The AI Service Triage Agent: How It Sorts Customer Enquiries and Drafts Replies in Odoo and MYOB Acumatica

The support inbox at a wholesale distributor looks much the same on any Monday morning: sixty-odd unread emails, a handful of website form submissions and a voicemail light. “Where’s my order” sits next to “please resend the invoice” sits next to “the unit arrived with a cracked housing and the site is waiting on it”. Whoever opens the inbox first does the sorting in their head, answers the easy ones, and the cracked housing waits behind twelve invoice requests because it was politely worded. Nothing in that pile is hard. All of it is slow, and the slow part is the sorting.

That sorting is the job an AI service triage agent is built for. This post covers what the agent reads, what it drafts, what Odoo 19 and MYOB Acumatica already ship for support teams, and the Australian rules that decide why a person still sends every reply.

What an AI service triage agent does

An AI service triage agent is software that reads each inbound customer enquiry, works out what kind of request it is and how urgent it is, finds the matching records in your ERP, and drafts two things: a categorised ticket for your team and a suggested reply for the customer. It reads, sorts, looks up and drafts. It does not send, refund, credit, reschedule or promise anything until a person approves.

The difference between this and the helpdesk AI tools you’ll see advertised is where the facts come from. A generic triage tool sees the email. An agent wired to your ERP also sees the sales order, the delivery status, the invoice, the credit hold, the serial number and the warranty date, which is most of what the customer is asking about. Support questions in a product business are usually ERP questions wearing an email.

Triage covers the whole front door: order status, invoice copies, pricing and stock questions, account changes, complaints and returns. Returns and warranty claims then branch into their own workflow, which we covered in the RMA and warranty triage agent post. This post is about the sorting layer in front of it.

A worked example: one email, four records, one approval

An email lands on the support alias at 7:52am from a site manager at a customer account: the pump ordered last week has arrived with a cracked housing, the site is due to start Thursday, and could someone call. Here is what the agent does before anyone has opened the inbox.

1. Reads the message and identifies the customer. It matches the sender’s domain and name to the customer record, then finds last week’s sales order, the delivery that shipped against it and the invoice raised on dispatch.
2. Classifies and prioritises. Damaged on arrival, a live site deadline, a request for a call. It proposes the category “damaged goods”, the priority “urgent” and the field service team, because Thursday is two days away.
3. Checks what the ERP knows. Stock of the same pump at the nearest warehouse (two on hand), whether the account is on credit hold (it isn’t), whether the carrier has logged a damage claim (it hasn’t), and the serial number on the delivery so the warranty branch can pick it up later.
4. Drafts the ticket and the reply. The ticket carries a three-line summary, the linked order, delivery and invoice, and the proposed category and priority. The reply acknowledges the damage, says a replacement can ship today from the nearest warehouse once the team confirms it, asks for two photos of the housing and the carton, and promises a call before 10am.
5. Waits. The ticket sits in a stage called “AI triaged, awaiting review”. A support person reads the summary, checks the request for photos makes sense, changes “before 10am” to “before 9:30am” because they can, and approves. The reply goes out under their name and the ticket moves to the field service team.

The agent has done the reading, the looking up and the writing. The person has done the deciding, which is the only part that needed them. The twelve invoice-copy requests in the same inbox got the same treatment in parallel, each with a drafted reply and the PDF attached from the ERP, and the person approved those as a batch.

What Odoo 19 gives you today

Odoo’s Helpdesk app does a lot of the plumbing without AI. Each team gets an email alias that turns incoming messages into tickets, plus a website form and live chat as channels, per the receiving tickets documentation. Tickets carry four priority levels from low to urgent, service level agreement deadlines can be set per team, and new tickets can be assigned automatically by equal share, by fewest open tickets, or by tag.

On top of that plumbing, Odoo 19 ships three AI patterns that apply to tickets, described on Odoo’s own AI in support workflows page. All three are shipped, and we’ve checked them against the v19 Enterprise source.

• AI agents on the ticket. A configurable agent with a system prompt and knowledge sources (product guides, FAQs, Knowledge articles, PDFs) can be opened from a ticket record, with the ticket as context. Tick “Restrict to Sources” and it can only answer from the documents you gave it. Without topics assigned it can only provide information, which for a first deployment is exactly what you want.
• AI automations on ticket creation. An automation rule can run when a ticket is created and use a prompt to fill a field: the category, the priority, a one-line summary. Odoo’s documentation is blunt that the AI receives text extracted from fields and produces an output from the prompt. It doesn’t know your business rules, so keep the fields it writes advisory and let your stages do the gating.
• AI fields. A field on the ticket whose value is generated by a prompt, such as a three-bullet summary of a long customer email. A daily scheduled action fills empty AI fields, and a person can refresh one on demand.

Two caveats. The AI live chat agent that answers website visitors is real, and it’s instructed to say when it lacks information rather than guess, but its built-in escalation creates a CRM lead rather than a helpdesk ticket; a human operator still turns a chat into a ticket. And Odoo’s AI app talks to OpenAI or Gemini, with Odoo.sh and on-premise databases needing your own API keys, so ticket text leaves your database for a provider you chose. Odoo 20 will be shown at Odoo Experience on 24 to 26 September 2026; until it ships, treat anything beyond the above as roadmap. Our AI in Odoo guide covers the wider picture app by app.

What MYOB Acumatica gives you today

MYOB Acumatica handles support through Cases. Each case has a class, a business account, a contact, an owner and an optional workgroup, and incoming email can be set up to create cases, per MYOB’s managing cases job aid. Assignment, release and closure are steps on the case form, done by a person or by the rules you configure.

