Assurance hasn't caught up with AI - and regulators have noticed
Traditional assurance methods were built for systems that behave consistently. AI does not. Models drift, adapt and degrade over time in ways that point-in-time audits and sample-based testing cannot reliably detect. APRA has identified this gap directly - observing that assurance practices are not keeping pace with the scale, speed and complexity of AI adoption. For internal audit and risk functions, this is both a capability challenge and an urgent one.
The challenge
Internal audit and second-line risk functions are struggling to keep pace with AI deployment. Many lack the specialist skills, tools and methodologies to assess probabilistic models, agentic workflows or AI-assisted code generation. APRA observed that assurance activities are frequently lagging AI deployment - meaning systems go live and operate for extended periods before any independent scrutiny is applied. In a fast-moving environment where AI touches critical operations and customer outcomes, that lag carries real risk.
Key insights
Point-in-time assurance is structurally insufficient for AI
Sample-based, periodic audits were designed for systems that behave the same way each time they run. AI models are probabilistic - their outputs vary, their performance degrades, and their behaviour can shift as data patterns change. An audit that assessed a model at deployment tells you very little about how it is performing six months later. APRA expects continuous monitoring proportionate to the criticality of each AI use case, with defined thresholds that trigger review or intervention when breached.
AI risk cuts across functions in ways existing assurance structures do not handle
AI risk sits at the intersection of operational risk, cyber and information security, data governance, model risk, change management, legal and regulatory compliance, privacy and conduct risk. Existing assurance structures tend to be siloed by domain. APRA expects integrated assurance for AI across cyber security, data governance, model performance, operational resilience, privacy and conduct risk - which requires deliberate coordination between functions that do not typically operate as a team. Fragmented assurance produces blind spots, and blind spots in AI produce incidents.
Second-line and internal audit need specialist capability - now
APRA observed that many internal audit and risk functions lack the technical capability to independently assess AI systems. Where agentic behaviour, automated decision-making or AI-assisted code generation are involved, the skills gap is particularly acute. Building this capability is not optional - it is a prerequisite for meeting APRA's assurance expectations. Entities need to assess honestly whether their current functions can do this work, and what resourcing or specialist support is required to close the gap.
Change management controls are under strain from AI-generated code
APRA noted that the volume and speed of AI-assisted software development is placing strain on the effectiveness of change and release management controls. When developers use AI to generate code, existing review and testing processes may not adequately assess what has been produced. Security testing programmes need to explicitly cover AI-generated code and software components. Entities that have not reviewed their change and release controls in light of AI-assisted development have a gap that is likely growing.
Assurance must cover the full lifecycle - including decommissioning
Most assurance attention is focused on AI deployment. APRA's expectations extend across the full lifecycle: design, development, deployment, ongoing monitoring and decommissioning. Models that are retired without structured governance - where training data, access controls and dependencies are not formally closed out - can create residual risk that lingers undetected. A complete assurance framework follows the AI from first approval to final retirement, not just from testing to go-live.
What this means for your organisation
Encompass works with risk and assurance functions in regulated entities to design AI assurance frameworks that are continuous, integrated and calibrated to the criticality of each use case. This includes assurance maturity assessments, methodology design and support in building the internal capability needed to independently scrutinise AI systems - including probabilistic models and agentic workflows - with the rigour that regulators now expect.