From 10 December, your privacy policy has to explain what your software decides

For the last few years, the honest answer to "what are the rules on using AI in Australia?" has been "mostly guidance". That changes on 10 December 2026.

From that date, businesses covered by the Privacy Act must include new information in their privacy policies whenever they have arranged for a computer program to make, or help make, decisions that could significantly affect people. The OAIC published its final guidance, a fact sheet and a flowchart on 30 September. There is no grace period: from 10 December, a privacy policy that says nothing about in-scope automated decisions is out of date.

It sounds like a paperwork change. In practice it's the first time many businesses will have to write down, publicly, where software is shaping decisions about their customers, applicants and staff. Plenty will discover they don't know.

What actually has to change

New APPs 1.7 to 1.9 apply when three things are all true. A computer program makes a decision, or does something substantially and directly related to making it. The decision could reasonably be expected to significantly affect an individual's rights or interests. And personal information about that individual is used in the program's operation.

Where that's the case, the privacy policy must describe three things:

• the kinds of personal information used by those programs
• the kinds of decisions made solely by a program
• the kinds of decisions where a program does something substantially and directly related to the decision.

It's about kinds, not a catalogue. The APP 1 guidelines allow related decisions to be grouped, ask for plain language rather than technical detail, and don't require commercially sensitive information about how a system works. What they do require is enough for a person to understand that software is involved and, ideally, how to question an outcome.

Who it applies to

The obligation applies to APP entities. For most private businesses that means annual turnover above A$3 million, plus some smaller businesses covered regardless of size: private health service providers, businesses that buy or sell personal information, credit reporting bodies, Commonwealth contractors, and businesses related to a covered entity, among others. The small business exemption hasn't been removed. If you're under the threshold and none of the exceptions apply, this obligation doesn't bind you yet, although your larger customers' obligations may still reach you through their contracts.

Three things businesses will get wrong

1. "A human reviews it, so it's not automated"

This is the most common assumption, and the OAIC's guidance rejects it. A decision is in scope if the program does something that is a key factor in the decision and directly connected to it, even when a person makes the final call. Scoring, ranking, triaging, recommending and flagging can all qualify. The guidelines suggest asking how much the decision-maker relies on the output, how often they override it, whether it's advisory or effectively determinative, and how deeply it's built into the process.

A human in the loop is still good practice. It just doesn't take the decision out of the privacy policy.

2. "It's the vendor's software, so it's the vendor's problem"

The obligation sits with the entity that arranged for the program to be used. The guidelines' examples are the kind of things many businesses now buy off the shelf: AI that screens job applications, a chat assistant used to draft performance assessments that feed promotion decisions, a third-party tool that auto-approves refunds, a case management system that escalates complaints automatically. The OAIC expects businesses to identify and oversee these programs during and after procurement, and to make sure contracts are clear about who retains control over decisions.

That has a practical consequence. You can't describe decisions in your privacy policy if you don't know which of your tools are making or shaping them.

3. "We don't use AI for decisions, just simple automation"

"Computer program" is read broadly. The guidelines include rules-based systems and spreadsheet formulas alongside machine learning, generative AI and chatbots. Ordinary tools like word processing aren't caught. But an automated rule that suspends an account, sets a price for an individual, ranks applicants or decides who gets a call back could be, if the effect on the person is significant.

What "significantly affect" looks like

The threshold is more than trivial, and it's judged by the effect on the person, not how sophisticated the software is. The guidelines point to decisions about access to credit, employment, housing, insurance, healthcare, essential services such as telecommunications and utilities, and differential pricing. The more vulnerable the people affected, the more likely an effect is significant.

These are illustrations, not legal tests. The edge cases are where you need advice.

Technology businesses have two jobs

If you build or sell software, this change lands twice. You have your own privacy policy to update, covering your hiring, sales, pricing and support tools. And your Australian customers now have a reason to ask you, in detail, what your product decides.

Expect it in security reviews and procurement questionnaires, which are already adding AI sections. A-LIGN's 2026 Compliance Benchmark Report found four out of five organisations now field direct customer questions about their AI risk management. If your product scores, ranks, flags, approves or prices anything about your customers' customers or staff, have a short, plain-language description ready, covering what the feature does, what personal information it uses, whether a human can override it, and what control the customer keeps. Vendors who can hand that over on the first ask will close faster than those who can't.

A plan for the next seven weeks

There's enough time to do this properly if you start now.

1. Find everything. List every tool, feature and workflow that uses personal information to make or inform a decision about a person, including vendor tools and anything teams adopted without approval. A simple AI register works.
2. Screen it. For each one, ask whether the program's role is substantial and direct, and whether the effect on the person could be significant. Record the reasoning either way.
3. Ask your vendors. For in-scope tools you buy, get a description of what the software decides and what data it uses. Check contracts for decision control and notice of changes.
4. Write it plainly. Group related decisions, describe the kinds of personal information, and separate decisions made solely by software from those it substantially assists.
5. Give people a way to ask. It isn't strictly required, but a clear path to question an outcome or ask for human review makes the disclosure credible and lines up with the national AI guidance.
6. Get it reviewed, then keep it current. Have your lawyer review the wording before 10 December, and make the register the trigger: any new tool that makes or shapes decisions about people prompts a privacy policy check.

This is a governance question, not a drafting one

The disclosure is the visible part. The real work is knowing where software is shaping decisions about people, who owns each one, and how someone would challenge it. That's also the core of Australia's Guidance for AI Adoption, which asks organisations to keep an AI register and give people a way to contest AI decisions.

Businesses that treat 10 December as a wording exercise will update a paragraph and hope it's right. Businesses that treat it as a prompt to find and own their automated decisions will end up with a privacy policy they can stand behind, and a much better answer when a customer, investor or regulator asks how they use AI.

This article is general information only and isn't legal advice. Whether the new obligations apply to you, and how your privacy policy should describe your automated decisions, depends on your circumstances. Get advice from a privacy lawyer before relying on it. Regulatory references checked 17 October 2026.