AI Agents for Australian Businesses: What Actually Works in 2026

Most of what’s written about AI agents comes from people who have never connected one to a live business system. The demos look effortless. The LinkedIn posts promise a workforce that never sleeps. Then you ask a practical question, like “what happens when the agent books a journal entry against the wrong tax code?”, and the room goes quiet.

We run AI agents in production, on our own operations and against ERP systems, so this guide is written from the unglamorous side of the demo. Here’s what AI agents can actually do for Australian businesses in 2026, where Odoo and MYOB Acumatica sit today, and the governance work that separates a useful agent from an expensive incident.

What an AI agent actually is (and what it isn’t)

An AI agent in a business context is software that uses a large language model to work towards a goal across multiple steps: it can read information from your systems, decide what to do next, use tools like your ERP’s API to act, and hand the result to a person for approval. That last part matters. A chatbot answers questions. An agent does work.

The distinction from traditional automation matters too. A workflow rule follows a fixed path: if X, then Y, every time. An agent handles the messy middle, like reading a purchase order PDF that’s formatted differently by every customer, and still produces a structured result. That flexibility is the value, and it’s also the risk, because language models make mistakes. Anyone selling you an agent that “never gets it wrong” hasn’t run one for long.

Adoption is no longer fringe. The Australian Bureau of Statistics reports that 12% of Australian businesses used AI in 2024-25, up from 1% two years earlier, with 35% of large businesses on board. In the sectors we work with, intent is stronger again: MYOB’s research found 78% of wholesale distribution businesses plan to use AI in their ERP. The question has shifted from whether to use AI to how to wire it into the systems that run the business without breaking anything.

What AI agents are doing in Australian businesses right now

Forget the abstract use-case lists. Here’s the level of specificity that matters when you’re scoping this work.

Reading inbound purchase orders and drafting sales orders

A wholesale business receives purchase orders as PDF attachments, each customer using their own format. An agent watches the inbox, reads each PDF, matches the customer against the ERP, checks the products and pricing against the price list, checks stock, and drafts a sales order. A person reviews the draft, fixes anything odd, and confirms it. The agent never posts the order itself. That single workflow routinely saves hours of rekeying a day, and because a human approves every order, a misread line item is an annoyance rather than a shipped mistake.

Flagging reconciliation exceptions before month end

Finance teams running an ERP alongside e-commerce, point of sale and a warehouse system spend days each month finding out why the numbers don’t agree. An agent can compare transactions across those systems daily, flag the exceptions, and propose a resolution for each one: a missed payout fee here, a duplicated order there. The accountant reviews the proposals and applies the ones that are right. Nothing touches the ledger without sign-off, which also keeps your Business Activity Statement (BAS) position clean, because every correction is a human decision with an audit trail.

Chasing suppliers and watching stock

On the operations side, agents draft the boring correspondence: chasing suppliers for delivery confirmations on open purchase orders, flagging purchase orders that will land late against promised customer dates, and preparing replenishment suggestions from sales velocity and lead times. The pattern is identical in every case. The agent reads, drafts, flags and proposes. A person decides.

If you’re evaluating any agent product or consultancy, that pattern is the test. Ask exactly where the human approval sits. If the answer is vague, keep your API keys in your pocket.

Where Odoo and MYOB Acumatica actually are with AI

Vendor announcements and shipped software are different things, so here’s the current state of both platforms we implement, with each capability labelled as shipped or roadmap.

Odoo: useful AI shipped in v19, agents on the v20 roadmap

Odoo 19, the current production release, ships AI embedded across the suite: AI-assisted fields and server actions, document extraction in accounting, AI in CRM, and AI-powered livechat and knowledge features. We’ve verified these against the v19 Enterprise source code, not the marketing site, and we cover the detail in our guide to AI in Odoo for Australian businesses.

