AI Risk Register for Australian Businesses: What to Record for Each Agent on Your ERP

An AI risk register is a short list, kept per agent, of what could go wrong, how you would notice, how far the damage could spread and which control stops it. As at 13 October 2026, Australia’s National AI Centre asks for the pieces without using the name: seven screening questions first, then a documented risk assessment and treatment plan for each use.

At a glance

• Who it is for: owners, CFOs and operations managers of Australian businesses on Odoo or MYOB Acumatica who are about to let an AI agent read or draft against live ERP data, and the person who will approve its drafts.
• Problem it solves: most AI risk talk stays at the level of “hallucination” and “privacy”. A register turns it into named failures of one agent, each with a way to notice and a control with an owner.
• Platforms: Odoo 20 and Odoo 19, and MYOB Acumatica. The register is a document, so it works the same on both; the controls it points to are platform settings.
• Trigger: before a business case or pilot (screening), before go-live (full lines), and again whenever the agent gets a new permission, data source or audience.
• Inputs read: the agent’s row in your AI register, its permission list, the records it can reach, and the approval steps a person performs.
• What gets written: a screening result, then one line per failure with the record it touches, how you would notice, the blast radius, the control, the owner and a review date.
• Approval gate: the agent’s owner signs the register before the agent leaves the sandbox. A yes to any screening question goes to senior leadership first, as the National AI Centre’s tool suggests.
• What it never does: replace a privacy impact assessment, legal advice or the approval gate itself. It records the controls; it doesn’t perform them.
• Typical cost: no licence. It is staff time, and when we run the work it sits inside a pilot or proof of concept from $5,000 ex GST, per the AI agents for ERP in Australia hub as at 13 October 2026.
• Typical timeline: written during discovery, which the same hub page lists at 1 to 2 weeks, and revisited at the end of the sandbox pilot.
• Last verified: 13 October 2026 against the National AI Centre’s Guidance for AI Adoption: foundations, its implementation guidance, its AI screening questions and the OAIC’s automated decision-making resources.

The risk register that fails is the one written once, for a board paper, about “AI” in general. It lists bias, hallucination and cyber risk, rates each one amber, and never gets opened again. An agent that drafts sales orders has different failures from one that drafts overdue reminders, and the register only earns its keep when the person approving the drafts recognises every line.

What is an AI risk register?

An AI risk register is the working record of what one AI system could get wrong and what you’ve done about it. For an agent over an ERP, each line names a failure, the record it would touch, how you’d notice, how much could go wrong before a person sees it, the control that stops it, and who owns that control.

It is a different document from the two it usually gets confused with. The AI agent register is the inventory: which agents exist, who owns each one and what it’s allowed to do. The AI usage policy is the rulebook staff sign, one of the documents in our AI agent governance guide. The risk register sits between them. The inventory says the credit control agent exists. The risk register says what happens when it emails the wrong customer.

Does Australian guidance ask for an AI risk register?

Not by that name. We read the National AI Centre’s Guidance for AI Adoption on 13 October 2026, and neither the foundations page nor the implementation page uses the words “risk register”. What they ask for are the parts of one.

The foundations guidance lists “measure and manage risks” as one of six essential practices. Its starting step is to create a risk screening process that flags AI systems and use cases posing unacceptable risk. Its next steps are to conduct risk assessments and create mitigation plans for each specific use case, apply controls based on the level of risk, and set up a process to investigate, document and analyse AI incidents.

The implementation guidance goes further for teams that already use AI. It asks for an organisation-level risk tolerance, a triage system for systems that may pose enhanced or unacceptable risk, a documented risk assessment that weighs the likelihood and consequence of each risk, and a risk treatment plan, with risks reassessed after the controls are in. It also says the AI register should record the identified risks, the treatment plan and the dates of review for each system. Put those together in one place per agent and you have a risk register, whatever you call it.

Neither page describes itself as mandatory, and neither is a substitute for legal advice. The laws that already apply to your business apply to what your agent does.

What are the seven screening questions?

The National AI Centre publishes AI screening questions as a planning tool. There are seven, each answered yes or no, and the page is clear that there are no scores or automatic outcomes. It’s meant to be used before a business case or any spend, and before a full risk assessment, not instead of one.

In our words, not theirs, the seven ask whether the AI use:

1. Handles personal, sensitive or confidential information.
2. Deals directly with end users with a lot of autonomy and little human oversight.
3. Acts on its own at scale, or in a way that makes human intervention hard.
4. Makes or supports decisions that affect vulnerable or under-represented people.
5. Operates in a regulated area or can have legal effects on a person.
6. Could cause harm that is hard to contest, reverse or get redress for.
7. Is built for many purposes, or could easily be adapted beyond the intended one.

If every answer is no, the tool says to follow your normal adoption process and AI policy. A yes to any of them means consulting senior leadership or a governance committee before detailed planning. The page gives its own contrast: an AI system that automatically handles refunds needs more control than a chatbot taking restaurant bookings.

Here’s the part that matters for an ERP. Almost every useful agent over Odoo or MYOB Acumatica answers yes to the first question, because customer contacts, supplier contacts and employee records are personal information. So screening rarely ends with “nothing to see”. It ends with a list of which questions were a yes, and that list is the start of the register.

