AI Automated Decisions: 2026 Privacy Rules
A quiet deadline that touches almost every business
Most of the attention on artificial intelligence regulation in Australia has focused on what did not happen. The mandatory guardrails floated in 2024 were shelved, replaced in the December 2025 National AI Plan with voluntary guidance. That created a widespread and mistaken impression that AI in business remains largely unregulated for now.
It does not. While the AI-specific mandatory regime stalled, an existing law was quietly amended to reach directly into how businesses use AI to make decisions about people. The Privacy and Other Legislation Amendment Act 2024 (Cth) added new transparency obligations to the Privacy Act 1988 (Cth), and those obligations commence on 10 December 2026. From that date, any organisation covered by the Privacy Act that uses a computer program to make or substantially support decisions affecting individuals must say so, in plain terms, in its privacy policy.
This is not a niche rule for technology companies. It captures lenders assessing loan applications, insurers pricing premiums, employers screening job applicants, and any midsize business that has quietly wired an AI model or automated rules engine into a customer-facing process. If your organisation is in that position and has not yet mapped where automated decisions happen, the window to prepare is closing. This guide explains exactly what changes, who is captured, and the practical steps to be ready. For businesses weighing whether to bring in outside help, it also clarifies where structured AI advisory support earns its place and where it does not.
The regulation you were waiting for is already here The absence of a standalone AI Act does not mean AI is unregulated. The Privacy Act, the Corporations Act, anti-discrimination law and consumer law all already apply to automated systems. The 2026 amendment simply makes one of those obligations, transparency, explicit and enforceable.
What the new APP 1.7 actually requires
The amendment works by expanding Australian Privacy Principle 1, the principle that governs what an organisation must put in its privacy policy. A new APP 1.7 requires an APP entity to include information about automated decision-making in its privacy policy where that entity uses personal information in a relevant automated process.
The trigger is specific. Under the amended Act, the obligation applies where an APP entity arranges for a computer program to be used, with personal information about an individual, to make a decision, or to do a thing that is substantially and directly related to making a decision, that could reasonably be expected to significantly affect the rights or interests of that individual. Where that test is met, the privacy policy must set out the kinds of personal information used in the operation of the automated decision-making technology, and the kinds of decisions made solely or substantially by that technology.
A companion provision, APP 1.9, clarifies the reach of the phrase significantly affect. It confirms that a decision counts whether it affects the individual adversely or beneficially, and gives examples of the kinds of decisions in scope: a decision under an Act or legislative instrument to grant or refuse a benefit, a decision that affects an individual's rights under a contract or arrangement, and a decision that affects an individual's access to a significant service or support. The breadth of that language is the point. A great many routine commercial decisions are covered.
Which Automated Decisions Are In Scope
| Metric | Likely in scope | Likely out of scope | Improvement |
|---|---|---|---|
| Lending | Automated credit approval or decline | Internal fraud-flag review by a human | In scope |
| Insurance | Algorithmic premium or eligibility decision | Aggregate actuarial modelling | In scope |
| Employment | AI shortlisting or ranking of applicants | Roster optimisation with no personal impact | In scope |
| Services | Automated eligibility for a benefit or support | Product recommendation with no rights impact | Assess |
The distinction that matters most is between a decision that a program makes or substantially supports, and one where a person genuinely exercises judgement using a tool. A model that scores an applicant and a human who then rubber-stamps the score is far closer to automated decision-making than the presence of a person in the loop suggests. The Office of the Australian Information Commissioner has flagged that meaningful human involvement, not token sign-off, is what distinguishes the two, and it ran a public consultation through 2026 to shape the detail of how disclosures should be written.
Why transparency is only the visible part
The privacy policy disclosure is the letter of the new obligation. It is also the smallest part of the work. Writing two paragraphs about automated decisions is trivial once an organisation actually knows what those paragraphs should say. The hard part, and the reason 10 December 2026 is a genuine project rather than a copy-editing task, is that most businesses do not have a current, accurate inventory of where automated decisions are being made.
Automated decision logic accretes quietly. A rules engine added to a loan origination system in 2019, a third-party scoring API bolted onto an onboarding flow, a recruitment platform whose ranking feature was switched on by default, a pricing tool a vendor upgraded to include a model last year. Each was a sensible operational choice. Collectively they mean that the honest answer to "where do we make automated decisions about people?" is often "we are not entirely sure." You cannot disclose what you have not mapped.
