Skip to main content
Order Fulfilment / Document IntelligenceLive within the ORCA group

Automating Order Fulfilment for a Medical Products Distributor

A Western Australian medical products distributor was processing every purchase order by hand — opening emails, reading PDFs, checking stock in their accounting system, creating invoices manually. We built an AI pipeline that does the entire process in under 60 seconds, 24 hours a day.

<60s

Per order (was 15 min)

95%

Hands-free processing

24/7

Processing window

$60K+

Annual savings (est.)

The Problem

The company distributes medical consumables and equipment across Australia. Purchase orders arrive via email as PDF attachments — between 15 and 30 per day — from hospitals, clinics, and aged care facilities. Each order follows a different format. Some are machine-generated from procurement systems. Others are scanned handwritten forms.

The manual process for each order involved five distinct steps: open the email, download the PDF, read and interpret the line items, check stock levels in MYOB AccountRight Live for every product, then either create an invoice (if everything is in stock) or compose a backorder notification email to the customer. A single order with 8–10 line items took 12–19 minutes to process.

Orders received after business hours — roughly 40% of total volume — sat unprocessed until the next morning. During peak periods (end of month, seasonal demand), the morning backlog regularly exceeded 40 emails. Staff were spending the first two hours of every day just catching up, before they could address any new orders, customer calls, or procurement tasks.

The manual process also introduced errors. Transcription mistakes — wrong quantities, incorrect product codes — occurred in an estimated 3–5% of orders, leading to invoice corrections, return processing, and eroded customer trust. Duplicate orders (from email forwards or resends) slipped through approximately 1–2% of the time.

What We Built

An end-to-end automated pipeline that monitors the orders inbox, extracts structured data from PDF purchase orders using AI, checks real-time stock availability, and generates invoices or backorder notifications — all without human intervention for standard orders.

Step 1: Email Monitoring & Classification

The system monitors the orders inbox via Microsoft Graph API webhooks. When a new email arrives, it triggers within 30 seconds. An AI classification step determines whether the email contains a purchase order (as opposed to general correspondence, supplier invoices, or marketing). Non-PO emails are left for manual handling. PO emails move to extraction.

Step 2: PDF Parsing with AI

The PDF attachment is sent to an LLM-based extraction engine that reads the document and outputs structured JSON: customer name, PO number, delivery address, requested delivery date, and every line item with product code, description, quantity, and unit price.

The parser handles the full range of PO formats the company receives — from clean machine-generated documents to scanned handwritten forms. Each extraction includes a confidence score. Orders scoring above 90% confidence proceed automatically. Orders between 70–90% are flagged for quick human verification. Below 70% are routed entirely to manual processing.

Step 3: Real-Time Stock Check

For each line item, the system queries MYOB AccountRight Live via API for current stock on hand. This happens in parallel — all items checked simultaneously rather than sequentially. The system also handles product code mismatches, where a customer references their own product code rather than the distributor's internal SKU, using a maintained mapping table.

Step 4: Invoice Generation or Backorder Notification

If all items are in stock, the system generates an invoice in MYOB with the correct pricing, tax treatment, payment terms, and delivery details. The invoice is held for the general manager's one-click approval before being sent to the customer — maintaining human oversight without human effort.

If any items are out of stock, the system automatically composes and sends a backorder notification email that clearly lists which items are available for immediate shipment, which are backordered, and expected restock dates where available. The customer can reply to proceed with a partial shipment.

Step 5: Duplicate Detection & Audit Trail

Every processed PO is logged with a hash of its contents. If the same PO number from the same customer arrives again (common when customers forward or resend), the system detects the duplicate and alerts staff rather than creating a second invoice. A full audit trail tracks every step of processing — email receipt time, classification result, extraction confidence, stock check results, and final action taken.

