The AI Agent Register: How to Stay in Control When Your Business Runs More Than One Agent

An AI agent register is a single record of every AI agent your business runs: what each one does, who owns it, which systems it can touch, what it costs, and who approves its work. It sounds like admin. It is admin. It’s also the difference between running agents and being run by them, and with 12% of Australian businesses now using AI (22% of medium-sized ones, per the ABS), plenty of teams are about to find that out.

We’ve covered what AI agents can do inside Australian businesses and the controls that make a single agent safe. This post is about the next problem, the one that arrives quietly: what happens when there’s more than one.

One agent is a tool, five agents are a fleet

Nobody plans agent sprawl. It happens one useful idea at a time. The sales team gets an agent that drafts orders from emailed PDFs. The buyer wants the same treatment for replenishment. Meanwhile someone in accounts has built a Copilot flow nobody else has seen, and a manager is pasting debtor lists into a chatbot because it’s faster than asking for a report.

Each might be fine on its own. Together they raise questions that get harder every month. What’s actually running? Who owns the one that touches pricing? What did they cost last month? Which ones can write to the ledger, and who approves what they write?

Auboros runs research, content and QA agents over our own operations, drafts-first with a person approving the output, and even at that scale we lost track of one for a fortnight. It kept working. That was the unsettling part. An agent that fails loudly gets fixed; one that hums along unwatched is the real risk.

“The first register we build with a client is never surprising because of the agents IT knows about. It’s the two nobody wrote down. You can’t govern what you haven’t listed.”

Bill Alvarez, Practice Manager, Auboros

What goes in an AI agent register

The good news is that the fix is a table, not a platform. Ours has seven columns, and it fits on a page:

• Name and job. What the agent does, in one sentence a non-technical director would understand. “Reads emailed purchase orders and drafts sales orders for approval.”
• Owner. A named person, not a department. The owner answers for the agent’s output.
• Systems and scope. Which systems it connects to, whether access is read-only or write, and whose permissions it acts under. An agent should act with a real user’s data scope, never a shared super-user.
• Model and data path. Which AI model or vendor sits behind it, and where your data goes. If customer records leave Australia, this is where that gets written down.
• Approval gates. What it’s allowed to draft, and who approves each draft before anything posts. Ledger, tax, pricing and stock write-backs need a human on the gate, every time.
• Cost. The subscription plus last month’s usage. Agent costs drift, and the register is where drift becomes visible.
• Escalation and review. Who can switch it off, what it should do when it’s unsure, and the date of the next review.

You don’t have to invent the format. The National AI Centre publishes a free AI register template as part of its Guidance for AI Adoption, which condensed Australia’s earlier ten voluntary guardrails into six essential practices in October 2025. The guidance expects approved AI systems to be recorded in a register. For agents that touch an ERP, we’d add the approval-gate and cost columns; the template covers the rest.

The two dates that make a register worth having

The register pays for itself operationally, but two dates give it teeth.

The first has passed: 21 October 2025, when the National AI Centre released the guidance above. It’s voluntary, and Australia still has no general AI Act. It is, though, the standard a customer, insurer or board member will reach for when they ask how you govern AI, and “we follow the national guidance, here’s our register” is a short, convincing answer.

The second is coming: 10 December 2026. From that date, businesses covered by the Privacy Act must say in their privacy policies what kinds of personal information they use in automated decisions that could significantly affect people, and what kinds of decisions those are. The OAIC consulted on draft guidance over May and June 2026 and says it intends to publish the final version by September 2026. You cannot write that disclosure without knowing what your agents do. The register is where the answer comes from. We covered the wider control set in our guide to AI agent governance for Australian businesses; the register is the page that makes the rest of it auditable.

What Odoo and MYOB Acumatica already track for you

If your agents live inside your ERP, part of the register already exists.

In Odoo 19, every AI agent is a record in the system: its name, its instructions, which model it uses, which tools it’s allowed to call and which knowledge sources it can read. Odoo’s own documentation confirms that agents act within the signed-in user’s access rights and that the default Ask AI agent can’t create or alter data. That’s shipped today. One telling detail from the source code: the call summariser agent ships switched off. A human decides to turn it on. We’ve written up what Odoo’s AI agents can do today in more depth.

MYOB Acumatica is earlier in the same journey. The current ANZ release is 25 R2, which includes AI Studio in technology preview, built with AWS and Anthropic with data processed in local data centres. In the preview release, agents run when a user asks and can’t create records or documents on their own; general availability is planned for the 26 R1 release, and MYOB hasn’t committed an ANZ date. Globally, Acumatica’s 2026 R1 shipped in March 2026 with its AI Assistant under managed availability, and global features typically reach the ANZ product 6 to 9 months later. Label your register entries accordingly: shipped, preview or roadmap.

What neither platform can see is everything outside itself: the Copilot flow in accounts, the chatbot project a manager runs from a browser, the automation a contractor left behind. That gap is why Microsoft now sells Agent 365, a control plane for discovering and governing agents that reached general availability in May 2026. For a large organisation deep in Microsoft 365 it’s worth a look; for most mid-market businesses it’s more platform than the problem needs. The one-page register comes first either way, because no tool can govern agents you haven’t decided to own.

A Tuesday morning in mission control

Here’s the register alive, drawn from patterns we build for wholesale clients.

A Queensland distributor runs three agents over its ERP. At 7:20am the operations manager opens the approval queues. The replenishment agent proposed eleven purchase orders overnight; the buyer approves nine and sends two back because a supplier has changed its minimum order quantity. The sales order agent read fourteen emailed purchase orders and drafted thirteen; the fourteenth was an unreadable scan, so it escalated to a person instead of guessing. The credit agent drafted six payment reminders; the credit controller approves five and holds one because that customer is mid-conversation about a payment plan.

Then the cost column earns its place. Usage on the sales order agent doubled last week. The register names an owner, the owner looks, and the cause is a supplier now attaching a 300-page catalogue to every order email. The fix takes ten minutes. Without the register, that cost drifts until someone spots it in the P&L, with no name attached.

Twenty minutes, one page, three humans making every decision that matters. That’s mission control at mid-market scale.

Start with a page, not a platform

If you run even one agent against live business systems, start the register this week. List everything that runs, including the unofficial ones; an amnesty gets better answers than an audit. Name an owner for each. Write down scope, gates and last month’s cost. Book a monthly half-hour review, and give new agents a sandbox before they touch production. Our AI agent services page describes how we design these patterns, drafts-first with role-scoped permissions, and the register is the first artefact we build on every engagement.

Running more than one agent, or about to?

Auboros designs, governs and runs AI agents over Odoo and MYOB Acumatica for Australian businesses, from our base in Brisbane and the Gold Coast. We’ll help you build the register, set the approval gates and wire the agents in safely.

If you’d like a second pair of eyes on what’s already running in your business, book a free consultation. Bring the unofficial agents too. We’ve seen worse.