The AI Replenishment Agent: How It Drafts Purchase Orders in Odoo and MYOB Acumatica

Most wholesale and distribution businesses live with a strange contradiction: too much stock of some products and not enough of others, at the same time, in the same warehouse. The reorder report says one thing, the sales team says another, and the buyer spends Monday morning reconciling both in a spreadsheet. This is the job an AI replenishment agent is built for, and it’s one of the capabilities Australian businesses ask us about most.

An AI replenishment agent is software that reads sales velocity, stock levels, supplier lead times and minimum order quantities from your ERP, then drafts purchase recommendations for a human buyer to review. It doesn’t place orders on its own. It does the analysis a good buyer would do if they had three spare hours per product line, and it presents the result as a draft with its reasoning attached.

The appetite is real. MYOB’s ERP Trends research found 78% of wholesale distribution businesses plan to use AI in their ERP, and the Australian Bureau of Statistics reports that 12% of Australian businesses used AI in 2024-25, up from 1% two years earlier. Most of that adoption is still chatbots and document drafting. Replenishment is where it starts touching real money, which is exactly why the design matters.

What an AI replenishment agent actually does

The pattern has four verbs, and only three of them belong to the agent.

• It reads. Sales orders by product and period, current and incoming stock, supplier lead times as quoted and as actually delivered, minimum order quantities, carton multiples, landed cost and margin. All of it already lives in your ERP.
• It flags. Lines trending toward a stockout inside the supplier’s lead time. Slow movers tying up cash. Lines where the quoted lead time and the last six receipts no longer agree.
• It drafts. Purchase recommendations grouped by supplier, with quantities that respect MOQs and pack sizes, and a short note explaining each suggestion.
• A person approves. The buyer reviews the draft, edits quantities, rejects lines, and only then does a request for quotation or purchase order go anywhere.

That last step isn’t a training-wheels phase to grow out of. It’s the design.

What Odoo and MYOB Acumatica already do without AI

Here’s the part most AI content skips: both platforms already ship serious replenishment logic, and none of it is AI.

Odoo 19 handles this through reordering rules: set a minimum and maximum per product, and Odoo either creates the purchase or manufacturing order automatically or lists the line on a replenishment report for a person to action. Its just in time logic schedules orders using forecasted stock and the combined lead times on the product. It’s deterministic. We’ve been through the v19 Enterprise source code, and there is no AI in the replenishment path. The v20 roadmap is more AI-forward, but nothing agentic has shipped for purchasing yet.

MYOB Acumatica, built on the Acumatica platform, goes further on the statistics. Its replenishment engine supports reorder points, safety stock, seasonality factors and demand forecast models built on historical sales, and Distribution Requirements Planning plans across locations. These are statistical models rather than machine learning, which is no criticism: they’re predictable, explainable and cheap to run. For what the platform’s actual AI features cover today, see our guide to MYOB Acumatica’s AI capabilities.

If you haven’t configured these built-in tools yet, start there, not with an agent. An agent reasoning over wrong lead times and missing MOQs will draft confident nonsense. We covered the groundwork in our guides to Odoo’s purchasing module and Odoo inventory management.

Where the agent adds judgement the rules can’t

A reordering rule evaluates one product at a time, against parameters someone typed in months ago. It works until the world drifts away from the parameters. The agent’s job is noticing the drift.

Rules can’t tell you that a supplier’s last four receipts averaged twelve days late, so the quoted lead time is fiction and the reorder point is too low. They can’t weigh whether topping up a low-margin line is worth the freight. They won’t spot that you’re simultaneously out of stock in Brisbane and overstocked in Melbourne on the same SKU, where the right move is a transfer, not a purchase. And they can’t read context, like the doubling of sales on three lines that might be a trend or might be one customer’s one-off project.

Here’s what that looks like in practice. Monday, 6am, the agent runs against live ERP data with read-only access. By 7am the buyer has a draft in their inbox: fourteen lines need attention this week. Nine are routine top-ups, grouped into three supplier orders that respect each supplier’s MOQ. Two are early orders because of that supplier with the drifting lead times, with the receipt history attached as evidence. Three are flagged as questions rather than orders, because sales velocity doubled last month and the agent can’t tell whether that’s repeatable, so it presents both order quantities and asks. The buyer approves nine, adjusts one, rejects one, and answers the three questions. Total elapsed time: about twenty minutes, most of it on the three lines that deserved human thought.

“The reorder point tells you one number went below another number. It can’t tell you the supplier has been shipping late for three months, or that the margin on a line no longer covers the freight to restock it. That judgement layer is what the agent drafts, and it’s exactly why a buyer still signs off on every order.”

Bill Alvarez, Practice Manager, Auboros

Why a buyer approves every order

Language models make errors. They misread a spreadsheet column, treat a spike as a trend, or hallucinate a pack size. On a chatbot that’s an annoyance. On a purchase order it’s a container of stock you didn’t need, paid for with real cash and stored in real racking.

The approval gate is what makes the agent safe to run at all, and in practice it means three things. The agent’s access is read-only and scoped to the data a buyer could see anyway, never a shared super-user login. Everything it produces lands as a draft, and the ERP’s own workflow handles release. Odoo Enterprise even ships an Approvals app that generates the RFQ from an approved purchase request, so the gate is native to the platform rather than bolted on. And every recommendation carries its reasoning and sources, so when the buyer disagrees, you can see why the agent got it wrong and fix the data or the prompt.

This is the same drafts-first pattern we apply across our AI agent work: agents propose, humans approve, and everything is auditable.

How to start without betting the warehouse

Don’t switch this on across your whole catalogue. Pick one category or one supplier group, ideally one with steady demand and a co-operative supplier. Check the data first: real lead times, current MOQs, correct costs. Then run the agent’s recommendations alongside your existing process for a full ordering cycle and measure the gap between what it drafted and what the buyer actually ordered. Where they differ, work out which one was right.

If the agent earns trust, widen the category coverage. If it doesn’t, you’ve lost nothing except a few weeks of parallel running, and you’ve usually found data problems worth fixing anyway. For wholesale distributors starting from scratch, this pairs naturally with a platform foundation like the MYOB Acumatica FastStart program, which gets the underlying purchasing and inventory data structured properly in the first place.

---

Thinking about a replenishment agent for your warehouse?

We’re a Brisbane-based ERP consultancy that implements Odoo and MYOB Acumatica across Queensland and beyond, and we run drafts-first AI agents over both platforms. If you want to know what a replenishment agent could draft from your own sales and supplier data, book a free consultation. We’ll tell you plainly whether your data is ready for one.