This is why the amendment functions as a forcing mechanism for something businesses should do regardless. Producing a defensible privacy policy disclosure requires an AI and automation inventory, an assessment of which entries meet the significant-effect threshold, and a view on whether meaningful human involvement exists in each. That work surfaces risks well beyond privacy: discrimination exposure in automated screening, consumer-law exposure where automated decisions mislead, and governance gaps the board would want to know about. The transparency rule is the trigger, but the real deliverable is knowing your own systems.
From Trigger to Compliant Disclosure
A worked example of the threshold in practice
Consider a typical 200-person Australian non-bank lender. It runs an online application flow where an applicant enters income, expenses and identity details. A scoring model returns a risk grade, and the origination system automatically declines any application below a set threshold and auto-approves any above a higher one, referring the middle band to a human credit officer.
Two of those three paths are automated decisions in the APP 1.7 sense. The auto-decline and the auto-approve are made solely by the program, they use personal information, and they plainly affect the applicant's rights or interests under a prospective contract. The referred middle band, where a credit officer genuinely assesses the file, is closer to human decision-making, though the lender would still want to be confident the officer exercises real judgement rather than deferring to the score.
For that business, the privacy policy from 10 December 2026 must describe the kinds of personal information the scoring model uses, and state that lending decisions are made solely by automated technology in defined circumstances. Getting there means the lender first has to document the model's inputs, confirm which decision paths are fully automated, and decide whether the referred path counts. None of that is a legal drafting exercise. It is a systems-mapping exercise that the drafting depends on. This is the same discipline that underpins effective AI advisory work for boards and executives: you cannot govern, or disclose, what you have not mapped.
The pattern generalises. Swap the lender for an insurer setting premiums, a labour-hire firm ranking candidates, or an aged-care provider running automated eligibility checks, and the analysis is structurally identical. In fact, providers navigating the sector's other reforms will recognise the overlap: the same mapping rigour applies to AI automation under the new Aged Care Act, where automated eligibility for support is precisely the kind of "significant service" decision APP 1.9 calls out.
The cost of getting it wrong
The transparency requirement arrives alongside strengthened enforcement powers. The 2024 reforms gave the OAIC new tiers of civil penalty and, relevant here, the ability to issue compliance notices and infringement notices for specified contraventions. A privacy policy that does not meet the automated decision-making requirements from 10 December 2026 is exactly the kind of shortfall those notice powers are designed to address, without the regulator needing to run full-scale litigation to act.
The reputational exposure is arguably larger than the direct penalty. An organisation that discloses, accurately, that it makes automated lending or eligibility decisions invites scrutiny of whether those decisions are fair. That is uncomfortable, but it is far safer than the alternative: being found to have run automated decisions that materially affected people while saying nothing, then having that surface through a complaint, a journalist, or a regulator. The businesses that fare best treat the disclosure deadline as a reason to get their automated decisions defensible, not merely described.
What the Preparation Work Buys You
Does this obligation apply to you?
Two thresholds decide whether an organisation is captured at all. First, is the organisation an APP entity, meaning most businesses with annual turnover above three million dollars, plus many smaller ones in health, credit reporting and related fields. Second, does it make automated decisions that meet the significant-effect test. The decision guide below is a starting filter, not legal advice, but it captures the practical logic most businesses need to work through.
Are You Captured by APP 1.7?
The most dangerous answer is the last one. An organisation that cannot say with confidence whether it makes automated decisions is not thereby exempt; it is simply exposed and unaware. Regulated sectors such as financial services tend to have the clearest exposure because automated decisioning is core to lending, insurance and advice, but the reach of APP 1.9 means retailers, labour-hire firms, health providers and membership organisations should all check rather than assume.
A realistic path to 10 December 2026
There is enough time to do this well, and not enough to leave it. The work divides cleanly into discovery, assessment and implementation, and the discovery phase is the one businesses consistently underestimate. Mapping automated decisions across a midsize organisation with several systems and a few third-party tools is weeks of work, not days, because the knowledge is distributed across teams and some of it lives only in vendor configurations.
Getting Ready in the Time Remaining
The sequencing matters. Disclosing before remediating means publishing an accurate description of practices you have not yet made defensible, which is worse than useless. The point of doing discovery and assessment first is that remediation, tightening human review, documenting model inputs, adding fairness checks, happens before the practice becomes a matter of public record. By the time the privacy policy is updated, the underlying decisions should already be ones the organisation is comfortable defending.
