Conversational ERP in Australia: What Odoo 19 and MYOB Acumatica Ship, and How to Keep It Safe

Updated 15 September 2026: MYOB Acumatica now ships a built-in AI Assistant that answers plain-language questions, so the “mostly roadmap” verdict from August is corrected; the AI Studio preview wording is replaced with AI Automation and MYOB’s own version numbers; MYOB’s statement on where AI data is processed is updated; Odoo 19’s Ask AI, agent and AI server action behaviour is restated from the 19.0 documentation; added the short answer, an At a glance block, question headings, a captioned options table, “Goes well with” and a rebuilt FAQ.

Conversational ERP is asking a live business system questions in plain English and getting answers from its data. As at 15 September 2026 both platforms ship a read-only version: Odoo 19’s Ask AI and MYOB Acumatica’s AI Assistant. Both are built to answer questions, not post entries. The safety question is not the model but whose login the assistant answers under, and what happens when it drafts a change.

At a glance

• Who it is for: owners, CFOs and operations managers on Odoo or MYOB Acumatica who want staff to ask the system questions without learning the report builder, and who want to know what that exposes.
• Problem it solves: the gap between the question a person has (“which customers are over their credit limit and still ordering?”) and the filtered report someone has to build to answer it.
• Platforms: Odoo 19 Enterprise (Ask AI and the AI app); MYOB Acumatica 2025.2 and 2026.1 (AI Assistant and AI Automation); any ERP with an API through a bring-your-own model connected over MCP.
• Trigger: a person types a question, in the ERP’s own assistant or in a hosted chat tool connected to it.
• Inputs read: whatever records the asking person’s login can already see: customers, invoices, orders, stock, purchase orders, ledger balances.
• Systems: the ERP; optionally a hosted model (OpenAI or Gemini in Odoo; MYOB-managed AWS Bedrock in MYOB Acumatica; your choice via MCP) and an MCP server in front of the ERP.
• What the agent drafts: an answer with links to the underlying records; on request, a draft email, note or proposed record change for a person to review.
• Approval gate: reads return immediately; any write to customers, orders, stock, pricing or the ledger is drafted and waits for a named person to approve it.
• Audit trail: who asked what, what the assistant read, what it proposed and who approved it, kept with the record (chatter in Odoo, notes and the AI Automation trace in MYOB Acumatica) or in the MCP server’s log.
• Typical cycle time: seconds for a question; a drafted change waits as long as the approver takes.
• What it never does: run as a shared administrator login, post to the ledger, change prices or credit limits, or answer from data the asking person could not open themselves.
• Typical cost: the native assistants are part of the platform (Odoo Online works without your own key; Odoo.sh and on-premise need your own provider key per Odoo’s AI API key documentation; MYOB’s features are “subject to licensing” per MYOB’s LLM provider documentation). A scoped agent pilot over the API typically starts from a few thousand dollars ex GST per our AI agent services page.
• Typical timeline: switching a native assistant on is an administrator task measured in hours; a bring-your-own-model pilot is 4 to 10 weeks from discovery to a drafts-first go-live per the same page.
• Last verified: 15 September 2026 against Odoo’s Ask AI documentation, MYOB’s AI features page and the OAIC’s automated decision-making consultation page.

Ask a question in plain English, get an answer from live business data. That is the pitch for conversational ERP, and in 2026 it has moved from conference demo to shipping feature on both platforms we implement. The ABS reports that 12% of Australian businesses used AI in 2024-25, rising to about 35% of large businesses, so a chat box on top of the ledger is coming to your industry whether you plan for it or not.

Here is the part the vendor demos skip: the model is not the hard bit. The hard bit is deciding whose eyes the answer is allowed to see through.

What is conversational ERP?

Conversational ERP is natural-language querying of a live business system. Instead of building a filtered report, you type the question the report was meant to answer: “which invoices over $10,000 are more than 30 days overdue?”, “what did we sell to Queensland customers last month?”, “which purchase orders are late?” The answer comes back from live data, not from a spreadsheet someone exported last Tuesday.

The people who benefit most are the ones who never learned the report builder. A warehouse lead who wants to know what is arriving this week. An owner who wants one number on a Saturday morning without phoning the bookkeeper. A sales rep prepping for a call who needs the customer’s order history and open credits in thirty seconds. That value is real, and it lands on day one because answering questions does not require the agent to change anything.