The bigger step is Odoo 20, which is expected to move from assistive AI towards agentic AI: agents that execute multi-step work across modules. That capability is roadmap, not product. It’s due to be shown at Odoo Experience in late September 2026, with general availability expected after that. If a vendor or partner tells you Odoo agents are shipping today, they’re describing the roadmap as the product. Our Odoo v20 roadmap breakdown covers what’s confirmed versus expected, and none of it changes our standing advice: implement on the current version now, and treat v20’s agents as an upgrade decision once they’re real and stable.

MYOB Acumatica: AI arriving on the usual release lag

Globally, Acumatica’s 2026 R1 release reached general availability in March 2026 with an AI Assistant and AI Studio, AI-powered anomaly detection in reporting, and early AI-assisted workflows, some of it in experimental or early-access form. The important caveat for Australian and New Zealand businesses: global Acumatica releases typically reach MYOB Acumatica customers six to nine months later, and the release most local customers run today is 2025 R2. “Acumatica has it” and “your MYOB Acumatica instance has it” are usually two different statements, so confirm with your partner what’s actually enabled on your version before you plan around it.

In the meantime, agents don’t have to wait for native features on either platform. Both expose APIs, which means the agent layer can sit outside the ERP under your own governance. That’s where most of the practical work happens in 2026.

How agents connect to your ERP without handing over the keys

The connection layer is where most of the security and governance questions live, and it’s matured a lot in the past year. The Model Context Protocol (MCP) has become the common standard for connecting AI models to business tools. It started at Anthropic and was donated to the Linux Foundation’s Agentic AI Foundation in December 2025, with the major AI vendors backing it. In plain terms, MCP means you can build one governed connection to your ERP and let your choice of AI model use it, rather than wiring each tool to each system separately.

That gives you a real architectural choice:

Hosted assistants, like Microsoft Copilot or Anthropic’s Claude, connected to your ERP through MCP or an integration layer. Fastest to start, and the vendor handles the model. You still control what the connection can see and do.
Self-hosted agent frameworks, like the open-source OpenClaw, which run on infrastructure you control. More work to operate, and you carry more of the security responsibility, but data flows and model choice stay entirely in your hands.
Bring-your-own-LLM builds, where the agent logic is yours and the underlying model is swappable. This is how we build: if a better or cheaper model ships next quarter, you change a setting, not the architecture.

Whichever route fits, the non-negotiable is scope. The agent gets its own credentials, with the same data visibility as the person it works for, never a shared administrator login. An agent that helps a sales rep should see exactly what that rep sees. Nothing more.

The governance layer: what keeps agents out of trouble

This is the part most AI content skips, and it’s the part your accountant, your auditor and your board will ask about. The good news is that Australia now has practical guidance to anchor it. The National AI Centre’s Guidance for AI Adoption, released in October 2025, replaced the earlier Voluntary AI Safety Standard and sets out six essential practices, including accountability, risk management and human oversight. It’s voluntary, but it’s the reference point we’d expect customers, insurers and larger trading partners to start measuring against.

Privacy is not voluntary. The Privacy Act and the Australian Privacy Principles apply whenever personal information goes through an AI workflow, and the regulator has published specific guidance on using commercially available AI products. There’s also a date to put in your diary: from 10 December 2026, privacy policies must disclose the kinds of decisions made by substantially automated processes that significantly affect people. If an agent helps decide who gets credit terms, that’s your problem to document before it’s the OAIC’s problem to investigate.

In practice, the governance that matters on an agent-over-ERP build comes down to five controls:

Role-scoped permissions. The agent acts with the requesting user’s data access, never a super-user account.
Approval gates on writes. Anything touching the ledger, tax codes, pricing or stock is drafted by the agent and posted by a person.
Audit trails. Every agent action is logged: what it read, what it proposed, who approved it, and when.
Sandbox-first deployment. Agents earn production access by proving themselves against a copy of your data first.
A usage policy staff have signed. Which tools are approved, what data can leave the building, and who owns each agent.

“The first decision on any agent build is what it’s not allowed to touch. An agent that drafts a sales order and waits for a person is an asset. An agent that can post to the ledger unsupervised is a liability with an API key. The approval gate isn’t a compromise, it’s the design.”

