Your AI supply chain is longer than you think
When a regulated entity deploys an AI capability, it rarely has full visibility of what sits beneath it. Foundation models, training data sources, hosting infrastructure and software platforms each introduce dependencies that may be opaque, untested and outside the entity's direct control. APRA has identified third and fourth-party AI supply chain risk as a material governance gap - and expects entities to map it, manage it and plan credibly for when it fails.
The challenge
AI capabilities are increasingly embedded in software, platforms and developer tools that regulated entities procure as finished products. What lies beneath - the foundation model, its training data, the hosting environment, the subcontractors involved - is often invisible. Contractual arrangements rarely address audit rights, model updates, incident notification or data handling changes with the specificity that prudential risk management requires. The result is exposure that entities cannot fully see, assess or control.
Key insights
Fourth-party risk is real and largely unmanaged
APRA observed that upstream dependencies - foundation models, training data providers, and platform subcontractors - are frequently opaque, limiting entities' ability to independently assess model performance, bias, resilience and security. Most third-party risk frameworks focus on direct suppliers. The AI supply chain adds layers beneath that: a vendor's AI capability may depend on a foundation model from a hyperscaler, trained on data from a separate provider. Each layer introduces risk that contractual arrangements with the direct vendor may not address.
Concentration risk is a systemic concern, not just an entity-level one
APRA noted that some entities are heavily dependent on a single provider for multiple AI use cases. Where multiple regulated entities share the same AI provider or foundation model, a single failure, outage or security incident creates systemic exposure across the sector. Boards and risk functions need to assess not just their own concentration, but what happens if a commonly used platform fails across the industry simultaneously. Plausible substitution and exit strategies must be credible, not theoretical.
Contracts written for traditional technology do not work for AI
Standard technology procurement contracts were not designed for AI. They typically lack provisions for model updates and deviations, audit rights over model behaviour, incident notification obligations, data handling changes and performance degradation thresholds. APRA expects contractual and governance arrangements to provide sufficient transparency, auditability and assurance over AI services. Reviewing and updating supplier contracts against these expectations is a practical first step that many entities have not yet taken.
You need to be able to understand what your AI is doing - and why
APRA's expectation is that entities can understand model behaviour, material changes, performance issues and outcomes across the service lifecycle. For many deployed AI capabilities, this level of transparency does not currently exist. Entities cannot explain outputs, detect drift, or identify when model behaviour has materially changed. Where AI touches critical operations or customer decisions, the inability to explain what the model is doing - and why - is both a governance failure and a conduct risk.
Exit and substitution strategies must be tested, not assumed
APRA found few entities had demonstrated robust contingency planning or tested exit and substitution strategies for critical AI providers. "We could switch providers if needed" is not a substitution strategy. Genuine resilience requires understanding the portability of data and models, the time and cost to substitute, and whether a credible alternative actually exists. For entities where AI supports critical operations, the absence of a tested fallback is a material operational resilience gap.
What this means for your organisation
Encompass helps regulated entities map their AI supply chains, assess concentration and fourth-party exposure, and design contractual and governance arrangements that meet APRA's expectations. The work begins with visibility - understanding what you have deployed, who sits behind it, and what would happen if it stopped. From that foundation, credible risk management becomes possible.