SecurityBrief Australia - Technology news for CISOs & cybersecurity decision-makers
Australia
From 10 December, your privacy policy has to say what the software decides

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

Tue, 17th Nov 2026 (Today)
Nick Lim
NICK LIM Founder Switchboard Finance

On 10 December 2026 a quiet change to Australia's Privacy Act switches on. If a business covered by the Act uses a computer program to make, or substantially help make, a decision that significantly affects a person, its privacy policy has to say so. Many of the businesses that will be caught do not think of themselves as running automated decisions. Plenty of them run a chatbot.

What the rule says

The change, usually called the automated decision-making rule, sits in Australian Privacy Principle 1, added by the Privacy and Other Legislation Amendment Act 2024. Three things have to be true. The business has arranged for a computer program to make a decision, or to do something substantially and directly related to making one. The decision could reasonably be expected to significantly affect a person's rights or interests. And personal information about that person is used along the way.

When all three apply, the privacy policy has to set out three things: the kinds of personal information the programs use, the kinds of decisions they make on their own, and the kinds of decisions where they carry out a step that feeds the outcome.

The privacy regulator published its final guidance on 30 September. Three points stand out from the law firm summaries. A human signing off at the end does not take a tool out of scope if its output is a key factor. Generative AI and chatbots count as computer programs. And the duty sits with the business using the tool, not with the vendor that built it.

It is a transparency rule. It creates no new right to contest a decision. It does create a public document that customers, journalists and regulators can hold against what the business actually does.

A customer deserves a plain answer to what happens to what you type into a finance chat before the first message, not after a breach.

A business should also check whether the Act covers it at all. It generally applies above A$3 million in annual turnover, and to some smaller businesses as well. Any firm large enough to buy decision software should assume it is covered and confirm.

Why lending is the obvious test

Finance is on the regulator's list of interests that matter, beside health, housing and employment. A decision about credit changes what a person can buy, where they can live and whether a business makes payroll.

It also runs on the most sensitive file a small business ever assembles. Tax returns. Bank statements. Identity documents. Details of a partner's income, a tax debt, sometimes an illness that explains a bad year. That file is forwarded between owner, accountant, broker and lender like a lunch order.

More of it now meets software before it meets a person, and customers have noticed. When Forrester surveyed credit-active Australians for Experian in July, 80 per cent named their personal information being accessed or misused as the main worry about using an AI agent.

The file was always sensitive. What is new is the number of machines that read it.

What a chat window collects without asking

A form asks for six fields and gets six fields. A chat box gets whatever the customer decides to type, and people explain money the way they talk. They name their bank. They mention the tax office. They describe a former partner, a parent's guarantee, a creditor who has stopped being patient.

None of that was requested and all of it is now held. Often it has also gone to a model provider in another country to produce the reply. The policy should already cover that cross-border disclosure. From December it may also have to say what the software does with the information.

Whether a given chatbot is caught turns on what happens next. An assistant that answers general questions and decides nothing sits at one end. An assistant whose output ranks an applicant, sets a price or filters who gets a call back sits at the other, and a person approving its work at the end does not move it.

A list for the weeks that are left

Map the tools. List every piece of software that scores, ranks, triages or answers customers. Include the ones a vendor runs for you.

Trace each one to a decision. If its output is a key input to who gets credit, a contract, a price or a service, write the three disclosures in plain words. If it is not, write down why, and keep the note.

Put a line at the chat itself. Say it is AI, that the conversation is recorded, and what not to type. A privacy policy nobody opens does not protect anyone at the moment they are typing an account number.

Strip what you do not need. Remove names, phone numbers, email addresses, street addresses and long numbers from transcripts before they are stored or sent to a model. Data that was never kept cannot leak.

Ask the vendor three questions. Where is the information processed? How long is it kept before it is deleted or de-identified? Is it used to train anything?

Decide what the software must not decide. Then test that it does not.

The businesses that find this easy will be the ones that already know what their software does. The rest have until 10 December to find out.

This piece is general commentary from the lending side, based on published guidance and summaries of it as at November 2026. It is not legal advice, and a business unsure whether the rule applies should ask a privacy lawyer.