MYOB Acumatica AI Automation: How a Prompt Hands the Work Back to a Person

MYOB Acumatica AI Automation cannot post a document. A prompt runs only when a person clicks its command, it writes only fields that person could already edit on that form, and the record still has to be saved and released by a human. As at 22 September 2026 that is the approval gate, and this note shows how to design around it, like our AI agents for ERP in Australia.

At a glance

• Who it is for: finance, sales and service teams on MYOB Acumatica 2025.2 or 2026.1 who want to use the built-in AI Automation feature and need to know, before switching it on, exactly where a person stays in the loop.
• Problem it solves: the marketing copy says “automate approvals and document processing”; the form reference says a command a person clicks that fills in editable fields. Both are true in different senses, and the gap between them is where governance questions get asked in the wrong meeting.
• Platforms: MYOB Acumatica only (the ANZ product, versions 2025.2 and 2026.1); the global Acumatica release runs ahead of it and is labelled separately below. Odoo 19’s equivalent is covered for comparison.
• Trigger: a person opens a record on the source form (Bills and Adjustments, Sales Orders, Cases, Stock Items, Service Orders and so on) and clicks the prompt’s command on the More menu or its toolbar button.
• Inputs read: only fields on that one record, plus repeating lines the prompt author has chosen to include; no related records, no other documents, no web content.
• What the prompt drafts: text into one or more output fields on the same form, such as a bill description, an order acknowledgement note, closure notes on a case or a standardised problem description on a service order.
• Approval gate: the person who clicked reads the result in the field, edits it, regenerates it or discards it, then saves. Releasing, posting or sending the document is a separate action that AI Automation does not perform.
• Audit trail: the AI Automation History form logs every connection test, prompt test and command run with its token counts; the Trace page shows the exact text sent to and returned from the model, including what was masked.
• What it never does: create a new document, record or detail line; change a related record; release or post anything; run on a schedule or on a workflow event without a person; see fields the prompt author did not include.
• Typical cost: the AI Automation feature is licensed separately (MYOB’s documentation says to consult the sales policy, no price is published); the MYOB-managed model has no separate API bill; a scoped design-and-governance engagement with us starts from $5,000 ex GST per the AI agents for ERP in Australia hub, verified 22 September 2026.
• Typical timeline: an administrator can activate a built-in prompt in an afternoon on 2026.1; a custom prompt with masking, system instructions and a security review typically takes 1 to 2 weeks including testing on real records.
• Last verified: 22 September 2026 against MYOB’s LLM provider integration overview, prompt creation guide, data masking guide, system instructions guide and MYOB’s AI page.

This week we wrote up how to make an Odoo 19 AI server action wait for a person, because Odoo’s documentation says a called tool runs unconditionally. This is the MYOB Acumatica counterpart, and the answer is almost the opposite. AI Automation is built so that it cannot get ahead of the person, which is reassuring for a CFO and mildly frustrating for anyone hoping to automate a whole workflow. Knowing which of those you are before you build saves a lot of prompt engineering.

What is an AI Automation prompt in MYOB Acumatica?

An AI Automation prompt is a saved set of instructions, attached to one data-entry form, that sends selected fields from the open record to a large language model and writes the model’s answer back into selected fields on the same record. MYOB’s integration overview describes the feature as generating text “to help the user with the specific task”, with examples such as stock item descriptions and case closure notes. It launched as AI Studio; MYOB’s AI page and knowledge base now call it AI Automation, and the Enable/Disable Features screen still shows both names.

The moving parts, using MYOB’s own form names so you can find them:

• LLM Connections (ML201000): the link to a model provider. On 2026.1 the MYOB-managed connection, which runs through AWS Bedrock with no API key to manage, is available immediately; on 2025.2 an administrator sets it up first, per MYOB’s setup article. You can instead bring your own Azure, Amazon Bedrock, OpenAI or Anthropic connection.
• LLM Prompts (ML202000): where a prompt is written, tested against a real record and activated, which adds the command to the source form.
• LLM System Instructions (ML203000): reusable safety, security and house-style text that a security expert maintains and that is inserted into every new prompt by default.
• AI Automation History (ML400000): the log of every test and every run, with token counts.

Requests travel through a MYOB-run proxy that translates them into the provider’s API format and passes the response back; MYOB’s documentation says the proxy does not store or retain the data and that existing access controls, permissions and audit logging apply before anything is sent. That’s the plumbing. Now the part buyers ask us about.

Where exactly does the person come back in?

Four places, and none of them is optional. This is the piece of the design you don’t have to build, because MYOB built it in.