What conversational ERP is not: a reason to hand an AI system open access to everything. It is the entry point to AI agents for ERP in Australia, and the permission decisions you make for the chat box are the ones every later agent inherits.

Whose login does the assistant answer under?

Every conversational ERP setup answers one question, either deliberately or by accident: whose login does the agent use?

If the agent runs as a shared super-user account, anyone who can type a question can read anything the ERP contains. Payroll. Margins. The warehouse casual asking “what does everyone here get paid?” is one autocomplete away from an answer no permission matrix would ever have shown them. The rule we set out in our AI agent governance guide applies with full force here: an agent acts with the requesting user’s exact data scope, never a shared privileged account.

Both vendors have built their native assistants this way. Odoo’s Ask AI opens from the signed-in user’s session, and the standard agent cannot make changes to the database: it can open views and display reports, but it cannot create leads or alter data. MYOB says its AI features respect the same user permissions, access controls and governance settings already in place, and that only authorised users can configure or run AI Automation prompts. The place the rule gets broken is not the native assistant. It is the bring-your-own integration someone wires up with an administrator API key because it was quicker.

“Whose login does the agent run under? That single question does more for safety than any model choice. Get it wrong and you have built a search engine for everything the ERP was carefully hiding.”

Josh Craig, Director, Auboros

What do Odoo 19 and MYOB Acumatica ship today?

It pays to be precise about shipped versus promised, and the answer changed between August and September. Everything below is labelled shipped, global only or roadmap, and traces to a vendor page checked on 15 September 2026.

Odoo 19: shipped. Ask AI opens from anywhere in the database with Ctrl+K or the AI button. Per Odoo’s documentation, it answers questions, opens views, summarises chatter threads and drafts follow-up messages, and each response can be sent as a message, logged as a note or copied. The standard agent cannot change the database. Building your own agents needs the AI app, which is Enterprise; per the agents documentation an agent with no Topics assigned “is only able to provide information”, and Restrict to Sources limits it to the documents you approve. Odoo Online works without your own key; Odoo.sh and on-premise databases need your own OpenAI or Gemini key. We covered the mechanics in our Odoo AI agents deep-dive, and the multi-step agents Odoo has talked about for version 20 stay roadmap until Odoo Experience on 24 to 26 September.

MYOB Acumatica: shipped, and this is the correction. In August we called MYOB’s conversational side “mostly roadmap”. As at 15 September 2026 MYOB’s AI features page lists a built-in AI Assistant that lets you “ask questions in plain language about revenue, customers, invoices, inventory and financial performance, or use a ready-to-go prompt”. It sits alongside AI Automation (previously called AI Studio), which lets an administrator attach custom prompts to any form; MYOB’s knowledge base documents the setup for version 2025.2 and says the MYOB-managed connection is available immediately on 2026.1. Our MYOB Acumatica AI review walks through all eight features MYOB now lists. What is still roadmap is the natural-language querying MYOB and Microsoft promised under their April 2026 partnership, due “late 2026”; as at September MYOB’s feature list attributes nothing to it.

Global Acumatica: not in your tenant yet. Acumatica’s global AI page (modified 26 August 2026) describes Model Context Protocol access so Claude, Gemini, Cursor and other MCP-compatible tools can read permitted data under existing role security. MYOB has not published an ANZ date for it. Global features typically reach the MYOB release 6 to 9 months later, so plan on the API route below until the MYOB release notes say otherwise.

The bring-your-own route: available now, with work. You can connect a hosted assistant to either platform through the Model Context Protocol, an open standard for wiring AI tools to business systems. An MCP server sits in front of the ERP with its own scoped credentials, so you decide what it can see and do, and you can swap the underlying model later without rebuilding. This is how we run our own agents. Odoo 19 does not ship a first-party MCP server, so the server talks to Odoo through its JSON-RPC API under a per-user API key; MYOB Acumatica is reached through its API under a named API user, where MYOB’s API licence page sets the included user at 1,500 calls a day and sells unlimited access separately. Our MCP for ERP guide covers the setup.

What happens when the question becomes a request to change something?

Reads are the easy half. The moment someone types “put that customer on hold” or “reorder 200 of those”, a conversational interface becomes a write tool, and the write path is where the design either has a gate or does not.

