By Dave Keys, Founder and Chief AI Officer, aify. Position as at 9 September 2026. General information, not legal advice.

Most of the executives I talk to are waiting for Australia's AI law before they take AI governance seriously. That law is not coming. The government dropped its proposed mandatory guardrails, and the National AI Plan says the existing regulators will handle it with the laws they already have.

The law they already have is the Privacy Act. It has applied to every AI tool your staff have used since the day they started using them, and on 10 December 2026 it gains the first obligation written specifically for AI-assisted decisions. This is what that means for AI tools, and what to do about it.

What the Privacy Act is, and what it is for

The Privacy Act 1988 is the Commonwealth law that governs how organisations collect, hold, use and disclose personal information. Its working core is thirteen Australian Privacy Principles, which cover the lifecycle of information about a person: telling people what you collect and why, only collecting what you need, using it for the purpose you collected it for, keeping it accurate and secure, and letting people see and correct it.

Its intent is simple. People should be able to deal with organisations without losing control over information about themselves, and organisations should be able to demonstrate that their handling of that information is reasonable. The Act does not care what technology you use. It cares what happens to the person.

That is why it works as an AI law without mentioning AI. Putting a client's file into an AI tool to summarise it is a use of their personal information. An AI system that infers something about a person is collecting personal information. A system that produces confident but wrong output about someone engages the obligation to keep records accurate. A vendor that hosts your data offshore and trains on it is subject to security and overseas disclosure rules. The regulator spelled all of this out in October 2024, along with its view that organisations should not enter personal information into publicly available generative AI tools at all.

What has changed, and what is still coming

The Act was amended in December 2024, and three of those changes matter for AI tools.

Penalties are now tiered

A privacy policy that does not say what it should is a low-tier breach, up to $330,000. An interference with someone's privacy that is not serious is a mid-tier breach, up to $3.3 million. A serious interference is up to $50 million, three times the benefit gained, or 30 per cent of turnover, whichever is greatest. The regulator can now issue infringement notices without going to court. And since June 2025, an individual can sue directly for a serious invasion of their privacy.

Security requirements are more explicit

Security now expressly requires ‘technical and organisational measures’. That phrase is the one a regulator will reach for when AI was deployed with no policy, no register and no owner.

AI-assisted decisions must be disclosed from 10 December 2026

From 10 December 2026, if a computer program makes, or substantially assists in making, a decision that could significantly affect a person's rights or interests, and personal information is used to run it, your privacy policy must say so. It must disclose the kinds of personal information the program uses, which decisions are made solely by the program, and which decisions the program substantially assists. It is a disclosure obligation only. There is no right to a human review yet.

The next round of reform is a draft. An exposure draft released on 31 August 2026 proposes a fair-and-reasonable test for collecting and using information, stricter consent, 72-hour breach notification and an audited obligation to destroy what you no longer need. It keeps the small business exemption. It is expected to commence in 2027 or 2028. Build towards it.

Three assumptions I keep hearing that are wrong

"We turn over less than $3 million, so the Act does not apply to us." 

If you provide a health service and hold health information, it applies at any turnover, and the regulator's list of health services includes disability and aged care providers. Since July 2026, it also applies to accountants, lawyers and real estate agents regardless of size. The exemption covers fewer organisations than most people think.

"A person signs off on every decision, so it is not automated." 

The regulator has signalled a broad reading. If the program's output is a key factor in the decision, the disclosure obligation applies whether or not a human approves it: rostering, triage, prioritisation, and risk flagging all fall within that reading.

"Nobody audits privacy, so the risk is theoretical." 

Nobody audits it on a schedule. But the regulator now runs proactive sector sweeps, opens its own investigations in response to media coverage, and has named AI a priority. And for care providers, the Commission auditor tests information management every cycle. 

Your AI governance will be examined by an auditor long before the privacy regulator does.

How it lands on Australian businesses

For most organisations, the practical impact of AI tools is not a new system. It is being able to answer three questions with evidence: which of our tools use AI or automation to make or shape decisions about people, what personal information do those tools hold and where, and who approved their use.

The organisations that struggle are not the ones that made a deliberate decision about AI. They are the ones whose staff adopted a dozen AI tools with no decision at all: a note summariser here, a scoring feature switched on in the CRM there, a rostering optimiser the vendor added in an update. None of those was assessed, none are in the privacy policy, and each is a low-tier breach waiting for a sweep.

What to do, and by when

Four things by 10 December. 

Inventory every AI and automated tool that touches personal information, including AI features inside software you already own. Classify each one against the decision test: does it significantly affect someone's access to services, entitlements, pricing, employment or care? 

Rewrite the privacy policy so it carries the required disclosures and a plain account of how AI is used. And get written disclosure from each vendor: how the system determines it does not train on your data, where it is hosted, and what happens in the event of a breach.

Three things after. 

Run a privacy impact assessment for each higher-risk use and put a human check in front of any AI output that becomes a record about a person. Put controls around the AI tools: an approved tools policy, access controls, logging, and a retention rule for prompts and outputs. And name one accountable owner who keeps the register, assessments, and vendor checks up to date.

Where it bites hardest

NDIS

Almost everything a provider holds is sensitive information, which means consent-based handling and a higher security bar than most sectors. The December obligation captures the AI tool providers that are most keen to adopt: rostering, incident severity classification, referral prioritisation, and participant risk flags are all decisions about access to support. 

The NDIS Commission has published no guidance on providers' use of AI, so the Privacy Act, plus the national six practices for AI adoption, are the standards you will be held to. The Practice Standards audit is where it will be tested first.

Aged care

The same coverage applies, plus the new Act. Outcome 2.7 of the strengthened Quality Standards requires an information management system that is ‘in line with applicable privacy laws’, informed consent to the collection, use, storage, and disclosure, and residents' ability to access and correct their information in a form they understand. 

Nothing in the Standards mentions AI, so the defensible position is to treat every AI tool as part of that system and evidence it accordingly. The draft reforms reclassify precise location and biometric data as sensitive, which is where wander alerts, fall detection and facial recognition at entry will land.

The finance function in a small or mid-sized business

A business under $3 million that is not otherwise caught may be exempt from the Privacy Principles. Four things apply anyway:

• Tax file numbers: These are governed by a rule that binds every entity, so they must never enter an AI tool outside payroll-grade controls.
• Employee records: The exemption covers acts directly related to employment. It does not cover applicants, contractors, or handing employee data to an AI vendor for the vendor's purposes.
• AI-assisted credit decisions: These engage the credit reporting rules at any business size.
• Accounts payable, expense approval and collections: These can involve decisions about individuals, including sole-trader suppliers, customers and staff. They are December disclosures if they affect entitlements.

The question to take to your next leadership meeting

If a regulator or auditor asked tomorrow which of your systems use AI tools to make or shape decisions about people, could anyone in the room produce the list?

How aify can help

The four December actions are a bounded piece of work with a hard date, and they are exactly what aify's Gateway Discovery Workshop is built for. In a single structured session, we identify every AI touchpoint in your decision-making, classify each one against the December test, and leave you with the privacy policy language and the vendor questions ready to send. 

For providers, we map the output to the NDIS Practice Standards or the Aged Care Quality Standards so the same evidence serves the Commission audit. We have published a full client reference and a three-page executive brief. 

If you want to know where your organisation stands, start with the Gateway and talk to aify.