For businesses that lack the internal capacity to run this end to end, this is a defensible reason to engage outside help, and a good test of whether an advisory partner is worth the money. The right partner accelerates discovery, brings a tested assessment framework, and hands back an organisation that understands its own systems. The wrong one produces a compliance document and leaves the underlying decisions exactly as opaque as before. If you are evaluating providers, our guide on how to choose an AI consultancy in Australia and the broader AI consulting buyer's guide both apply directly to this kind of engagement.
Four mistakes that turn a manageable task into a scramble
The organisations that struggle with this deadline tend to make the same avoidable errors, and each one is easy to see coming once named.
The first is assuming that a human somewhere in the process takes the decision out of scope. It usually does not. If a person's role is to confirm what the system already decided, rather than to weigh the inputs and reach an independent conclusion, the decision remains substantially automated in the sense the Act cares about. Businesses that lean on token human review as a compliance shield are the ones most likely to be caught out, because the presence of a human is exactly the fact pattern regulators have signalled they will look behind.
The second is treating third-party tools as someone else's problem. A recruitment platform that ranks candidates, a credit-scoring API, a pricing engine supplied by a vendor: the organisation that deploys these to make decisions about people carries the transparency obligation, not the vendor. Outsourcing the tool does not outsource the accountability, and vendor contracts written before 2024 rarely give the buyer the information now needed to disclose accurately. Getting that information may require going back to suppliers, which takes time.
The third is scope creep in the wrong direction, disclosing everything out of caution. Over-disclosure sounds safe but creates its own problems: it invites scrutiny of processes that are not actually automated decisions, and it dilutes the disclosures that matter. The obligation is specific, and a precise disclosure of the decisions genuinely in scope is both more compliant and more defensible than a vague catch-all.
The fourth is leaving the board out until the end. The automated decisions in scope are, almost by definition, decisions that significantly affect people, which makes them a governance matter as much as a privacy one. A board that first learns what automated decisions the organisation makes when it is asked to approve the disclosure has been poorly served. The mapping work should reach the board well before the deadline, as assurance rather than surprise.
How to write the disclosure itself
Once the discovery and assessment work is done, the disclosure is genuinely straightforward, and it helps to know what a good one looks like. APP 1.7 asks for the kinds of personal information used in the automated decision-making technology, and the kinds of decisions made solely or substantially by it. The word kinds is doing useful work: the obligation is to describe categories, not to publish the model or its logic. A lender does not have to disclose its scoring algorithm; it has to disclose that it uses information such as income, expenses and credit history to make automated lending decisions in defined circumstances.
Plain English matters more than legal precision here. The audience for a privacy policy disclosure is the individual affected, not a court, and the OAIC's consultation through 2026 emphasised comprehensibility. A disclosure that is technically accurate but incomprehensible to an ordinary reader misses the purpose of the reform even if it satisfies its letter. The test to apply is whether a customer reading the policy would come away genuinely understanding that automated decisions are being made about them, what information feeds those decisions, and what kinds of decisions they are. If the answer is yes, the disclosure works. This is the same clarity discipline that good AI strategy work brings to every customer-facing system, not just the privacy policy.
Where this sits in the bigger regulatory picture
The automated decision-making rules do not stand alone. They sit alongside the December 2025 National AI Plan, the National AI Centre's Guidance for AI Adoption and its six essential practices, and directors' existing duties under the Corporations Act. An organisation that treats the 10 December deadline as an isolated privacy task will do the minimum and miss the point. One that treats it as the concrete first instance of a broader shift, toward being able to explain and defend how AI makes decisions about people, will be positioned for whatever comes next.
That broader shift is why the smartest response is not defensive. Businesses that can demonstrate, credibly, that their automated decisions are mapped, fair and explainable will hold an advantage over competitors still hoping the question does not come up. Transparency, done properly, is a trust asset. It signals to customers, regulators and boards that the organisation knows what its systems do. In a market where AI scepticism is rising, that is worth more than the compliance box it ticks. It is also the foundation that any serious AI consultancy engagement should be building toward, well before a deadline forces the issue.
The deadline is fixed. The systems are already running. The only real variable is whether an organisation reaches 10 December 2026 having genuinely mapped and improved its automated decisions, or having merely described them. The first is governance. The second is exposure with better paperwork.