Odoo’s own documentation is refreshingly blunt about this. In AI server actions, the AI action is “the manager” that reads the record and decides which tool to call; the tool is “the worker” that performs the update. The manager “does not enforce business rules, modify records directly, or guarantee the correctness of the operation”, and if a tool is called “it will execute unconditionally, unless the code itself prevents it”. Business rules must be enforced explicitly in the tool’s Python code. In plain terms: if you want a credit hold to wait for a person, the waiting has to be built into the tool, because the model will not add it for you.

MYOB Acumatica’s AI Automation is drafts-first by construction: the prompt runs when a person clicks the action on a form, the output lands in front of that person, and data masking keeps marked fields inside the platform with a trace page showing what was sent. The catch to know is elsewhere in the platform: on the Process Bank Transactions form, payments the system creates from matched lines skip approval workflows, which is why our reconciliation agent guide puts the gate on the match list.

Over the API, the gate is yours to build, and we build it the same way every time: the MCP server exposes read tools freely and write tools only as “propose” actions that create a draft, a note or an approval task, never a posted record. The AI agent register records which tools each agent has and who approves what.

A worked example: credit questions that stay read-only

Here is how the pattern looks for a Queensland wholesale distributor. An accounts receivable officer asks the assistant: “which customers are over their credit limit and still placing orders?” The assistant queries the ERP under her login, and only her login. It returns four customers, each with the exposure amount, the last order date and a link to the underlying records so she can check the numbers herself.

Then she asks it to do something about it: “draft hold notices for the two largest.” The assistant writes two emails and proposes a credit-hold flag on each account. And there it stops. Nothing sends and no flag changes until she has read each draft and approved it. Reads are instant, writes wait for a human. That one rule is the difference between a useful assistant and an incident report, and it is the same rule the credit management agent runs on when the questions become a scheduled job instead of a chat.

Does conversational ERP trigger the December 2026 privacy disclosure?

There is a compliance clock on this. The Privacy and Other Legislation Amendment Act 2024 introduced an automated decision-making obligation: from 10 December 2026, organisations covered by the Australian Privacy Principles that use personal information in automated decisions with the potential to affect people’s rights or interests must describe, in their privacy policy, the kinds of personal information used and the kinds of decisions made. The OAIC consulted on its guidance until 15 June 2026 and, as at 15 September, has not published the final version.

A read-only question box about stock levels sits at the low-risk end. Once the same assistant starts recommending credit holds on named customers, or reading staff records to answer HR questions, the disclosure obligation is squarely in frame. Design the permission model and the policy wording now; retrofitting both in December is the expensive order.

The data path matters too. MYOB’s page says customer data is stored in Australia and that, depending on the feature, AI processing happens in Australia or through global infrastructure run by Anthropic, AWS and Microsoft, which can process data to generate a response but cannot store it or train on it. Odoo sends prompts to OpenAI or Gemini under your key or its own. Either way, know which model processes which prompts before finance or personal data flows through it, and use masking or Restrict to Sources for anything the model does not need to see.

Five questions before you switch it on

Put these to any vendor, integrator or enthusiastic staff member proposing a chat interface over your ERP:

• Whose login does it use? The agent should act with the requesting user’s permissions, never a shared account with broad access.
• Can it write, and where are the gates? Read-only is a fine starting point. Any write-back to ledgers, pricing, stock or customer records needs a named human approving each change, built into the tool, not assumed from the prompt.
• What gets logged? You want a record of who asked what, what the agent read and what it proposed. Auditors will eventually ask for it.
• Where does the data go? Know which model processes your prompts, in which country and under what retention terms before finance data flows through it.
• What is it allowed to answer from? Source and tool allowlists beat an open connection to everything. Start narrow and widen deliberately.

If those five have good answers, conversational ERP is one of the lower-risk places to start with AI: mostly reads, and easy to trial in a sandbox before anyone touches production. We covered the wider picture of what works in our guide to AI agents for Australian businesses. If the five do not have good answers, the chat box can wait.

Goes well with

• AI agents for ERP in Australia: what to build once the question box has earned trust, and what a scoped pilot costs.
• MCP for ERP: how to connect Claude, Copilot or your own model to Odoo or MYOB Acumatica through a scoped user today.
• AI agent governance for Australian businesses: the controls to have in place before the first prompt runs on customer data.

Thinking about a plain-English layer over your ERP?

Auboros designs and runs drafts-first AI agents over Odoo and MYOB Acumatica from Brisbane, with role-scoped permissions and human approval on anything that writes. If you want to scope conversational ERP for your business, book a free consultation. We’ll tell you what’s shipped, what’s roadmap, and what’s worth waiting for.