AI Regulation is Coming to Australia: Most of what it will ask, you can already work out.

Australia spent seven years deciding not to regulate AI. The Ethics Principles were voluntary. The AI Safety Standard was voluntary. The proposed mandatory guardrails for high-risk AI were shelved as recently as December 2025. Then on 15 July the government changed its mind: an Office of AI established inside PM&C, and a mandatory set of Australian AI Standards to be legislated in early 2027.

Which raises the only question that matters commercially. What will those Standards actually require? The announcement says very little, beyond data centre energy and water rules and a firm line that Australian creative work cannot train models without consent. Everything else is a heading.

The shape is easier to infer than it looks, for two reasons.

The first is that the government has told us how narrow it intends to be. The Prime Minister was explicit that the goal is not to “legislate for every possible eventuality or risk,” and fast-tracked approvals sit alongside the new obligations. That rules out the one template a lot ofpeople assumed Australia could borrow: EU.

Since 2024 the EU AI Act has been the world’s reference implementation for AI regulation: sort systems into risk tiers, then require the high-risk ones to pass conformity assessment before they go anywhere near a user. It is comprehensive, it is heavy, and it is now being walked back by its own authors. The EU’s Digital Omnibus on AI, agreed this year, defers those high-risk obligations from August 2026 out to December 2027 and trims parts of the regime, driven by concern that Europe was regulating itself out of the AI economy. Australia is declining to copy a model that Europe is currently softening.

Strip away the possibility of an EU-style conformity regime and what remains is the concern Australian policy has centred since 2019. What happens to personal information, and where data lives.

The second is that government is already demonstrating its own answer. The APS is mid-rollout of AI to more than 220,000 public servants, with Chief AI Officers appointed across agencies from July 2026 and GovAI Chat running on Australian sovereign infrastructure so data stays in-country. A coordinating office sitting in PM&C makes those internal patterns the likely reference point for the national Standards. Sovereign hosting, data residency, controlled access, auditable use. If you sell to government or operate in a regulated sector, that bar reaches you commercially well before it reaches you legally.

So the direction is legible. And on one point you don’t have to infer at all.

December, not 2027

From 10 December 2026, under the new APP 1.7, APP entities must disclose in their privacy policy the kinds of personal information used in automated decision-making and the kinds of decisions being made. The trigger is arranging for a computer program to use personal information to make, or to do something substantially and directly related to making, a decision that could reasonably be expected to significantly affect someone’s rights or interests.

Read “substantially and directly related to making” carefully. A human signing off at the end does not put you outside this.

You cannot disclose what you cannot see, so the work starts with knowing which systems in your estate shape decisions about people. That is the same inventory the 2027 Standards will assume you already hold.

The playbook is already published

The Voluntary AI Safety Standard has been sitting there since 2024, largely ignored because it was optional. It is the most likely basis for the mandatory version, which makes it the closest thing available to an early draft of your 2027 obligations.

The better argument for following it is that almost nothing in it is compliance work. Each item earns its keep on delivery grounds alone.

1. Inventory your AI. Which systems influence decisions about people, who owns them, what they touch. This is also how you find the shadow pilots already using customer data and the four teams solving the same problem.
2. Catalogue data and trace lineage. This is the accuracy mechanism for RAG and agents. Provenance is how you know an answer came from the current policy document and not the 2019 draft.
3. Enforce access control properly. Row and column-level control is what lets real data near a model at all. Without it, your valuable use cases either get blocked by risk or leak something.
4. Record training data provenance and consent. This decides whether you can commercialise the output, which is precisely the exposure the copyright position creates.
5. Build for explainability. Reconstructing inputs, model version and data as at the moment of a decision is what gets systems through legal review. It is architectural and cannot be retrofitted cheaply.

Do these and December takes care of itself, 2027 is mostly evidence-gathering, and your AI programme gets faster in the meantime. That is an unusually good trade.

About the author

Ricardo Neves is a Principal Data Consultant at DNX Solutions, based in Sydney. Over seven years in data and cloud, he has worked across engineering, architecture, data science, and consulting, which is the vantage point this piece draws on: knowing what a system touches is the first step to knowing what it owes a regulator.