The click. A prompt does nothing until someone opens a record and chooses its command. MYOB’s prompt guide says the command appears on the More menu of the source form once the prompt is active, and Acumatica’s product management confirmed in its 2025 R2 AI Automation FAQ that automation “is always user-initiated through an action button”. No schedule, no event trigger, no batch mode.

The field. The result lands in an editable field on the form, in front of the person who clicked. MYOB’s built-in prompts guide states the rule plainly: the generated command can modify only the output fields you can edit manually on the form. If your role can’t edit the description on a released bill, neither can the prompt.

The edit or the retry. MYOB’s own instructions to the user are to edit the generated text if it isn’t right, or click again for a new version. The text sits in the field unsaved until the person decides.

The save and the release. Saving the record is a separate step, and releasing or posting a financial document is a further one that AI Automation does not touch. A generated bill description reaches the ledger only when the AP officer releases the bill, exactly as if she had typed it.

So AI Automation is a drafting tool with the review already wired in, not a workflow engine. If you want a prompt to run when a case is created, route it to a queue and wait for approval, that is an external agent working through the API under a scoped user, the pattern our MCP for ERP guide covers.

What can a prompt write, and what can it never touch?

The limits are narrower than the marketing suggests, and for a first AI feature inside an ERP that’s the right way round.

• Writes: existing fields on the open record, including text boxes and description tabs, and user-defined fields where the version supports them.
• Reads: fields on the same record and its repeating lines (bill lines, order lines, case activities) when the prompt author includes them with a foreach block.
• Cannot create: a new document, record or detail line. Acumatica’s 2025 R2 FAQ lists creation as unsupported, and MYOB’s current prompt library contains no prompt that creates anything.
• Cannot reach: related records. A prompt on a sales order can’t read the customer’s credit history or write to the customer record; the FAQ says automation operates within the current document only.
• Cannot browse: the web or attached files. Acumatica’s product manager said in January 2026 that web search wasn’t available and file processing was roadmap with no date; treat both as roadmap until MYOB’s release notes say otherwise.
• Cannot post: anything. There is no release, no approval and no send in the feature.

Those lines are worth a paragraph in your AI usage policy, because they answer the question your auditor will ask. An AI Automation prompt cannot create a journal, change a tax code or move stock. Its blast radius is a text field on a record a named user was already editing.

Who approves the prompt itself, and who can run it?

There are two approval gates in this feature, and people usually only notice the first. The gate on each run is the person who clicks. The gate on the prompt is the person who activates it, and MYOB has thought about that one too.

MYOB’s system instructions guide describes two roles. A Security Expert writes and maintains the reusable system instructions (tone, what never to expose, response format) and reviews, tests and approves prompts. A Prompt Engineer writes the prompts, inserts the system instructions and tests against real records. Default system instructions land in every new prompt automatically, and the engineer can move or trim them without changing the master copy.

Testing is built into the form. You pick an Example Record, the system shows the exact prompt it will send with that record’s real values in place of the placeholders, you click Test Prompt, and you see the raw response and a preview of how it lands on the form. Only when you tick Active and save does the command appear for users, and deactivating removes it straight away. Anyone with access to the source form can then run it; MYOB’s AI page says AI features respect the permissions and governance settings already in place.

For a Queensland business that means a prompt goes through the same change control as any other configuration: a named owner, a test tenant, a sign-off, an entry in the AI agent register and a review date, because MYOB warns that providers retire models with little notice and a connection can stop working.

A worked example: the AP bill description that survives an audit

MYOB’s prompt library includes a finance prompt called Standardise AP Bill Description, and it’s a good first prompt because it touches money without touching the ledger. Here’s how it runs at a wholesale distributor with a two-person accounts payable team.

A supplier’s tax invoice for cleaning consumables arrives and the AP officer keys it on the Bills and Adjustments form: supplier, invoice number, date, lines, tax codes. The description field says “INV-90431”, which is what everyone types when they’re busy. She clicks Generate Bill Description on the More menu. The prompt sends the supplier name, the existing description, the invoice number, the total, the currency and each line’s description, amount and account. The model returns one plain-text line, something like “Cleaning consumables and paper products from the supplier, total AUD 1,284.50 (inv INV-90431)”, and it lands in the description field.

She reads it. The prompt’s instructions tell the model to invent nothing and editorialise about nothing, but she checks anyway, because the line will be what the approver, the auditor and next year’s AP officer see when they search. She trims it, saves the bill and sends it for approval. The approver sees a bill they can recognise at a glance instead of an invoice number. Then, and only then, the bill is released.