The AI layer is AI Automation, the feature MYOB announced in April 2026 as AI Studio and has since renamed. It lets an administrator add a prompt as a command on almost any form, and MYOB ships two built-in prompts that matter for service, documented on its built-in prompts page: Refine Reply, which drafts or improves a reply to the latest email in a thread opened from a case, and Closure Notes, which summarises the case’s activities into the closure notes tab. Both are user-initiated. A person opens the case, clicks the command, reads the draft, edits it and sends. Nothing fires on its own, and the generated text can only land in fields the user could edit by hand.

Three details are worth knowing.

• Status. MYOB’s announcement describes an early access technology preview from May 2026, with wider availability later in the year. Version 2025.2 needs the MYOB-managed LLM connection configured by an administrator; version 2026.1 has it available without setup. Treat “wider availability” as planned until it lands.
• Where the data goes. The managed model runs through Amazon Bedrock in local data centres, which MYOB positions as keeping data in-region. You can also connect your own provider (Azure, AWS, OpenAI and Anthropic are supported), at which point the data terms are yours to check.
• Masking. AI Automation can mask sensitive fields such as account numbers and prices before a prompt leaves the system, then swap the real values back into the response. The documented limit is that it can’t mask values buried inside a larger text field, such as an email body. For triage, that means the customer’s own words go to the model as written.

Globally, Acumatica’s 2026 R1 release adds AI-generated summaries of cases with sentiment and recommended next steps. Global features typically reach MYOB Acumatica 6 to 9 months later, so plan on the built-in prompts above for now. Our MYOB Acumatica AI post tracks what has landed in ANZ.

The Australian rules behind drafts-first

Two regulators and one tribunal decision explain why the person, and never the agent, sends the reply.

The OAIC’s guidance on commercially available AI products treats putting a customer’s email into an AI tool as a use or disclosure of personal information under Australian Privacy Principle 6. Using it as part of customer service, the purpose it was collected for, is generally within bounds; sending it to a provider who can access or keep it is a disclosure you need to be able to justify, and an offshore provider brings APP 8 into play as well. The OAIC also says public-facing AI tools such as chatbots should be clearly identified to customers, and recommends against entering personal information into publicly available generative AI tools at all. A triage agent that reads from your ERP under a restricted account, drafts into your helpdesk and sends nothing meets that guidance far more easily than a support rep pasting emails into a browser tab.

The ACCC’s position on false or misleading claims covers any information a business provides, including delivery timeframes, stock availability, warranties and consumer rights, whether it comes from staff, a website or any other channel. A drafted reply that quotes the planned delivery date as a promise is a misleading claim waiting to happen. In February 2024 the British Columbia Civil Resolution Tribunal found Air Canada liable for wrong fare advice its website chatbot gave a customer, in Moffatt v Air Canada, rejecting the argument that the chatbot was responsible for its own words. It’s a Canadian small claim and Australian courts aren’t bound by it, but the reasoning is the one the ACCC already applies: the business owns what it tells customers.

Drafts-first is the cheapest way to comply with rules that already exist.

Where the AI service triage agent goes wrong

Five failure modes turn up in practice, and each has a design fix.

• The confident wrong answer. The agent’s knowledge sources include a returns policy from 2023, so the draft quotes a 14-day window that is now 30. Fix: restrict the agent to sources, date-stamp them, and make policy pages somebody’s job to maintain.
• The polite emergency. A courteous email about a gas fitting that “seems a bit loose” gets filed as a general enquiry. Fix: give the prompt an explicit list of words and product classes that force “urgent” and a human read, and never let the agent lower a priority a person has set.
• Two customers, one name. The sender matches two contacts at similar trading names and the draft quotes the other customer’s order. Fix: the agent proposes the match with a confidence note, and any draft that references order data from a weak match is held back rather than shown.
• Planned dates read as promises. The ERP’s expected delivery date becomes “your order will arrive Wednesday”. Fix: templates that say “currently scheduled”, and a rule that the agent never states a date the person hasn’t confirmed.
• The slow slide to autonomy. The vendor pitch is to let the agent handle the easy ones alone once you trust it. Fix: decide in writing which categories, if any, may ever auto-send, log every draft and approval, and review the log monthly. Our AI agent governance guide has the register and permission patterns.

“Triage done well means the right person sees the right email first, with the order already open. Once that part works, the drafting is a bonus.”

Josh Craig, Director, Auboros

How to start without betting the inbox

Start with one channel and one team, in a sandbox database first. Give the agent read access under a dedicated user whose permissions are narrower than any support rep’s, with no ability to send email or post to the ledger. Run it for a month on categorisation and summaries only, and measure two things: how often a person changed the category, and how long the oldest urgent ticket waited. Add drafted replies for one low-risk category (invoice copies are the usual pick) once the sorting holds up.

The ABS reports 12% of Australian businesses used AI in 2024-25, rising to 22% of medium-sized businesses. A support queue where the urgent email is always at the top, with the order already attached, is still unusual, and it’s a smaller project than it sounds.

---

A support inbox that sorts itself, with your team still in charge

We design drafts-first service triage agents over Odoo and MYOB Acumatica from Brisbane, for support and service teams across Queensland and beyond, as part of our AI agent services. Every reply is approved by a person before it leaves.

If your support inbox is sorted by whoever opens it first, book a free consultation. We’ll go through a week of your actual tickets with you and show which ones an agent could have triaged, and which ones it should never touch.