Automated decision making is where a computer program an organisation has arranged for uses a person’s personal information to make a decision, or to do something substantially and directly related to making one, where the decision could reasonably be expected to significantly affect their rights or interests. Covered organisations must describe those decisions in their privacy policy from 10 December 2026. This guide shows where to look in Power Automate, Copilot, CRM and Excel.
Table of Contents
What counts as automated decision making under the Privacy Act?
A computer program counts when three things are true. The organisation has arranged for it 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. Personal information about that person is used in the program’s operation.
Those three conditions are APP 1.7, which the Privacy and Other Legislation Amendment Act 2024 added to the Privacy Act. The Act applies the obligation to decisions made from 10 December 2026, including decisions made by systems set up before that date.
“Computer program” is a wide term. The OAIC’s APP 1.7-1.9 transparency obligation fact sheet defines it to include pre-programmed rule-based processes, AI and machine learning, everyday software and apps, plus generative AI such as chatbots. Because the Act also treats refusing or failing to make a decision as making one, a flow that drops a request without a reply needs the same check as one that rejects it.
This guide isn’t legal advice. It explains the test and where to look in your systems, while your lawyer settles your policy’s wording. Our AI for Australian business guide covers the wider privacy and governance picture for AI projects.
Does a person signing off take a decision out of scope?
Not by itself. The OAIC’s fact sheet says a decision may be within scope even where a program’s output is subject to human review. The question is whether the program is a key factor in the person’s decision and directly connected to it. A manager clicking approve at the end doesn’t settle that.
In the OAIC’s own example, managers of a sales team use a generative AI program to suggest bonuses and salary adjustments. A director approves every final decision in writing, yet the OAIC still treats the AI output as substantially and directly related to the decision because managers use it as a key factor in each remuneration decision.
The same example shows what could change the answer. If staff interrogate the outputs, review the underlying performance indicators, narrow the program’s parameters, use more evidence and document why they departed from a recommendation, that could indicate the output was not substantially related.
Human review matters to the public as well as the regulator. In the OAIC’s Australian Community Attitudes to Privacy Survey 2026, 81% of Australians named the right to have a human review decisions as an essential condition before an organisation uses AI to make a decision that might affect them.
We see the sign-off question in our own work. For a NSW health and community services provider, we built a flow that approves or rejects a client’s service request. A staff member signs off before the outcome takes effect. What matters is whether staff treat the flow’s outcome as a key factor, how often they depart from it and whether they record why.
Because the OAIC’s bonus example is about staff pay, private sector employers should also check the OAIC’s employee records exemption. The exemption can cover an employer’s handling of an employee record where it is directly related to a current or former employment relationship, but not, for example, records about unsuccessful job applicants or volunteers.
What are examples of automated decision making in business?
Decisions the OAIC would generally consider in scope include recruitment screening, approving or rejecting credit, assessing insurance eligibility, prioritising health or disability services, AI-ranked performance and bonus outcomes and personalised pricing of significant goods by online retailers. Any of them can sit inside software a business configured itself.
Eligibility decisions are where the public is least accepting. Only 25% of Australians in the same OAIC survey view AI use for automated eligibility or risk-based decisions, such as loan approvals or benefit eligibility, as acceptable.
The table below sets common business decisions beside the systems they often run in and how each one sits against the test.
| Decision | Where it often runs | How it sits against the test |
|---|---|---|
| Shortlisting job applicants | Recruitment software and screening tools | On the OAIC’s in-scope list |
| Approving or refusing a credit application | Approval flows and finance rules | On the OAIC’s in-scope list |
| Prioritising who gets a health or disability service | Triage spreadsheets, intake flows and agents | On the OAIC’s in-scope list |
| Ranking staff for bonuses | AI-generated performance reports | On the OAIC’s list (see the exemption above) |
| Setting a price for each customer | Pricing rules and online stores | On the OAIC’s list for significant goods |
| Sending a form reminder | Scheduled flows | Usually decides nothing about the person |
The pricing row comes from the OAIC’s example of an online store for essential goods that prices by postcode. In that example, the pricing is significant whether the price difference is small or large because small differences can add up over time.
Where do automated decisions hide in Power Automate, Copilot, your CRM and Excel?
