What APP 1.7 Requires Australian Organisations to Disclose

The question I get most often is the wrong one: do we have to disclose our AI. APP 1.7 is not an AI register. It is a privacy-policy obligation that can apply where a computer program uses personal information to make, or substantially and directly support, a decision that significantly affects the person.

The Office of the Australian Information Commissioner publishes the Privacy Act and the Australian Privacy Principles. Start there, not with a vendor slide. The statute talks about computer programs. A scoring engine that has been in the loan book for a decade can be more relevant than a chatbot launched last month.

What the policy has to carry, in substance, is two lists. The kinds of decisions. The kinds of personal information used in those decisions. Kinds, not a dump of every field and every threshold. A person who might be hired, declined, priced, flagged or restricted should be able to read the policy and recognise the decision type that touched them.

What the policy does not have to do is explain the algorithm. You are not writing a model card for the public internet. You are saying that software is in the decision, what sort of decision it is, and what sort of personal information feeds it. The technical documentation, the vendor contract and the override rules stay inside the organisation, where legal, risk and the system owner can actually use them.

Who is in scope is a legal question. The operational question that sits underneath it is simpler, and most boards have not answered it: where does software already influence an outcome about a person, using information that identifies them or can be used to identify them. Until that list exists, the privacy policy is a guess.

Human approval does not automatically take a workflow off the list. If a person clicks accept on a recommendation they have not independently assessed, the program may still be substantially supporting the decision. Treat that as a facts question, not a slogan.

10 December 2026 is a hard date on the calendar. After that date, a privacy policy that is silent on automated decisions is not merely out of date. It may be incomplete against an enforceable APP.

OAIC will keep the statute. Law firms will keep the legal opinion. The gap I see in mid-market organisations is the inventory. Privacy counsel cannot draft kinds of decisions they have not been shown. Operations cannot show them if nobody has mapped the workflows. Vendors will say they are not the APP entity. All three can be true at once, and the organisation still has a policy to publish.

The commercial cost of getting this late is not a public price list. It is compressed legal time, incomplete lists, and a board that finds out in November that the claims engine, the rostering tool and the collections scorer were never in the draft. Early mapping is cheaper than a December scramble because the facts do not change when the calendar does. Only the available hours do.

There is also a trust cost. Customers and staff already experience these decisions. If the policy is the first time the organisation admits software is involved, the disclosure reads as a surprise rather than a control.

A mid-market lender assumed APP 1.7 was an AI project. The first pass found the live decision points in a rules engine and a bureau-score workflow, not in the generative tools the digital team had just bought. The privacy policy needed those decision kinds named. The chatbot did not.

A professional-services firm used a vendor screening tool that ranked candidates. Hiring managers usually accepted the first page. The legal question was whether that ranking substantially supported the shortlist. The operational fact was that nobody could name the ranking inputs. Both had to be resolved before anyone could write a disclosure sentence.

An insurer’s claims team described a “system recommendation” that staff were expected to follow unless they wrote a reason. That is not a human decision in any useful sense until you can show the person actually reviewed the file. The disclosure question followed the workflow, not the org chart.

A retailer had personalisation and fraud flags in the same vendor platform. One affected what a logged-in customer was offered. The other affected whether an order was held. Those are different kinds of decisions. Dumping them into one sentence about “AI” would have been the dishonest version of disclosure.