Your business already runs on AI. When the board asks you to prove it, what will you show?
For most regulated Australian businesses in financial services, health, insurance, and government, the honest answer today is a shrug. Australia has assigned accountability for AI faster than anyone can prove it is being met. The technology is useful. The problem is proving it is safe, to a board, to an auditor, and increasingly, to a regulator.
The regulator has already moved
That is no longer hypothetical. On 30 April 2026, APRA issued its first AI-specific Letter to Industry, setting out expectations for boards and accountable executives across four areas: cyber and information security, governance, supplier risk, and change management and assurance. Its finding was blunt. AI governance is not keeping pace with adoption. APRA signalled stronger supervisory action and, where appropriate, enforcement in reserve for entities that fail to manage AI risk. This is the acute pressure, and it applies now.
And the next obligation is already scheduled. Under the Privacy and Other Legislation Amendment Act 2024 (Cth), automated decision-making transparency obligations commence on 10 December 2026. Businesses using AI to make decisions that significantly affect someone’s rights will need to disclose what data was used and how the decision was reached. Most organisations have no mechanism to produce that evidence today, and the record you will need starts accumulating now.
The three obligations that matter
Three regulatory pressures converge on any regulated business sending data to an AI model:
- APRA CPS 234. Information security for APRA-regulated entities. Section 15 covers third-party arrangements. If you send data to an overseas AI provider, you are expected to demonstrate active oversight of that arrangement, not just a contract on file. The April 2026 letter is APRA telling boards it now expects this for AI specifically.
- Privacy Act 1988, APP 8. Cross-border disclosure of personal information. The moment personal information leaves Australia for an overseas AI provider, APP 8 obligations are triggered. The question an auditor asks is simple: can you show, per request, what was disclosed and where it went?
- AI transparency (from 10 December 2026). Disclosure of the data inputs and decision basis behind automated decisions. This is the later, softer driver, but the record you will need to satisfy it starts building the day you turn AI on.
The common thread is evidence. Each obligation assumes you can produce a record of what happened, when, and under what controls, at the level of an individual AI call.
Enforcement is not theoretical either.
$5.8 million. Australia’s first Privacy Act civil penalty, against Australian Clinical Labs in October 2025. Most of it for one thing: failing to take reasonable steps to protect personal data under APP 11.1.
The statutory maximums are larger still (penalties for serious or repeated breaches can reach $50 million), but the more instructive signal is simply that the regulators have started to act.
Why post-hoc log analysis isn’t enough
The standard answer is to analyse logs after the fact. Pull the request logs, run a classifier, generate a report. That works until a regulator asks you to prove the control was active at the time of the request, not reconstructed afterwards from logs you could, in principle, have edited.
The shift that matters is from after-the-fact analysis to inference-time evidence: a record generated at the moment of the AI call, cryptographically signed, and stored so that it cannot be altered without detection.
A record generated at the moment of the call, not reconstructed from logs you could, in principle, have edited. That is the difference between a report and evidence.
Evidence is permission
There is an upside to getting this right that the compliance framing hides. The brake on AI in most regulated businesses is not the technology. It is that the risk function cannot say yes defensibly, because there is nothing to review.
Per-call evidence changes that answer from “no, we can’t see it” to “yes, and here’s the proof.” The businesses producing evidence by default are the ones putting more AI into production, sooner, because each use ships with its own record.
What defensible actually looks like
A defensible AI posture for an Australian regulated business produces, for every single AI call:
- A scan for Australian PII (Tax File Numbers, Medicare numbers, ABNs, BSB and account details) using validation, not just pattern matching.
- A check for prompt injection in the request, including content hidden inside uploaded documents.
- A signed attestation that maps what happened to the relevant control (CPS 234 S15, APP 8, AI transparency) so the evidence speaks the regulator’s language.
- An immutable, tamper-evident audit trail retained for seven years.
None of this should change how your team works. The right architecture captures those calls, by proxy or by ingest, and does the compliance work transparently. One API URL changes. Nothing else.
Readiness is an architecture decision
APRA’s posture is the forcing function today, and December 2026 adds to it rather than starting the clock. The underlying obligations, CPS 234 and APP 8, already apply. The businesses that will be ready are the ones that treat AI compliance as an architectural decision now, so that when the board asks the question, the answer is a record rather than a scramble.
If you’re weighing how to put per-call evidence in place, get in touch. It’s the conversation we’re built for.