On 30 September the privacy regulator told Australian businesses how it will read the automated decision-making rule that starts on 10 December, and the answer is wider than most expected. The OAIC’s new fact sheet and flowchart say a spreadsheet formula counts, a chatbot counts, software you rent from a vendor counts, and a manager signing off on the output does not automatically take you out of scope. If your systems score, rank, price, screen or shortlist people, you probably have privacy policy work to do in the next 68 days.

Digital One Agency runs that check as a working session: we walk the OAIC flowchart against the software you actually use and tell you which systems need disclosure, and which need changes so they can stand up to a question. You can book a meeting to start.

We covered the basics of the rule when it was still a date on the calendar, in our earlier piece on the 10 December privacy policy change. This post is about what changed this week: the regulator has now published its own examples, and they move the line.

A spreadsheet formula is a “computer program”

For APP 1.7 to 1.9, a computer program includes pre-programmed rule-based processes, AI and machine learning, ordinary “software, apps or word-processing tools”, and generative AI used to produce text, images, video or code, chatbots included.

Remember the first worked example. A fictional health and aged care provider uses a spreadsheet to score clients on date of birth, address, marital status, blood pressure, medication and other details, and the sheet ranks them automatically. Staff then call people in rank order. The OAIC says that is in scope, because the formula determines who gets services.

It goes further. If staff used the same ranking to decide who not to call, that is still in scope, because a decision includes refusing or failing to do something.

Most businesses would never list a spreadsheet in a software inventory. That is the trap. Tools that quietly sort customers, tenants, applicants or patients are often the oldest and least documented things in the building: a lead-scoring rule in the CRM, a “priority” column with a formula behind it, an old macro. Those are programs now.

“A human signs off” is not the exit you think it is

The usual objection to this rule is that a person always makes the final call. Regulators saw that coming.

The fact sheet says a decision can be in scope “even where a computer program output does not replace the entire decision-making process or is subject to human review.” It defines substantially as a key factor in the human’s decision and directly as having a direct connection to it. And it states that the OAIC considers machine learning or generative AI outputs used in significant decisions would generally be in scope “unless subject to extensive human oversight and control.”

Its example is a sales team of 58 whose managers use a generative AI tool to draft performance profiles and suggest bonuses and pay rises. HR has oversight, and a director approves every final decision in writing. Still in scope, because the AI’s recommendation is relied on to document the reasoning.

What would change the answer? The OAIC lists the behaviours: staff interrogating outputs, reviewing the underlying indicators, setting the tool’s parameters more narrowly, using additional evidence, and documenting why they departed from the recommendation. That list is the useful part, because every item on it is something your software either records or does not.

One practical consequence goes unstated. If your defence is “extensive human oversight”, you will need to show it. A review screen that records what the reviewer looked at, an override field with a required reason, and a log that ties the final decision to the person who made it turn a claim into evidence. Without those, the honest position is that you are in scope and should disclose. Law firm Hamilton Locke made the same point in June, writing that a human in the loop “does not necessarily take an arrangement outside the new framework” (Hamilton Locke).

Bought it from a vendor? The disclosure is still yours

Many Australian businesses do not build decision software. They subscribe to it. The flowchart covers this in one line: entities that “procure, configure, authorise, integrate or rely on” a third-party program may still be in scope.

A housing example makes it concrete. An affordable housing provider in Gippsland buys an AI tool that matches applicants to vacant dwellings. Whether it pays for the premium tier with configurable parameters or the cheap tier that only recommends, it has “arranged for” the software, and the disclosure obligation sits with the provider, not the vendor.

Vendors are not off the hook entirely. The fact sheet says providers of third-party software “should provide clear, high-level information” so their customers can write accurate privacy policies. That gives you a reasonable request to make this month: ask each vendor whose product scores, ranks, filters or prices people for a short written description of the personal information it uses and the decisions it makes or informs.

Recruitment shows how far the chain runs. An engineering firm advertises graduate roles on a job platform that targets the ad by age and gender, then uses the platform’s AI video-interview tool to score voice, facial geometry and eye movement and produce a shortlist. The OAIC says both are in scope for the firm.

Small price differences are still “significant”

If you sell online and adjust prices by postcode, demand or customer profile, read Example 4 twice.

A fictional grocery platform charges subscribers in one Sydney suburb between $0.89 and $3.72 more per item than subscribers in another, based on address. Two friends work out that one paid $132 more for baby formula over a year. The OAIC says the pricing is significant “regardless of whether the price difference is small or large”, because small differences compound and the goods are essentials.

Put that $132 in weekly terms and it is about $2.54. Nobody would notice it at the checkout, and the regulator has still called it significant. Any business running dynamic pricing, personalised offers or programmatic advertising for significant goods should assume the OAIC means it. Its list of decisions generally in scope names differential pricing for significant goods, programmatic advertising for significant goods or services, insurance eligibility, loan and credit decisions, recruitment screening, and AI reports used to rank staff or set bonuses.

