The picture changed twice in eight months.
In December 2025 the government dropped the mandatory AI guardrails everyone had spent two years preparing for, and a lot of businesses read that as a reprieve and stopped thinking about it. Then in July 2026 the Prime Minister announced Australian Standards for AI that will carry mandatory obligations. In August, National Cabinet backed mandatory standards for large data centres, with legislation planned for early 2027.
Meanwhile one obligation with a fixed date has been sitting there the whole time. It starts on 10 December 2026.
What has the government decided?
Two things, and most commentary only covers the first.
December 2025. After consultation that drew more than 300 submissions, the government decided not to proceed with the proposed mandatory guardrails for high-risk AI. The National AI Plan of 2 December leaned on existing law instead, supported by non-binding guidance and an advisory AI Safety Institute.
July 2026. On 15 July the Prime Minister set out a different direction: Australian Standards for AI intended to apply across the economy and, for the first time, carry mandatory obligations. Until then, AI standards had relied on guidance and self-attestation. An Office of AI was established in the Department of the Prime Minister and Cabinet the same day.
The concrete requirements published so far are aimed at large AI data centres, covering energy, water and grid connection, plus protections for Australian creative work. On 26 August National Cabinet backed consistent mandatory standards for large data centres, which the Commonwealth will develop with the states. The government’s September consultation paper proposed standards covering large data centres and frontier AI training in Australia. Nothing in that paper applies to a business that simply uses AI, though whether anything lands on ordinary businesses later still isn’t settled. Legislation is expected in early 2027.
So it’s too early to say anyone dodged it. The shape of any future obligation is being decided now, and nobody can tell you yet what it’ll ask of a ten-person firm.
Is anything required today?
Yes, under laws with no “AI” in their titles.
Australian law is technology neutral. The Privacy Act doesn’t care whether a person or a model exposed a client’s information. Consumer law doesn’t care whether a misleading quote was typed or generated. Discrimination law doesn’t care whether a shortlist came from a hiring manager or a tool the hiring manager trusts. None of that softened when the guardrails were dropped, and none of it is waiting for 2027.
The burden of proof hasn’t moved. In practice, though, if something goes wrong you’ll be asked what steps you took, and “the AI did it” has never been an answer.
Which obligation has a date on it?
From 10 December 2026, an organisation covered by the Privacy Act has to say in its privacy policy where it uses personal information in automated decisions that could reasonably be expected to significantly affect an individual’s rights or interests. You describe the kinds of personal information involved and the kinds of decisions being made.
What catches people out is that the obligation isn’t limited to decisions a computer makes on its own. It also covers a program that does something substantially and directly related to making the decision, which includes recommending a decision to a human or guiding one. A person reviewing and approving the output doesn’t, on its own, put you outside it. If a tool is materially shaping the decision, it counts.
The OAIC published its final guidance on 30 September 2026: an updated chapter 1 of its APP guidelines and a fact sheet on the automated decision transparency obligation. The guidance says decisions that use AI outputs and significantly affect people will generally be in scope unless they’re subject to extensive human oversight and control. Its worked example suggests that means people questioning the output, checking the evidence behind it and recording why they departed from it. If you’re unsure whether you’re in scope, its advice is to take the cautious approach and put the information in your privacy policy.
The obligation isn’t limited to AI either. Rule-based systems and automated assessments are captured on the same terms. It carries civil penalty exposure, and it applies to decisions made after commencement even where the system and the data predate it.
Does any of this apply to a small firm?
It depends. The alarmist version of this article would be more persuasive and less true.
The Privacy Act’s small business exemption still exists, but it’s been narrowed, and three things can catch a small firm anyway.
The AML/CTF reforms, from 1 July 2026. A new group came into the reporting regime, and it reads like a list of my clients: real estate professionals, lawyers, accountants, conveyancers, trust and company service providers. But capture depends on providing a designated service, whatever your profession. A tax-only accountant or a litigation-only firm may not be caught at all.
And where a firm is caught, the exemption falls away for the personal information it handles for those obligations, not for the whole practice. So a conveyancer who’s now a reporting entity picks up Privacy Act obligations around their customer due diligence, and the December disclosure requirement bites only if something in that scope is automated. Using AI to draft client letters isn’t what this is about.
Commonwealth contracts, at any size. This one has been there all along, and most people never knew. Among the businesses the Privacy Act covers regardless of turnover, the Information Commissioner lists a contractor providing services under a Commonwealth contract. Its own guidance defines that as providing services to, or on behalf of, Australian Government agencies under a Commonwealth contract or subcontract. So if you do defence, infrastructure or professional work under a head contractor who holds the contract, the turnover threshold doesn’t protect that contract work, and the December date applies to it.
The statutory tort, since June 2025. This one ignores the exemption entirely. An individual can sue for a serious invasion of privacy, and the defendant doesn’t have to be covered by the Privacy Act at all. It requires intentional or reckless conduct and a serious invasion, so it isn’t a trap for ordinary mistakes. But it does mean “we’re under the threshold” stopped being a complete answer over a year ago.
I’m not a lawyer. If you think you’re in scope, this is the point to talk to one. What you can do without advice is answer the question underneath all of it: does an automated tool shape decisions about people in this business? If yes, the advice is worth paying for. If no, the rest of this is about doing the basics well.
What does governance look like for a firm my size?
Less than the phrase suggests, and more than nothing.
I work with Sydney construction, property and professional services firms of 5 to 100 staff, and this section is written for them. A large enterprise, or anyone running AI through hiring, lending or customer decisions at volume, is in different territory and needs more than four questions.
