advantage — rather than a cost to minimise — will compound that advantage with every quarter of outcome data they accumulate while others are still catching up.
Part 3 of 3.
Sources
Australian Prudential Regulation Authority. Prudential Standard CPS 230 Operational Risk Management. Effective 1 July 2025. Australian Prudential Regulation Authority. Prudential Standard CPS 234 Information Security. Australian Prudential Regulation Authority. CPG 230 Operational Risk Management Guidance. June 2024. AUSTRAC. Anti-Money Laundering and Counter-Terrorism Financing Reform. Effective March 2026. Australian Government. Scams Prevention Framework legislation. Australian Securities and Investments Commission. AI Model Conduct and Fairness Obligations. Australian Banking Association. Westpac AUSTRAC Settlement Reference. AUD 1.3 billion, 2021. Australian Government. Consumer Data Right / Open Banking framework.