What you can keep confidential, and what you cannot

Businesses with a genuinely clever model worry that disclosure means handing competitors the recipe. A lender example answers this.

A regional West Australian lender has built its own fraud and credit-scoring tool using name, employment records, credit information, biometric templates and location. According to the OAIC, the weighting the model gives each data point can be withheld as commercial-in-confidence. The fact that personal information powers fraud detection cannot, because it is standard across the industry and confers no edge.

Two rules fall out of that. Trade secrets and commercially sensitive information are excluded from what you must publish. Information that would merely expose you to “ridicule, embarrassment or public criticism” is not. One test the OAIC suggests is whether an arm’s-length buyer would pay for the information.

What the finished privacy policy section looks like

Sample wording is included, and it is shorter than many businesses fear. The lender’s version is a heading, “Decisions made by computer programs”, a list of five kinds of personal information, and a list of three kinds of decisions: account access, transaction processing, and loan or credit applications.

You may group decisions and data types. The detail must still be “meaningful to a reasonable person”, and sensitive categories such as health information or biometric templates should be named clearly rather than buried in a group. Your policy must also separate decisions made solely by a program from decisions where the program does something substantially and directly related to a human’s decision.

Getting this wrong is not theoretical. The OAIC began 2026 with its first compliance sweep of privacy policies, covering about 60 businesses across six sectors including property, pharmacies, car rental and car dealerships, and a privacy policy missing required content under APP 1.4 can draw an infringement notice of up to $330,000 for a company, according to Sparke Helmore. Whether the regulator uses that power on the new APP 1.7 content is a question for your lawyer. Either way, the sweep shows it is reading policies line by line.

Two points from our earlier post still apply. The rule covers decisions made from 10 December regardless of when the system was installed, and it applies to APP entities, which generally means turnover above $3 million plus the exceptions that pull smaller businesses in. For anyone unsure, the fact sheet’s default is sensible: take a cautious approach and disclose.

How we help, and what it costs

This is the work we do through our AI consulting services, and it splits into two parts.

First, the inventory. We sit with the people who run sales, operations, HR and finance, list every system that scores, ranks, filters, prices or shortlists people (spreadsheets and SaaS included), and run each one through the OAIC flowchart. You leave with a table: system, personal information used, decision made or informed, solely automated or not, and in or out of scope with the reason. Your lawyer then has something concrete to turn into policy wording, rather than a blank page.

Second, the fixes. Where you want to rely on human oversight, we build the review screens, override reasons and decision logs that make the oversight real and provable. Where a vendor tool is a black box, we help you write the information request and, if the answer is unsatisfactory, plan a replacement or an integration that keeps the decision in your hands. That second part is software work, and it is what our AI integration services are for. If you are in a regulated sector, our guide for APRA-regulated and critical infrastructure organisations covers the obligations that sit alongside this one.

On cost: the free AI audit is the starting point and costs nothing. Any build work after it is scoped and quoted before anything starts, so you decide with a number in front of you. We are not a law firm and do not write legal advice; we make sure the facts your lawyer works from are complete, and that your systems match what the policy says.

Your next step

Sixty-eight days is enough time to do this properly and not enough to do it twice. Before your next leadership meeting, open the OAIC flowchart, list the three systems in your business most likely to sort people, and see where each one lands.

Then book a meeting with Digital One Agency. In the first conversation we will go through those systems with you, tell you which look in scope and why, and set out what an inventory and any fixes would involve, so you can have a draft disclosure ready well before 10 December.

Frequently asked questions

Does using ChatGPT or another chatbot at work put us in scope?

Only if it is used, with personal information, to make or substantially and directly inform a decision that could significantly affect someone. The OAIC's fact sheet lists generative AI and chatbots as computer programs, so drafting a marketing email is not the issue, but using one to summarise job applicants or suggest bonuses can be.

If a manager approves every decision, do we still need to disclose?

Possibly. The OAIC says a decision can be in scope even when a human reviews the output, and that AI outputs used in significant decisions are generally in scope unless subject to extensive human oversight and control. Its example of staff interrogating outputs, checking the underlying data and documenting why they departed from the recommendation shows what that oversight looks like.

We use a SaaS product to screen applicants. Is that the vendor's problem?

No. The OAIC says the obligation generally stays with the business that arranged for the program, including by buying, configuring or relying on it. Vendors should give you clear, high-level information about how their software makes decisions, so ask for it in writing now.

Do we have to publish how our model works?

No. You must disclose the kinds of personal information used and the kinds of decisions made or informed, not the algorithm. The OAIC's lender example shows model weightings can be withheld as commercial-in-confidence, while the fact that personal information is used for fraud detection cannot.

When does this start and who does it apply to?

It applies to decisions made from 10 December 2026 by APP entities, which generally means businesses with annual turnover above $3 million plus some smaller businesses caught by exceptions. Systems installed before that date are still covered.