The question is no longer "which AI tool?"
The letter is not a new rule. It is guidance under existing prudential standards, drawn from a review of a group of large banks, insurers and superannuation trustees in late 2025. But it changes the question a board should be asking.
APRA expects entities to be "mapping and maintain visibility over the full AI supply chain, including material, third-party and fourth-party dependencies." And where AI supports critical operations, it says plainly that "credible fallback processes are required."
Most organisations can name their AI vendor. Very few can name what sits underneath it.
Mapping the chain.
Picture the chain behind a single AI answer. At the top is the product you bought and can name. Beneath it sit the plug-ins and connectors it uses to reach your data, the cloud it runs on, and the foundation model doing the reasoning. Each of those is someone else's supplier: your fourth parties.
Most firms can name their AI vendor. APRA expects the whole chain.
Your people and processes
Who asks, who decides, who is accountable
You
The AI application or copilot
The product you bought and can name
Third party: your vendor
Plug-ins, agents and data connectors
What the vendor builds on to reach your data
Fourth parties, often unnamed
Cloud hosting and compute
Where it runs, and under whose law
Fourth party
The foundation model
Versions, updates and training terms
Fourth party
APRA observed these are often opaque
Your knowledge: the one layer you can own outright
Held independently of every layer above, it is what makes substitution and exit credible.
APRA's review found what you would expect. It observed heavy dependence on a single provider across several AI use cases. It found contracts that "often lagged practice", with limited evidence of audit rights, model-update terms, incident notification and data-handling terms. And it noted that upstream dependencies such as foundation models are often opaque.
The mapping is the easy part. The harder question is what you would do if any layer failed, changed its terms or simply changed its model.
A fallback is only credible if the knowledge can move.
APRA expects entities to assess "the credibility and feasibility of substitution, portability or exit arrangements for critical AI providers." It also observed that "few entities had demonstrated robust contingency planning or tested exit and substitution strategies."
Most exit plans fail on one point. Switching the model is usually possible. Switching what the system has learned is not, because that learning was never yours to take. The prompts, the corrections, the decisions people made with it, the documents it was taught from: in most rented AI, all of that sits in the supplier's format, on the supplier's side of the line.
So the most credible fallback is not a second vendor waiting in the wings. It is knowledge you hold independently of every layer in the chain, in documented, exportable form. Then a provider change is a migration, not a rebuild, and "tested exit" is something you can actually test.
That is also where sovereignty and supervision meet. Running locally or on your own infrastructure removes some layers from the chain altogether. Owning the knowledge makes every remaining layer replaceable. For the four tests that separate real sovereignty from a postcode, see Sovereignty is control, not location.
A one-page map for each critical AI use.
For every AI use that supports a critical operation, fill in one row per layer. Any cell you cannot fill is a finding in its own right.
| Layer | Who provides it | What you can see (audit rights, model updates, incident notice) | If it fails or changes: your fallback | Tested? |
|---|---|---|---|---|
| AI application or copilot | ||||
| Plug-ins, agents and connectors | ||||
| Cloud hosting and compute (and jurisdiction) | ||||
| Foundation model (and version) | ||||
| Your knowledge: where it is held and in what format |
AI application or copilot
- Who provides it: ______
- What you can see (audit rights, model updates, incident notice): ______
- If it fails or changes: your fallback: ______
- Tested? ______
Plug-ins, agents and connectors
- Who provides it: ______
- What you can see (audit rights, model updates, incident notice): ______
- If it fails or changes: your fallback: ______
- Tested? ______
Cloud hosting and compute (and jurisdiction)
- Who provides it: ______
- What you can see (audit rights, model updates, incident notice): ______
- If it fails or changes: your fallback: ______
- Tested? ______
Foundation model (and version)
- Who provides it: ______
- What you can see (audit rights, model updates, incident notice): ______
- If it fails or changes: your fallback: ______
- Tested? ______
Your knowledge: where it is held and in what format
- Who provides it: ______
- What you can see (audit rights, model updates, incident notice): ______
- If it fails or changes: your fallback: ______
- Tested? ______
The last row decides whether the others can be replaced.
Start with the layer you own.
APRA says that where entities fail to adequately identify, manage or control AI risks, it will take stronger supervisory action. None of that requires waiting for a new standard. The map above takes an afternoon per use case. The harder work, and the more valuable, is making sure the bottom row holds: that what your organisation knows is yours, in a form you can take with you.
If you are APRA-regulated and unsure where to start, the first conversation is a free 30-minute consultation.
