AI Regulation

Privacy Act, from 10 December 2026

Is my system covered? The APP 1.7 checklist

Three limbs, one entity test and a handful of regulator examples decide whether a system belongs in your privacy policy. Here they are in the order you should run them.

Responsible AI Australia • 14 Oct 2026 • 7 min read

From 10 December 2026, Australian Privacy Principle 1.7 requires a privacy policy to disclose the kinds of decisions a computer program makes or substantially shapes about people, and the kinds of personal information it uses. The full guide explains the law. This piece is the shorter thing you need at your desk: a checklist you can run on each system, with the regulator's own examples of what falls in and what falls out. It is general information, not legal advice, and the judgement calls belong to your privacy adviser.

Step one. Does the Privacy Act cover your organisation?

The disclosure only binds APP entities. If none of the triggers below applies, you sit under the small business exemption for now and the checklist stops here, with two caveats: your larger customers are not exempt and will ask you the same questions in supplier reviews, and the Government has said it intends to remove the exemption in a later tranche of reform. The list is the OAIC's.

Annual turnover of more than $3 million. In the current or any previous financial year since the exemption was introduced.
An Australian Government agency.
A private sector health service provider that holds health information. Includes medical, allied health and complementary practitioners, gyms and weight-loss clinics, child care centres, private schools and private tertiary providers.
A business that sells or purchases personal information.
A credit reporting body.
A contracted service provider for an Australian Government contract.
Accredited under the Consumer Data Right system.
An employee association registered under the Fair Work (Registered Organisations) Act 2009.
Opted in to the Privacy Act.
Related to a business that is covered by the Privacy Act.
Operates a residential tenancy database, or is an AML/CTF reporting entity. The Act covers those particular activities even for a small business.

One tick is enough. Health service providers deserve a second look, because the OAIC reads that category to include gyms, child care centres and private schools, not only clinicians.

Step two. List every system that touches a decision

The three-limb test is run per system, so the inventory comes first. Cast the net wide and narrow it with the test, not the other way around. The obligation is not limited to artificial intelligence, so a scoring formula in a spreadsheet belongs on the list alongside the machine learning model.

Anything that makes, recommends, scores, ranks, triages, flags or escalates something about a customer, an applicant, a patient, a student or an employee.
Tools staff use informally, including general-purpose AI assistants used to draft assessments, shortlist candidates or summarise complaints.
Automation a vendor switched on inside a product you already use, which you may only find by asking the vendor.
Systems that refuse or fail to make a decision, and systems that approve people automatically. APP 1.9 counts both.

Limb one. What does the program do in the decision?

APP 1.7(a) is satisfied where you have arranged for a computer program to make a decision, or to do a thing that is substantially and directly related to making it. Pick the option that fits each system.

It makes the decision itself, with no person deciding. Automatic approval or decline of an application, refund or claim. This is the “solely” list in your policy.
A person decides, but the program's output is a key factor. It scores, ranks, triages, recommends, flags or summarises in a way that steers the decision. This is the “substantially and directly related” list, and a human signing off does not take you out of it.
It only does something incidental. Stores records, formats documents, calculates an age from a date of birth. Directly related, but not substantially. Out of scope on this limb.
It is not used in decisions about individuals. Out of scope.

The regulator's example draws the line. A spreadsheet formula that scores and triages callers to a crisis hotline, where the score is a key factor in a person deciding whose call to attend first, is substantially and directly related to the decision. The same spreadsheet used only to work out an age from a date of birth is not.

Also check who arranged the program. APP 1.7 fixes the obligation on the entity that arranged for the program to be used, even where another company operates it. Procuring a recruitment platform, directing staff to use an AI tool for performance reviews, or contracting a company to approve refunds automatically all count as arranging. Merely developing or hosting software that your customers use does not.

Limb two. Could the decision significantly affect the person?

APP 1.7(b) asks whether the decision could reasonably be expected to significantly affect the rights or interests of an individual. Beneficial effects count as much as adverse ones. The Explanatory Memorandum and the OAIC's examples point to these kinds of decisions.

Eligibility for a service, product, benefit or contract.
Pricing or terms, including differential pricing for significant goods or services.
Employment, shortlisting, promotion or performance outcomes.
Credit, insurance and other financial decisions.
Access to accommodation, education or health services.
Targeting of content or advertising where it changes what a person pays or limits their opportunities.

If the effect on the person is trivial, meaning the decision does not change what they can access, pay or do in any meaningful way, the system fails this limb and drops out. If you are unsure, treat the system as in scope for the draft and confirm the call with your adviser. The OAIC has not settled every edge, including how large a price difference has to be before it counts.

Limb three. Is personal information used in the program?

APP 1.7(c) requires that personal information about the individual is used in the operation of the program to make the decision or do the related thing. A program that works only on aggregate or non-personal data, with nothing about the affected person fed in, fails this limb. For the systems that pass, note the kinds of personal information, because they become the third list in your policy.