Automated decisions can hide in steps nobody calls a decision: an approval flow’s condition, a lead score that decides who is offered a service, an agent that routes an enquiry, a spreadsheet formula that ranks clients. Each system below has its usual hiding places.
Power Automate approval flows
Microsoft’s approvals guide describes the Start and wait for an approval action. It waits for the approvers’ response before the flow completes its run. The steps after it then act on the outcome.
The decision often sits in the branches around the approval. A condition that rejects a request before anyone sees it, or skips the approver when an amount is under a limit, is the program deciding. Microsoft caps a flow run at 30 days in its Power Automate limits. An approval still waiting at that point times out. A flow can also give an approval its own, shorter expiry with a branch that handles non-responses. If that branch approves or rejects the request, the program is deciding too. A request left without a reply when a run times out needs the same check as one the flow rejects. Auto-approving low-risk requests, a pattern described in our guide to AI in Power Automate, is the kind of step to put through the APP 1.7 test.
Microsoft Lists approvals
An approval triggered when a new item is added to a list in Microsoft Lists is one of the examples in Microsoft’s approvals guide. Check the conditions that run before and after the approval as well as the approval itself.
Copilot Studio agents and generative AI
Copilot Studio agents that triage an enquiry, check eligibility or decide who to escalate to are candidates to check. So is a generative AI assistant used to draft a recommendation about a person, such as a service plan. The OAIC considers that machine learning or generative AI outputs used in significant decisions would generally fall within scope unless subject to extensive human oversight and control.
CRM lead scoring
Dynamics 365 Sales, HubSpot and Salesforce can all score leads. In Dynamics 365, Microsoft’s predictive lead scoring guide says the model assigns leads a score between 0 and 100 from signals on the lead and related records such as contact and account. HubSpot scores are built from rules you set. With Marketing Hub Enterprise, contact scores can also be built with AI. The lead scoring overview on HubSpot’s knowledge base includes an example workflow that assigns an owner to unworked contacts scoring over 50. Salesforce’s Einstein scoring module describes AI lead scoring that analyses historical lead data to predict which leads are likely to convert.
None of these sources says a lead score is in scope. Whether one meets the test depends on what the score changes for the person. A score that only sorts a sales queue raises a different question from one that decides who is offered a price, a product or a service. Where scoring spans a CRM and other systems, our CRM integration work starts by mapping where each score is used.
Excel spreadsheets
Spreadsheets are the subject of the OAIC’s first worked example. An aged care provider’s formula scores clients’ need for care on their personal and health details and ranks them. Staff use the ranking to decide whom to contact for appointments or aged care support. The OAIC treats that as in scope. It adds that using the ranking to decide whom not to contact would be in scope too.
How do you find the automated decisions in your own systems?
Start from the systems, not the policy. List every system that touches personal information, find the steps that make or inform a decision about a person, then put each step through the OAIC’s questions. Record what you find in the terms APP 1.8 uses. The policy can then be written from that record.
- List every system that uses personal information, including software you bought or configured.
- In each one, list the automated steps that approve, reject, rank, score, route, price or recommend something for a person. Include steps that drop a request without a reply.
- Put each step through the questions in the OAIC’s transparency obligation flowchart: did you arrange for a program to inform or make the decision, does it use personal information, is it substantially and directly related to the decision and could the decision reasonably be expected to significantly affect the person’s rights or interests?
- For each step in scope, record the kinds of personal information it uses and the kind of decision it makes.
- Note whether the program makes the decision on its own or does something related to a decision a person makes. APP 1.8 asks for the two kinds separately.
- Review the list whenever a flow, agent or scoring model changes.
For Microsoft systems, Power Platform’s own inventory gives a head start. The Power Platform inventory page on Microsoft Learn describes a unified view of the agents and flows built on Power Platform across an organisation. Copilot Studio agents and Power Automate cloud flows are both covered. Seeing the whole inventory needs a Microsoft Entra role such as Power Platform administrator or Global reader.
Classic chatbots are left out of the inventory. The Power Platform admin centre lists them separately under Copilot Studio. Desktop flows, spreadsheets and CRM scores sit outside the inventory altogether. For those, ask each business team which workbooks, scores and AI tools they rely on, including any AI assistant used to draft recommendations about people. Because the inventory can’t tell you what each flow or agent decides, someone who knows the process still has to read the conditions.
