AI Agents and Automated Decision-Making: Which ERP Decisions Go in Your Privacy Policy from 10 December 2026

From 10 December 2026, a business covered by the Privacy Act must say in its privacy policy which kinds of decisions a computer program makes, or substantially helps a person make, and what personal information it uses, where the decision could significantly affect someone. The OAIC’s guidance of 30 September 2026 confirms that an AI recommendation a person signs off can still count.

At a glance

• Who it is for: owners, CFOs, operations managers and privacy officers of Australian businesses that run AI agents, AI features or scoring rules beside Odoo or MYOB Acumatica and have a privacy policy to update.
• Problem it solves: working out which agents fall under the automated decision-making rule, which don’t, and what your privacy adviser needs from you to write the disclosure.
• Platforms: Odoo 20 and Odoo 19, MYOB Acumatica, and any agent or tool wired in beside them. The rule looks at the decision, so the platform matters less than what the output is used for.
• The rule: Australian Privacy Principles 1.7 to 1.9, added by the Privacy and Other Legislation Amendment Act 2024, per the OAIC’s announcement of 30 September 2026.
• Applies from: 10 December 2026, which is nine weeks from the date of this post.
• Who is covered: APP entities. The OAIC says most small businesses with annual turnover of $3 million or less are not covered, with exceptions that include health service providers and businesses that trade in personal information.
• What goes in the policy: 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.
• What it does not ask for: the OAIC’s fact sheet doesn’t say you must name the model or the vendor, and commercially sensitive detail is excluded.
• Typical cost: the OAIC’s fact sheet, flowchart and guidelines are free. Your privacy adviser quotes the policy review. If the agent inventory and audit trail work needs us, a pilot or proof of concept starts from $5,000 ex GST per our AI agents for ERP in Australia page, as at 8 October 2026.
• Typical timeline: the date is fixed. Discovery on our agent engagements runs 1 to 2 weeks per the same page; leave room after the inventory for your adviser’s review and sign-off.
• Last verified: 8 October 2026 against the OAIC’s APP 1.7 to 1.9 fact sheet, its flowchart and Chapter 1 of the APP Guidelines (version 2.0, updated 30 September 2026).

We build agents that draft and wait for a person. That sentence has been our answer to most governance questions, and it’s still the right design. It is not, on its own, the answer to this one. We’re ERP consultants, not lawyers, so read this as a practitioner’s map of where your agents sit against the OAIC’s wording, and take the wording itself to your privacy adviser.

What does the automated decision-making rule require from 10 December 2026?

The Office of the Australian Information Commissioner (OAIC) sets out three conditions in its updated guidelines. The rule applies when you have arranged for a computer program to make a decision, or to do a thing substantially and directly related to making one; the decision could reasonably be expected to significantly affect an individual’s rights or interests; and personal information about that individual is used in operating the program.

When all three are met, the privacy policy has to describe the kinds of personal information used and the kinds of decisions involved. That is the whole of APP 1.7 to 1.9 as the OAIC describes it: a transparency rule. It asks you to describe what your programs do. It doesn’t tell you to switch them off.

Three definitions in the fact sheet do most of the work for an ERP-run business:

• “Computer program” is broad. It covers pre-programmed rule-based processes, machine learning, and generative AI including chatbots. One of the OAIC’s own examples is a spreadsheet that scores and ranks people. A credit-hold rule in your ERP that has nothing to do with AI is a computer program too.
• “Arranged for” includes software you bought. Procured third-party programs and configured off-the-shelf software count, and the guidelines say the obligation typically stays with the business using the output, not the vendor.
• “Significantly affect” means more than trivial. The fact sheet’s list includes approving or rejecting loans or credit, differential or personalised pricing of significant goods, and recruitment software. A wrong product recommendation doesn’t make the list.

Does a drafts-first agent with a human approver still count?

Often, yes. This is the part that will surprise businesses that did the careful thing and put a person on every write.

The fact sheet says advisory and determinative outputs both qualify, and that human review does not automatically take a decision out of scope. The test is whether the program’s output is “a key factor in facilitating the human’s decision making”. The guidelines list what to weigh: how much the person relies on the output, whether they can override it, how transparent the output is, and how tightly it’s built into the process.

Two of the OAIC’s worked examples are uncomfortably familiar. In one, a loan officer relies heavily on a risk score and almost always follows it. In another, staff at an online marketplace can override account suspensions and rarely do. Both are in scope. The fact sheet’s generative AI example goes the other way: staff who interrogate the output, use other evidence and document why they diverged may be able to show the output wasn’t substantially related to the decision.

“An approval button is a control, not an exemption. If your credit controller says yes to the agent 96 times in 100 without changing a word, the OAIC’s own examples say the program is doing a real share of the deciding. Better to write that down yourself than have someone else point it out.”

Josh Craig, Director, Auboros

This is where an audit trail earns its keep. Every agent we run logs what it drafted, who approved it, what they changed and what they rejected. Those numbers are the closest thing you have to evidence of how much the person relies on the output, in either direction.

Which ERP agent decisions are likely in scope?

The table below is our reading of the OAIC’s examples against the agent patterns we design, not legal advice. Note the second column. Much of what an ERP agent touches is company data, and the rule only bites where personal information about an individual is used, which in a B2B business usually means sole traders, consumers buying direct, guarantors and staff.