Identity and contact details.
Date of birth or age.
Demographic details such as gender, nationality or language.
Financial information such as income, assets or bank details.
Credit information.
Transaction and purchase history.
Employment history, qualifications and references.
Health information (sensitive information).
Biometric information (sensitive information).
Location data.
Online identifiers and device data.
Behavioural or usage data.
Images, video or voice recordings.
Government identifiers such as driver licence or passport numbers.

Worked examples, in and out

Each verdict below applies the three limbs to a common system. They are illustrations built on the regulator's examples, not rulings on your facts.

Likely in scope
A recruitment platform ranks applicants and a recruiter interviews the top of the list. Assists a decision, affects employment, uses employment history. Goes in the “substantially assists” list.
Likely in scope
A returns system approves refunds under a threshold with no person involved. Makes the decision, affects what the customer is paid, uses transaction history. Goes in the “solely” list.
Likely in scope
A fraud model flags transactions for an analyst to hold or release. Assists, affects access to funds, uses transaction and device data.
Likely in scope
A case management system escalates particular complaints on its own. Makes a decision about handling, affects the complainant's outcome, uses their details.
Likely in scope
Staff are directed to use an AI assistant to draft performance assessments that feed promotion decisions. Assists, affects employment, uses employee records.
Likely out of scope
A spreadsheet calculates a customer's age from their date of birth for a form. Incidental on limb one.
Likely out of scope
A demand forecasting model works on aggregate sales data with no individual's information. Fails limb three.
Likely out of scope
A chatbot answers questions about opening hours and policies but cannot approve, decline or escalate anything. Not used in decisions about individuals. Watch for the product update that changes that.
Likely out of scope
A document summariser condenses internal reports with no bearing on any decision about a person. Not used in decisions about individuals.
Likely out of scope
A payroll system calculates pay from timesheets under fixed rules. Usually incidental. If it also decides who is flagged for investigation, re-run the test on that feature.

Step three. Record the answer, especially the ones you leave out

The disclosure itself is three lists in the privacy policy. The record behind it is what you will want when the OAIC asks, when a supplier changes a feature, or when a new system arrives. For every system on the inventory, keep:

The system, its purpose and who arranged it.
Your answer on each limb and the reason, in a sentence.
The kinds of decisions and kinds of personal information, in plain words.
Whether a person can review the decision and how to ask.
A date, and a trigger in procurement and change management to revisit it.

The free builder produces this register and the privacy policy section from the same answers, in your browser, with nothing stored. For the wording itself, shaped to recruitment, lending, insurance, health, education, customer service and government, see disclosure examples by sector.

What the checklist does not decide

Whether a decision is lawful, fair or non-discriminatory is a question for consumer law, anti-discrimination law, credit reporting rules and the Australian Consumer Law, which apply to automated decisions as they do to human ones. APP 1.7 to 1.9 only decide what your privacy policy must say. How you govern the systems is what Responsible AI certification assesses, against the Australian AI Ethics Principles, with the result on the public register.

Questions people ask

Does a human in the loop take my system out of scope?

Not on its own. If the program's output is a key factor in the decision, for example a score, ranking, recommendation or flag that steers what the person decides, the arrangement is substantially and directly related to making the decision and APP 1.7 applies. Human review only takes a system out where the program's role is genuinely incidental.

Is a spreadsheet really a computer program for this purpose?

Yes. The Explanatory Memorandum says computer program takes its ordinary meaning and covers pre-programmed rule-based processes as well as AI and machine learning. The OAIC's own example of an in-scope system is a spreadsheet formula that scores callers to a crisis hotline.

We are a software vendor. Do we disclose our customers' decisions?

Generally no. The obligation sits with the entity that arranged for the program to be used to make or assist decisions about people, which is usually your customer. A company that develops and hosts software its customers use, but does not use it itself, is not the one who arranged the use. Expect your customers to ask you how the product uses automation so they can write their own disclosure.

What if I am not sure whether a decision is significant?

Treat the system as in scope for the draft so nothing is missed, record why you were unsure, and confirm the call with your privacy adviser. The regulator's examples of significant decisions include eligibility for services, pricing and terms, employment, credit and insurance, and access to accommodation, education or health services. Beneficial effects count too.

Do we have to list every decision and every data field?

No. APP 1.8 asks for kinds of decisions and kinds of personal information, and the OAIC frames the extent of disclosure as a balance: enough for people to understand how automated decision-making is used, without detail so excessive that it obscures the point. Commercial-in-confidence information about how a system works is excluded.

What should small businesses under the exemption do?

Nothing is required today, but running the checklist costs little and pays twice: larger customers covered by the Act increasingly ask suppliers the same questions, and the Government has said it intends to remove the small business exemption in a later tranche of privacy reform. Writing the register now is easier than reconstructing it later.

Go deeper

Sources

This guide is general information, not legal advice. It restates official instruments and regulator guidance as they stood on 14 October 2026. How a rule applies to a particular organisation is a judgement for its own adviser.