The Numbers

MetricBeforeAfter
Time per order12–19 min<60 seconds
Off-hours processingNext business dayImmediate (24/7)
Transcription errors3–5% of orders<0.5%
Duplicate orders1–2% slip-through0% (auto-detected)
Morning backlog40+ emailsZero
Staff time on order processing5–9 hours/day<30 min/day (approvals only)

At an average of 22 POs per day and a loaded staff cost of AUD 45/hour, the system recovers an estimated AUD 60,000–100,000 annually in direct labour savings alone — before accounting for reduced error correction, faster customer response times, and the elimination of weekend processing delays.

Technology Stack

N8N (workflow orchestration)
Claude AI (PDF parsing & classification)
Microsoft Graph API (email)
MYOB AccountRight Live API
Supabase (state & audit trail)
Webhooks (real-time triggers)

Built to Scale

The architecture was designed from the outset to support expansion. Adding a new mailbox is a configuration change, not a code change. The system currently monitors two Outlook mailboxes, but the architecture supports 20+ without modification.

Phase 2 scope includes integration with procurement portal systems for customers who submit orders through web portals rather than email — a different ingestion pathway that feeds into the same processing pipeline. The modular architecture means new input channels plug in without touching the core stock-check and invoicing logic.

Key Lessons

Confidence thresholds prevent expensive mistakes.In healthcare distribution, sending the wrong product isn't just inconvenient — it's a compliance risk. The tiered confidence system (auto-process above 90%, human-verify between 70–90%, manual below 70%) means the AI handles the easy 95% while humans focus on the 5% that actually need judgement.

Keep the human in the approval loop — but make it one click.The general manager still approves every invoice before it's sent. But instead of building the invoice from scratch, they're reviewing a pre-built invoice and tapping “approve.” The oversight is real, the effort is minimal.

Audit trails aren't optional in regulated industries. Every step is logged with timestamps, confidence scores, and decision rationale. If a customer queries an invoice, staff can trace exactly how the PO was parsed, what stock levels were at the time of processing, and why a particular price was applied.

Frequently Asked Questions

How does AI automation handle purchase order processing in healthcare?

+
The AI receives the PO as an email attachment (PDF, Word, or structured file), extracts line items using vision-capable LLMs, validates SKUs against the product catalogue, checks stock levels in the ERP (MYOB in this case), generates a confirmation with accurate delivery timing, and creates the invoice. End-to-end processing dropped from 15 minutes of manual work per order to under 60 seconds. A human only sees orders the AI flags as low-confidence — typically because the SKU is ambiguous or the quantity is unusual.

Can AI automation work with MYOB or other Australian accounting systems?

+
Yes. MYOB (AccountRight and Essentials) has a documented REST API that handles invoice creation, contact management, stock queries, and payment tracking. We also integrate with Xero, which is the more common choice for newer Australian businesses. For older on-premise versions of MYOB without an API, we use RPA or email-based triggers as a fallback.

How accurate is AI at reading purchase orders from different customers?

+
Accuracy on structured PO formats (standard customer templates) sits above 99%. Accuracy on unstructured or handwritten POs is lower — typically 85–92% on line-item extraction. We set confidence thresholds per customer: anything below the threshold is routed for human review rather than processed autonomously. The system also learns customer-specific quirks over time.

What is the ROI of automating order fulfilment for a medical distributor?

+
For the distributor in this case study, recovered staff time alone was worth approximately $60,000 AUD per year. Additional benefits — faster order confirmation leading to happier customers, zero pricing errors on invoices, and the ability to handle volume spikes without hiring temp staff — are harder to quantify but significant. Payback on the build investment was under four months.

Is customer data secure when AI processes medical supply orders?

+
Yes. We use enterprise LLM endpoints with zero-retention policies (no training on your data), encrypt credentials at rest, and host in Australian data centres. For TGA-regulated products or any workflow touching Protected Health Information, we can deploy a fully self-hosted configuration where no data leaves the customer network. Full audit logs of every AI action are retained for compliance review.

Processing orders manually?

If your team spends hours processing documents, checking stock, and creating invoices — we can automate it. Book a call and we'll scope your specific workflow.

Book a Free Call →