The first row deserves a second look, because it catches businesses with no AI at all. The AI credit management agent we’ve described reads aged receivables and drafts hold recommendations for a credit controller. An automatic credit hold that fires when an invoice passes 60 days does the same job with no person in between, which looks a lot like a decision made solely by a program.

What has to go in the privacy policy?

Three things, per APP 1.8: the kinds of personal information used in operating the programs, the kinds of decisions made solely by a program, and the kinds of decisions where a program does something substantially and directly related.

The OAIC asks for plain language and meaningful groupings and warns against “unnecessary amounts of granular technical detail”. Sensitive information, such as health information, should be clearly identified. The fact sheet also says vendors should give their customers clear, high-level information about how their programs work, so it’s reasonable to ask your ERP vendor, and whoever built your agents, for a paragraph you can hand your adviser.

As an illustration only, the notes a wholesaler takes to its adviser might read: “Decisions to place a trade account on hold or change its credit limit are informed by software that reads payment history, account balances and the account contact’s details. A staff member approves each change.” The adviser turns that into policy wording. We wouldn’t.

On enforcement, the OAIC’s infringement notices page, published 22 September 2026, lists the existing privacy policy duties in APP 1.3 and 1.4 as infringement notice provisions, at up to 60 penalty units for a body corporate ($21,840 at the current unit value). The page doesn’t mention APP 1.7 or 1.8, so how the new content is enforced is a question for your adviser, not something to assume from this post.

A worked example: four agents, one afternoon

Picture a Brisbane equipment wholesaler on Odoo with about 60 staff, a small webstore and four agents in its AI agent register. The finance manager and the operations manager sit down with the OAIC flowchart and go through the register line by line.

1. Sales order agent. Reads emailed purchase orders and drafts sales orders. Company customers, and the decision is how to key an order. Out.
2. Replenishment agent. Drafts purchase orders from sales velocity and lead times. No individual is affected. Out.
3. Credit agent. Drafts hold and limit recommendations. About 40 accounts are sole traders. The audit trail shows the credit controller approved 212 of 220 recommendations last quarter without an edit. That reads like the OAIC’s loan officer example. In, as a decision the program substantially informs.
4. Warranty triage agent. Classifies webstore claims and drafts a response, including a recommended rejection where the serial number is out of warranty. Consumers are affected. In.

Two lines go to the privacy adviser, each naming the kind of decision and the kinds of personal information. One process change comes out of it as well: the credit controller now records a short reason when approving a hold on a sole trader’s account. That is good practice whatever the Act says. Nothing is published until the adviser and a director sign off the wording, and the agents keep doing what they did before, which is draft and wait.

What Odoo and MYOB Acumatica change about the answer

The platform doesn’t decide whether you’re in scope, but it does decide how easily a decision becomes one made “solely” by a program.

In Odoo 20, released 24 September 2026, the release notes say agents can be summoned by automated and scheduled actions and can update records. If a scheduled agent changes a credit limit or a customer’s price list with nobody in the chat, there’s no person in the decision at all. We set out where the gate has to go in our note on Odoo 20 agents in automated and scheduled actions. The disclosure question is one more reason to keep those tools off unattended agents.

In MYOB Acumatica, AI Automation prompts run when a person clicks the action, as covered in our MYOB Acumatica AI Automation hand-off note, so decisions made solely by that feature are unlikely. MYOB’s AI page, read on 8 October 2026, lists eight AI features, including Anomaly Detection, which flags unusual activity in expenses, supplier invoices and other transactions. Whether a flag is “a key factor” depends on what your team does with it, which is the same reliance question as everywhere else.

What to do before 10 December

• List every program, not only the AI ones. Agents, AI features switched on in the ERP, scoring spreadsheets and automatic rules such as credit holds. If you don’t have a register yet, this is the week to start one.
• Run each line through the OAIC flowchart. Ask who the decision affects, whether that’s an individual, and whether the effect is more than trivial.
• Pull the approval numbers. For each drafts-first agent, how often was the draft approved unchanged? High reliance isn’t wrong. It just means the program belongs in the policy.
• Ask your vendors for their paragraph. The fact sheet expects them to help.
• Brief your privacy adviser with the list, and let them write the wording.
• Add a trigger for new agents. The rule we’d suggest: nothing new goes live until its register line says whether it touches a decision about a person. The controls behind that are in our guide to AI agent governance for Australian businesses.

Where you’re unsure, the fact sheet says to take a cautious approach. A sentence in a privacy policy about a credit decision costs very little.

Goes well with

• The AI agent register: the one-page list of agents, owners and approval gates that this exercise starts from.
• AI agent governance for Australian businesses: the four controls, including the audit trail that shows how much your approvers rely on a draft.
• The AI credit management agent: the agent pattern most likely to land in your privacy policy, and how its approval gate works.

Not sure which of your agents the 10 December rule catches?

Auboros designs and runs drafts-first AI agents over Odoo and MYOB Acumatica from Brisbane, for businesses across Queensland and Australia. We can build the agent inventory, pull the approval figures from the audit trail and hand your privacy adviser a list they can work from.

If you’d like that done before December, book a free consultation. Bring the automatic credit-hold rule too. It counts.