Bill Alvarez, Practice Manager, Auboros

Where agents still fail (plan for it, don’t pretend otherwise)

Language models produce confident, fluent errors. Not often, but never zero. An agent will occasionally misread a quantity, match the wrong customer, or summarise a document in a way that drops the one detail that mattered. Failure modes we design around, because we’ve seen them:

Confident misreads. A scanned purchase order with a smudged quantity becomes a wrong draft. The approval step catches it, which is why the approval step exists.
Stale context. An agent working from last week’s price list quotes last week’s prices. Agents need live system data, not exported copies.
Permission creep. An agent set up quickly with broad access “just to get it working” and never tightened. Schedule permission reviews the way you’d review user access.
Cost drift. Agents that retry, loop or over-read can quietly run up model usage bills. Set budgets and alerts per agent from day one.

None of this argues against agents. It argues against unsupervised agents. The businesses getting real value in 2026 aren’t the ones with the most autonomous setup, they’re the ones whose agents remove the repetitive reading, matching and drafting while people keep the judgement calls.

Where to start if you run Odoo or MYOB Acumatica

The sequence we recommend, and follow ourselves:

Fix the data first. An agent reasoning over duplicate customers and stale price lists just automates confusion. Data hygiene is the unskippable prerequisite.
Pick one workflow with a clear human checkpoint. Inbound order entry, reconciliation exceptions and supplier chasing are proven starters. Pick the one that burns the most hours.
Run it in a sandbox against real historical data. Measure how often the agent’s drafts are right before it goes anywhere near production.
Go live drafts-first. Keep the approval gate permanent on financial writes, and review the logs weekly for the first quarter.
Only then add the next agent. A registry of what runs, who owns it and what it can access stops “a few helpful agents” becoming an unmanaged crowd.

This is the same discipline as any ERP project: scope tightly, prove it, then expand. If your platform decision is still open, or your current system’s data isn’t ready for any of this, that’s a conversation about ERP foundations before it’s a conversation about AI.

Frequently asked questions

What is an AI agent in an ERP context?

An AI agent in an ERP context is software that uses a large language model to complete multi-step work against your business system: it reads data, decides the next step, acts through the ERP’s interfaces or API, and hands the result to a person for approval. It differs from a chatbot, which only answers questions, and from workflow automation, which follows fixed rules.

Are AI agents safe for Australian businesses to connect to accounting software?

They can be, if the connection is governed. Safe setups give the agent role-scoped credentials rather than administrator access, require human approval before anything posts to the ledger, log every action, and are tested in a sandbox first. An agent with unsupervised write access to your accounts is not a safe setup, whatever the vendor demo suggests.

Do Odoo or MYOB Acumatica have AI agents built in?

Not as shipped products in Australia as of mid-2026. Odoo 19 includes useful assistive AI, with agentic capability on the Odoo 20 roadmap expected at the end of September 2026. Acumatica’s 2026 R1 release added AI Assistant and early agent workflows globally, which typically reach MYOB Acumatica customers six to nine months after the global release. Agents connected through APIs work with both platforms today.

What does the Australian government require before a business uses AI?

There’s no dedicated AI Act. The National AI Centre’s Guidance for AI Adoption sets out six voluntary practices, while existing law still applies, most importantly the Privacy Act when personal information is involved. From 10 December 2026, privacy policies must also disclose certain substantially automated decisions that significantly affect people.

How is an AI agent different from workflow automation?

Workflow automation follows fixed rules: the same trigger always produces the same action. An AI agent interprets unstructured, variable input, like a differently formatted PDF order from every customer, and still produces structured output. That flexibility means agents handle work rules can’t, but also that their output needs human review in a way deterministic automation doesn’t.

Thinking about AI agents for your ERP?

Auboros designs, governs and implements AI agents for businesses running Odoo and MYOB Acumatica, from our base in Brisbane serving Queensland, NSW and Victoria. We run agents in production ourselves, with the approval gates and permissions described above, so the advice comes from operating experience rather than a slide deck.

If you want a clear-eyed view of what an agent could take off your team’s plate, book a free consultation. We’ll tell you what’s worth automating and what isn’t yet.