What must a privacy policy say about automated decision making?
Among Australia’s privacy policy requirements, APP 1.8 asks for three things: the kinds of personal information the programs use, the kinds of decisions made solely by a program and the kinds of decisions where a program does something substantially and directly related to the decision. The word “kinds” matters. It points to categories, not a list of every flow.
On detail, the fact sheet lets you group decisions and kinds of information, as long as the result is meaningful to a reasonable person. Where sensitive information such as health details is used, that should be clear. Commercially sensitive information and trade secrets about a program are excluded from what you must disclose. Information whose disclosure would expose an organisation to ridicule, embarrassment or public criticism falls outside that exclusion.
The same fact sheet includes two worked examples of policy wording, each a section headed “Decisions made by computer programs”. Start from those with your lawyer instead of a blank page.
Who is responsible when a vendor’s software makes the decision?
In most cases, you are. The OAIC generally expects the obligation to stay with the organisation using personal information to make the decision. Buying software, configuring or customising it and relying on its advice can all count as arranging for a program.
The fact sheet puts expectations on both sides. Providers should give clear, high-level information about how their software can be used to make decisions. Where a bought machine learning program provides predictive analytics or personalisation in a decision that significantly affects a person, you should disclose how it processed personal information, based on the information the vendor provides. These are the questions we put to vendors:
- Does any feature we use score, rank, grade or prioritise people, such as predictive lead scoring?
- Which personal fields does that score use? Can we see them?
- Is the model trained on our own records? Does it retrain on its own?
- Can an approval or routing rule act without a person seeing the outcome first?
- What does the product log about each decision? How long is that log kept?
How should new flows and agents be built so they are easy to disclose?
Build the record as you build the flow. Note which personal fields each decision step reads, log what it decided and why, put the human check where it can change the outcome and give every decision step an owner. A flow built that way already holds most of what APP 1.8 asks a policy to describe.
Where a person reviews the output, design the review so they can see the inputs and disagree. Make a reason field required when they do. Documenting why they depart is one of the practices in the OAIC’s bonus example above. Our Power Automate consulting builds approvals and logging this way from the first version.
Frequently asked questions
Does the small business exemption apply to automated decision making?
It can. As part of APP 1, the obligation binds only the organisations and agencies the Privacy Act covers. Under the OAIC’s small business guidance, a small business is one with an annual turnover of $3 million or less. The same guidance lists businesses the Privacy Act covers regardless of turnover, with health service providers first on that list.
Does APP 1.7 apply to systems set up before December 2026?
Yes. The Act applies the obligation to decisions made from 10 December 2026, whether the program was arranged before or after that date. It also applies whether the personal information was collected before or after that date. An approval flow built years earlier that makes a qualifying decision on 11 December 2026 is caught the same way as a new one.
Do chatbots and generative AI count as computer programs?
They can. Generative AI used to generate text, images, video or code falls within the OAIC’s definition of a computer program, chatbots included. A chatbot that only answers questions about opening hours makes no decision about anyone. One that decides whether an enquiry qualifies for a service, or what to offer, needs checking against the test.
Do you have to list every flow in the privacy policy?
No. APP 1.8 asks for the kinds of personal information and the kinds of decisions. Grouping is allowed under the OAIC’s guidance as long as the detail is meaningful to a reasonable person. Keep the full system-by-system list internally. It is how you keep the policy accurate as flows change.
What happens if a privacy policy leaves this out?
The Privacy and Other Legislation Amendment Act 2024 lists APP 1.7 in section 13K of the Privacy Act from 10 December 2026. Section 13K is a civil penalty provision. The OAIC can also issue an infringement notice or a compliance notice for a breach of it. An infringement notice for a single contravention can be up to 60 penalty units for a body corporate. The amount for a listed corporation is 200 penalty units instead. Those amounts come from the OAIC’s infringement notice guidance.
Find your automated decisions before you rewrite the policy
Many automated decisions were never designed as decisions. They sit in flow conditions, CRM scores or workbook formulas that someone built to save time. Finding them is technical work that comes before the legal wording.
OfficePro Consulting runs systems reviews that find the automated decision steps in Microsoft 365, Power Platform or CRM systems and record them in the terms APP 1.8 uses, ready for your lawyer to assess. Start by booking an automated decision systems review.