Almost none of the firms I work with need a framework. They need a clear answer to four questions about any AI that touches a client:
What goes in? If client information is being pasted into a tool, someone should have decided that on purpose, knowing where the data goes and whether it trains anything.
Who checks what comes out? A name. If the answer is “we all keep an eye on it”, nobody does. This is about output quality and accountability inside your business. On its own, it doesn’t change your position under the Privacy Act.
What happens when it’s wrong? It will be, occasionally. The question is whether that surfaces in a review step or in front of a client.
Would you be comfortable explaining it? If a client asked how their proposal was produced, is the honest answer one you’d happily give?
Answer those four for each place AI touches your work and you have governance. It might fit on a page, which is more than most firms have.
Where AI sits in your work decides how long those answers get. When I teach a team, AI starts on low-judgement work: research, drafting, transcription, scheduling. Checking what it produces is a standing condition. For work like that the answers are short. They get longer as the tool takes on more of the judgement, and longest when it shapes a decision about a person.
What do most businesses get wrong?
Three things, none of them about policy.
They buy tools before deciding who owns the output. The subscriptions arrive one person at a time, each solving somebody’s immediate problem, and ownership of what comes out the other end is a decision nobody gets around to making.
They write the document and stop. Until a policy’s been tried on one of your workflows, nobody knows whether it works. If it’s never changed how a single quote gets produced, it’s a file on a shared drive.
They treat consistency as optional. Different people using AI differently across one firm produces work that doesn’t sound like one business. Clients notice that long before a regulator would.
If you want a starting point that isn’t written for a large enterprise, the National AI Centre’s Guidance for AI Adoption updated the ten guardrails in the 2024 Voluntary AI Safety Standard into six essential practices. It’s non-binding guidance. Your own version should be shorter and more specific than anything published for general use.
When is this overkill?
Often. If your first AI project needs none of this, I’ll say so.
If the only AI in your business is one person using ChatGPT to tidy up their own emails, governance means making sure they know not to paste client financials into it. One conversation covers that.
If AI touches only internal work that never reaches a client and never involves personal information, the risk is a wrong answer and a wasted afternoon.
The threshold moves when AI output reaches a client, shapes a decision about a person or touches information you hold on someone else’s behalf. Below it, writing a policy is theatre. Above it, not writing one is a decision you’ve made whether you meant to or not. The same threshold is a good guide to how far you let an AI tool run unsupervised.
Where should you start?
With a list: every place AI touches your work, including the informal ones nobody wrote down. Most firms are surprised by how long it is.
Then answer the four questions above for each. Most are quick. One or two won’t be, and that’s where the work is.
Then write down only what you had to decide. That’s your governance. If it runs past a few pages for a firm your size, you’ve written it for an audience that doesn’t exist.
With clients I run it in that order: find out what’s happening, decide who owns what, then write down only the decisions. You can see how that runs in the Korbai Method, and what an engagement covers in services. Or tell us where AI shows up in your business on a discovery call. It’s half an hour, and we’ll come back to you afterwards with what, if anything, needs writing down.
FAQ
Is AI governance a legal requirement in Australia?
Not as a standalone requirement. There’s no Australian AI Act and the proposed mandatory guardrails weren’t taken forward. Mandatory Australian Standards for AI were announced in July 2026 with legislation flagged for early 2027, but the detail for ordinary businesses isn’t settled. What already applies is existing privacy, consumer, discrimination and work safety law.
What does the Privacy Act now require about automated decisions?
From 10 December 2026, organisations covered by the Act must disclose in their privacy policy where personal information is used in automated decisions that could significantly affect someone’s rights or interests, including the kinds of information and the kinds of decisions involved.
We have a human review every AI output. Are we exempt?
Not by that alone. The obligation reaches systems that recommend or guide a human decision as well as ones deciding on their own, and the OAIC’s final guidance says a decision can be in scope even when a person reviews the output. For AI outputs used in decisions that significantly affect someone, it generally expects those decisions to be in scope unless there’s extensive human oversight and control, which is a lot more than a sign-off. Human review is still worth having for the quality of the work.
Does the Privacy Act apply to what we put into AI tools?
If you’re covered by the Act, yes, in the same way as anything else you do with personal information. Putting client data into a third-party tool will in most circumstances be a disclosure, and needs the same thought as any other disclosure.
We’re under the $3 million turnover threshold. Are we outside all of this?
Less reliably than you were. The small business exemption still exists, but the AML/CTF reforms from 1 July 2026 brought firms providing designated services into the regime, and the exemption doesn’t apply to the personal information handled for those obligations. Separately, the statutory tort for serious invasions of privacy has applied since June 2025 regardless of turnover or whether you’re covered by the Act at all. Check with your own adviser before assuming either way.
Who should own AI governance in a small business?
Whoever owns the process the AI touches. If AI drafts your quotes, the person accountable for quotes is accountable for the AI in them. IT or a committee sits too far from the work to notice when it goes wrong.
Should we wait for the new standards before doing anything?
I wouldn’t. Knowing what you use, who owns it and what happens when it’s wrong is the same short piece of work whatever arrives in 2027, and the December Privacy Act obligation doesn’t wait for it.
Where does your business stand?
Whether an automated tool shapes decisions about people in your business is the question most owners have never worked through for anything specific. The AI risk assessment does it for one use case at a time: what data it touches, what it can reach, what its output affects, who owns it and whether it strays into the territory above. It’s free, takes about five minutes and needs no email to see the result.
General information only, not legal advice. Current as at October 2026: this area is moving quickly and the position may have changed by the time you read it.