What the prompt did not do matters more. It never saw or changed the tax codes, the GST amount or the account coding, because those weren’t in its output fields. The officer still checked the supplier’s document against the ATO’s tax invoice requirements, which for a purchase of $1,000 or more include the buyer’s identity or ABN, before the bill was released. Her Business Activity Statement (BAS) is built from the released bills, and nothing the model wrote touched it.

“Every write-back we design for a client needs an approval gate built into it. MYOB Acumatica’s AI Automation is the rare case where the gate is already there, because the feature can’t do anything except fill in a field for the person who asked. The work is deciding which fields, and who gets to ask.”

Josh Craig, Director, Auboros

What does it send to the model, and how do you keep data at home?

This is the CFO’s second question, and MYOB’s documentation answers it more precisely than most vendors do.

Where the data goes depends on the connection. MYOB’s AI page says customer data is stored in Australia; depending on the feature, AI processing happens in Australia or through global infrastructure operated by Anthropic, AWS and Microsoft, and those providers can process the data to generate a response but cannot store it or use it for training. The same page says AI Automation runs on MYOB’s managed infrastructure with models available through AWS Bedrock. Bring your own connection and the terms and data-use policy are between you and that provider; MYOB is the integration layer only.

What goes in is what the prompt author put in, field by field, and nothing else. MYOB’s data masking guide adds a second control: mark a field as a Masked Data Field and the system replaces its value with a MASKED placeholder before sending, then swaps the real value back into the response. The Trace page shows both versions, so you can prove what left the building. Masking works on lists of lines too.

One limitation is worth underlining because it bites the most obvious use case. You can’t mask a value that sits inside a larger text field, such as an email body. If a customer’s phone number is in the thread you’re asking the model to refine a reply to, masking the phone field elsewhere in the prompt won’t remove it from the thread. For the built-in Refine Reply prompt the email content goes to the model as written, and your privacy assessment should say so.

That assessment is not optional for customer or employee data. Names, addresses and purchase histories are personal information under the Australian Privacy Principles, and the OAIC’s guidance on commercially available AI products expects you to know what an AI feature does with personal information before you switch it on. Our governance guide has the one-page policy template we use.

How does this compare with Odoo 19’s AI server actions?

The two platforms made opposite bets, and if you run one of each, or you’re choosing between them, the difference decides how much governance you have to build yourself.

Odoo’s approach gives you more reach and more responsibility; our approval-gate build note shows how to keep the write in draft so a person still posts it. MYOB’s approach gives you less reach and almost no governance to build, which is why we tell MYOB clients to switch on two or three library prompts first and read the AI Automation History before asking for anything cleverer. When a workflow has to run without a click, on either platform, we build an external agent over the API with role-scoped access and drafts-first writes; the conversational ERP guide explains how that sits beside the native features.

Watch-outs before you switch it on

• Some prompts overwrite the source text. MYOB’s Standardise Problem Description prompt writes its clean version back into the same description field on the service order and says plainly that the raw notes are not retained. The dispatcher reviews before saving, but once saved the original intake text is gone. Keep the raw notes in a separate field or accept that the standardised version is the record.
• Licensing and version. The feature sits under Experimental Features on the Enable/Disable Features screen and is subject to licensing; 2025.2 needs the managed connection set up by an administrator, 2026.1 doesn’t. Check both before promising a go-live date.
• Model churn and token spend. MYOB warns that providers deprecate models with little notice, so put a quarterly connection test in someone’s calendar. A bring-your-own connection also carries its own API bill, and the History form is where a badly written prompt shows up burning input tokens on every click.
• Prompt drift. A prompt engineer can edit an active prompt at any time. Keep the instructions versioned outside the tenant, a text file in your change-control folder is enough, so the security expert can see what changed.
• The marketing gap. “Automate approvals” on MYOB’s AI page describes what a developer can build by customising the action behind a prompt, not what an activated library prompt does. Out of the box it’s a hand-off, and that’s a feature.

Goes well with

• MYOB Acumatica AI features: what’s real in 2026: the next question is which of the eight AI features are on your version and which are still global-only.
• Odoo AI server actions: building the approval gate: the Odoo half of this comparison, for businesses running both platforms or deciding between them.
• AI agent governance for Australian businesses: the usage policy, permission reviews and register that a prompt like this needs to live under.

---

Want AI Automation switched on without the governance headache?

We’re a MYOB Acumatica partner in Brisbane and we design, govern and run drafts-first AI over MYOB Acumatica and Odoo for businesses across Queensland and beyond. If you’d like us to set up three library prompts in a test tenant, write the masking and system instructions, and hand you a one-page policy your auditor will accept, book a free consultation. We’ll tell you plainly which prompts are worth the licence and which aren’t.