How AI is changing Accredo workflows

There’s a lot of breathless writing about AI in business right now. Most of it is irrelevant to anyone running a real ERP. This post is the opposite — a grounded look at where AI genuinely earns its place inside an Accredo business in 2026, and where it doesn’t.

The policy backdrop matters. New Zealand’s first national AI strategy, Investing with Confidence, was published by MBIE in July 2025 — adoption-focused and explicitly aimed at the SME gap. Public reporting around that strategy noted that while a majority of larger NZ enterprises are using AI in some form, most SMEs still aren’t. The opportunity is the gap between those two numbers.

What’s actually working

Three workflows are in production at Accredo customers today and have a clear ROI.

1. PO automation (PO documents → Accredo OE drafts)

This is the highest-leverage AI use in Accredo, full stop.

The pre-AI workflow: a customer emails a PDF PO. Someone in your AR/CSR team opens it, types the customer name into Accredo OE, finds the right debtor record, types each line item, matches it against IC, calculates the total, saves the draft, sends a confirmation. About 4–8 minutes per PO. Across a team of three doing 30–40 a day, that’s most of someone’s full-time week.

The AI workflow: the inbox is watched. PDFs (or scanned forms, or inline-email POs) get parsed automatically. The customer is matched against AR by name + ABN/NZBN + prior orders. Line items are mapped to IC SKUs with a confidence score. A draft order is created in OE with the right pricing applied (including Special Pricing). Exceptions — a SKU that doesn’t match, an unusual quantity — are flagged in plain English. Your team approves the drafts.

Real customer outcome: 40 POs/day → 15 minutes of approvals. The team’s afternoons come back. The error rate goes down because nobody’s keying numbers any more.

This is what Rapido’s PO2Order product is.

2. Invoice and remittance matching

The second-highest-leverage AI in an Accredo business is in AP.

The pre-AI workflow: supplier sends you an invoice; later they send a remittance. Someone has to match the invoice to the original PO (which you raised when you ordered), then later match the remittance to the invoice. Lots of transposing, lots of “is this the same supplier under a different name?”, lots of phone calls.

The AI workflow: invoices and remittances are matched automatically against POs and prior invoices. Exceptions — a price difference, a partial payment, a duplicate — surface with the reason. The matches just settle.

Customer outcome: Brittain Wynyard’s CFO told us remittances used to eat half their AP team’s week; now exceptions are the only thing they touch.

3. AR collections (drafted, not auto-sent)

The third workflow is at the edge of what most NZ businesses are willing to automate, but it works.

An AR-collections agent reads the customer history (when did they last pay, what’s their typical pattern, are they normally chatty or terse) and drafts a polite chase email at the right tone for each overdue customer. It doesn’t send. It drafts, you sign off, it sends.

Even with the human in the loop, you go from 30 minutes/day of writing chase emails to 30 minutes/week of approving them. And the tone is consistently better than what most teams produce when they’re tired.

Where AI doesn’t (yet) earn its place

Equally important — three places where AI in an Accredo business is mostly hype.

General “AI assistants” inside the ERP

A chatbot that lives in a sidebar of the Accredo desktop app and tries to answer arbitrary business questions is, today, more frustrating than useful. The desktop UI doesn’t surface context cleanly, the questions you actually want to ask cross modules in ways the chatbot can’t reason about, and you’ll catch yourself opening a SQL window or a report instead.

If you want a conversational interface to your Accredo data, the better answer is to expose Accredo as an MCP server and let your team use Claude or ChatGPT directly — they’re better at general Q&A than any in-ERP chatbot. That’s what the Rapido MCP server does (see Connecting Accredo to Claude and Connecting Accredo to ChatGPT).

Forecasting and predictive ordering

These products demo well and rarely survive contact with reality. An NZ wholesaler’s stock requirements depend on weather, market events, the Australian dollar, what your three biggest customers happen to be running, and what your sales rep promised on a phone call last Tuesday. No model has those signals.

The exception: very seasonal, very high-volume, very stable categories. Beverages, building supplies during certain windows, basic dry goods. If that’s you, the maths can work. Otherwise the human still beats the model.

Bot-on-bot purchasing

Some startup is going to pitch you “agent-driven procurement that orders from your suppliers automatically.” Don’t. The risk profile is wrong. A bug in a procurement agent costs you a six-figure mistake before anyone notices.

A pragmatic framework

For each AI workflow you’re evaluating, ask three questions:

What does failure look like? If the worst case is “we miss a draft order and someone has to type it manually” — fine, ship. If it’s “we ship 1,000 units we don’t have” — don’t.
Is the human in the approval seat? Drafts are great. Auto-approvals are scary. The Rapido products all keep you in the approval seat by default.
Is the data already in Accredo? If yes, you can start tomorrow. If no — if it’s also in Excel, in someone’s Outlook folder, in a manual spreadsheet — fix the data layer first. AI doesn’t fix bad data.

Where Rapido fits

We ship the three workflows above as production products. We don’t ship forecasting, predictive ordering, or auto-purchasing — partly because they don’t work well enough, partly because we’d rather be reliable than impressive.

If you want to talk about which workflow makes sense for your specific Accredo setup, book a 30-minute call. We’ll tell you which AI is worth turning on first, and which to wait on.