What should each line of the register record?

The guidance tells you to assess likelihood and consequence. For an agent over an ERP we add four columns that a generic template doesn’t have, because they’re the ones a finance or operations manager can actually check.

• The record it touches. A draft sales order, a customer email, a receipt date, a journal. If the failure can reach the ledger, tax codes, prices or stock, say so in the line.
• How you’d notice. The approval step, an exception report, a customer complaint. If the only way you’d find out is a customer telling you, the line isn’t finished.
• Blast radius. How many records, and how much money, before a person sees it. One draft waiting in a queue is a small radius. A scheduled run that posts overnight is a large one.
• The control and its owner. A named person, and a setting or step you can point to: a read-only connection, an approval gate, a daily cap, a suppression rule.

The table below is one agent’s register, written for the example in the next section.

Five lines is typical for a first pass. If you have twenty, you’re probably describing the technology and not the agent.

A worked example: screening and registering one agent

Picture a Brisbane equipment parts distributor on MYOB Acumatica with about sixty staff. Finance wants an agent that reads aged receivables each morning and drafts overdue reminders for the credit controller to approve, the pattern we describe in the AI credit management agent.

The finance manager runs the seven questions. Question one is a yes: debtor contacts are personal information, and some customers are sole traders. Question five is a maybe, because a credit hold has consequences for the customer. The rest are no, since the agent drafts and a person sends. Two answers aren’t a clean no, so the proposal goes to the directors before any build, as the screening tool suggests. They approve a pilot with one condition: the agent may recommend a hold, and may never apply one.

That condition becomes line three of the table above. It also raises a separate question for the privacy adviser. From 10 December 2026, a business covered by the Privacy Act must describe in its privacy policy the kinds of decisions where a computer program makes the decision, or does something substantially and directly related to making it, where personal information is used and the decision could reasonably be expected to significantly affect a person’s rights or interests, per the OAIC’s resources of 30 September 2026. Whether a recommended credit hold on a sole trader meets that test is the adviser’s call. The register’s job is to make sure someone asks.

Then the pilot runs in a sandbox, drafts only. In the second week the agent drafts a reminder for an invoice the customer had already disputed by phone, because nobody had set the dispute flag. The credit controller bins the draft. Line one gets a second control: a note field the agent must read before drafting. That’s the register doing its job. A failure showed up, a person caught it at the gate, and the line changed.

“A risk register nobody reopens is paperwork. The useful version is one page per agent, written with the person who approves its drafts, and reopened the day the agent gets a new permission. If a line can’t tell you how you’d notice the failure, it isn’t finished.”

Josh Craig, Director, Auboros

Where do the controls live in Odoo and MYOB Acumatica?

A register line is only as good as the setting behind it, so each control should point at something you can open and look at.

In Odoo 20, the controls we cite most are the agent’s skills, which decide what it can do beyond providing information (per Odoo’s 20.0 AI agents documentation, as we read it on 12 October 2026), and the access rights behind any outside connection. Our post on the Odoo 20 MCP server covers who a connection runs as and how to keep writes behind a person, and Odoo 20 agents in scheduled actions covers the unattended case, where the blast radius column matters most.

In MYOB Acumatica, MYOB’s AI page says, as at 13 October 2026, that AI Automation is configured by power users or administrators and can then be used by anyone in the business. For a register, that sentence is two lines: who may build an automation, and who may run it. The same page says customer data is stored in Australia and that its AI providers cannot store it or use it to train their models, which is the vendor statement to quote against screening question one. Our note on how an MYOB Acumatica AI Automation prompt hands work back to a person covers the approval side.

On either platform the strongest control is the same, and it’s the one every agent we build has: the agent drafts, and a person approves anything that changes the ledger, tax, prices or stock.

When should you reopen the register?

Set a review date on every line, and treat these as triggers to reopen it early:

• A new permission or skill. Read-only becoming draft, or draft becoming write, changes the blast radius of every line.
• A new data source. Pointing the agent at a second mailbox or another company’s records changes the answer to screening question one.
• A new audience. An internal agent that starts answering customers moves question two from no to yes.
• An incident or a near miss. The foundations guidance asks you to investigate, document and analyse AI incidents and use what you learn. A binned draft counts.
• A platform upgrade. An upgrade can add AI features and settings you haven’t screened. Our Odoo 20 upgrade checklist lists the settings to look at.

Language models make mistakes, and no register changes that. What it changes is whether the mistake lands in a draft queue with a name on it or in your ledger at 2am.

Goes well with

• The AI agent register: the inventory the risk register hangs off, with one row per agent, its owner and its permissions.
• AI agent governance for Australian businesses: the four controls the register’s last column keeps pointing to.
• What an AI agent costs to run in Australia: once the risks are written down, the budget line for the same agent.

Want a risk register that matches the agent you’re actually building?

Auboros designs and runs drafts-first AI agents over Odoo and MYOB Acumatica from Brisbane, and we write the AI risk register with the people who’ll approve the drafts, before anything leaves the sandbox. If you’d like to walk one agent through the seven questions with us, book a free consultation. We’ll tell you plainly which lines we’d worry about first.