AI Agents & Automation · Brisbane, Queensland
AI agents for your ERP, with a human in the loop
We design, govern and implement AI agents for businesses that run Odoo or MYOB Acumatica: agents that read inbound orders, flag reconciliation exceptions, chase suppliers and draft replies. Every one of them waits for a person to approve. We run agents in production on our own operations, so the advice comes from experience, not a slide deck.
How every Auboros agent is built
- ✓ Drafts-first: a person approves every write
- ✓ Role-scoped access, never a super-user login
- ✓ Full audit trail of reads, drafts and approvals
- ✓ Sandbox-proven before production
- ✓ Bring-your-own-LLM via MCP, so model choice stays open
Why this practice exists
The anti-hype AI consultancy
Most AI content is written by people who have never connected a model to a live business system. We have. Auboros combines an accounting and finance background with more than a decade of software development, and we implement the two ERP platforms our agents work against: Odoo and MYOB Acumatica. That combination matters, because an agent over an ERP is only as good as the person who understands both the ledger and the code.
We connect AI tooling to ERP systems through the Model Context Protocol (MCP) and platform APIs, with role-scoped access and human approval gates on writes. Where a native platform feature is the better answer, we say so: Odoo ships useful assistive AI today and has agents on its v20 roadmap, and Acumatica's AI features reach MYOB Acumatica on the usual ANZ release lag. We track both closely in the AI & ERP blog.
And we're honest about limits. Language models make mistakes: not often, never zero. Every agent we build assumes that, which is why the approval gate is the design, not a compromise.
Agent patterns we build
Real jobs, drafted by agents, approved by people
These are the patterns Australian businesses actually ask us for, and the level of specificity you should expect from anyone proposing agent work.
Inbound orders, read and drafted
An agent watches the inbox, reads PDF purchase orders in every customer format, matches customers, products and pricing against the ERP, and drafts the sales order. A person approves it before it posts.
Reconciliation exceptions, flagged daily
Compare transactions across ERP, bank feeds, e-commerce and POS every day. The agent flags what does not agree and proposes a resolution for each exception. Your accountant applies the ones that are right.
Credit control drafts
Watch overdue invoices and payment behaviour, draft the reminder emails and a recommended action per account. Judgement about relationships stays with your team.
Supplier chasing and replenishment
Chase delivery confirmations on open purchase orders, flag orders landing late against promised dates, and prepare replenishment suggestions from sales velocity, lead times and minimum order quantities.
Service and RMA triage
Classify inbound requests, check warranty or order status against live ERP data, route to the right person, and draft the reply. Drafts-first, human-send.
Conversational reporting
Ask questions of your live ERP data in plain language, under each user’s existing permissions. "What did margin do in Queensland last month?" answered from the system, not a stale export.
Governance first
The controls that keep agents out of trouble
This is the part your accountant, auditor and board will ask about, so it's where we start. Five controls, on every build, no exceptions.
01
Role-scoped permissions
The agent acts with the requesting user’s data access. Never a shared super-user account.
02
Approval gates on writes
Anything touching the ledger, tax, pricing or stock is drafted by the agent and posted by a person.
03
Audit trails
Every agent action is logged: what it read, what it proposed, who approved it, and when.
04
Sandbox-first deployment
Agents earn production access by proving accuracy against a copy of your data first.
05
Usage policy and registry
A signed staff policy, an owner for every agent, and scheduled permission reviews so access never quietly grows.
We work to the National AI Centre's Guidance for AI Adoption and the Australian Privacy Principles, including the December 2026 automated-decision disclosure changes. Governance advisory (usage policies, risk registers, permission reviews) is available as a standalone engagement.
Honest assessment
When agent work makes sense, and when it doesn't yet
✓ Good fit: worth a conversation
- Systems
- You run Odoo, MYOB Acumatica, or another ERP with a usable API.
- Pain
- Hours going into rekeying, matching, chasing or triage that follows a describable pattern.
- Data
- Your ERP data is in reasonable shape, or you are willing to fix it first.
- Appetite
- You want drafts-first automation with human approval, not a lights-out robot.
- Scale
- Enough volume that saving hours per day matters. One PDF order a week does not need an agent.
✕ Not yet: fix this first
- Expectations
- You want an agent with unsupervised write access to your accounts. We will not build that.
- Data
- Duplicate customers, stale price lists and untrusted stock numbers. Fix the foundations first, and we can help with that.
- Systems
- A closed legacy system with no API and no export path.
- Motive
- Replacing the finance team rather than removing their repetitive work. Wrong goal, and it fails.
If the honest answer is "your ERP foundations aren't ready", that's an implementation or integration conversation first. We do those too, which is rather the point.
How we deliver
Sandbox first, drafts first, then expand
-
1
Discovery
Free consultation: which workflow burns the most hours, what data it needs, where the approval gate sits, and whether the data is ready.
1 to 2 weeks
-
2
Sandbox pilot
Build the agent against a copy of your data. Measure how often its drafts are right before it goes near production.
2 to 6 weeks
-
3
Drafts-first go-live
Production with role-scoped credentials and permanent approval gates on financial writes. Weekly log reviews in the first quarter.
1 to 2 weeks
-
4
Expand deliberately
Add the next agent under the same registry: owner, permissions, cost budget and escalation path per agent.
Ongoing
FAQ
Common questions about AI agents and ERP
What can an AI agent actually do inside an ERP system?
An AI agent reads information from your ERP and connected systems, decides the next step, and drafts the work: a sales order from an emailed PDF purchase order, a proposed match for a reconciliation exception, a supplier chase-up email. A person reviews and approves before anything posts. Agents are good at reading, matching, drafting and flagging. People keep the judgement calls.
Is it safe to connect AI to our accounting system?
It can be, if the connection is governed. We give agents role-scoped credentials (never an administrator login), require human approval on anything that touches the ledger, tax codes, pricing or stock, log every action, and test in a sandbox before production. An agent with unsupervised write access to your accounts is not a setup we will build.
Do we need Odoo or MYOB Acumatica to work with Auboros on AI?
No, but it helps. We specialise in agents over Odoo and MYOB Acumatica because we implement both platforms and know their data models and APIs well. If you run another system with an API, we can usually still help, and if your ERP itself is the bottleneck, we will tell you that first.
How much does an AI agent project cost in Australia?
A single well-scoped agent pilot, for example inbound order entry or reconciliation exception handling, typically starts from a few thousand dollars for discovery and sandbox build, with production hardening priced after the pilot proves its accuracy on your real data. We quote fixed scopes after a free consultation rather than open-ended day rates.
Will AI agents replace our staff?
No. The agents we build remove repetitive reading, rekeying and chasing, the work nobody misses. Every build keeps people on approvals and judgement. In practice teams end up doing more of the work that needs a human: exceptions, customers, decisions.
What data do the AI models see, and where does it go?
That is a design decision we make together. Options range from hosted models with contractual data controls to self-hosted open-source frameworks where nothing leaves your infrastructure. Either way, agents run under the Australian Privacy Principles: we scope what personal information an agent can read, and document it in your privacy policy where required.
Wondering what an agent could take off your team's plate?
Book a free consultation. We'll tell you what's worth automating, what isn't yet, and what it would take. If the answer is 'nothing yet', you'll hear that too.
30 minutes. No obligation. Brisbane-based team, AEST business hours.