Automating supplier invoices with OCR in your ERP

The short answer

OCR (optical character recognition), usually combined today with AI-based field extraction, can read a supplier invoice, create a draft purchase invoice in ERPNext or Prometheus, match it to the purchase order and receipt, and route it for approval. In a well-designed set-up, clean invoices that match need only a quick confirmation, and a person's time goes into the mismatches, duplicates and unreadable scans. It is one of the most reliable automation wins in an ERP, but only if you design the review step, the matching rules and the tax fields carefully. Capture that posts straight to the ledger without review is not automation; it is a new source of errors.

How it works in an ERP

A typical flow in ERPNext or Prometheus looks like this:

1. Intake. Invoices arrive at a dedicated accounts payable mailbox, a scanner folder or an upload screen.
2. Extraction. The OCR or extraction service reads the supplier, invoice number, date, line items, subtotal, GST/HST or QST, and total.
3. Draft record. A draft purchase invoice is created, with the original file attached.
4. Matching. The draft is matched to the purchase order and purchase receipt; quantity, price and tax differences are flagged.
5. Review and approval. Clean matches go to the accounts payable clerk to confirm. Mismatches, new suppliers and invoices over a threshold go to an approver.
6. Posting. A person submits the invoice, and it is scheduled for payment.

The Canadian details that matter

Generic OCR tools are often trained on US invoices. In Canada, your capture needs to handle a few things properly.

Sales tax fields. Invoices may show GST, HST, QST or PST, sometimes several at once, and some suppliers show tax only as a total. Your extraction should capture each tax separately where it appears, because your ERP's tax templates and your input tax credit claims depend on it.

Information required for input tax credits. The CRA sets out the information you need to support an input tax credit, in tiers by amount. Under $100, you need the supplier's name, the date and the total. From $100 to $499.99, you also need the GST/HST charged and the supplier's GST/HST registration number. At $500 or more, you also need your business name, a description of the goods or services, and the terms of payment. Designing capture around these fields means missing details are flagged when the invoice arrives, not during an audit.

Bilingual invoices. Suppliers in Quebec and elsewhere may send invoices in French. Test extraction on French documents, including accented supplier names and date formats.

Keeping the original. The CRA explains when imaged paper records can replace the originals, which depends on following the relevant national imaging standard, in its guidance on imaging paper documents. Invoices received as PDFs are already electronic, but you still need to keep them. Attach the original file to the ERP record, keep it for the required period, and back it up.

An audit trail. The CRA may review your business systems and expects a trail from source document to financial records. Your set-up should show who confirmed or changed each extracted value.

This is general information, not tax advice; confirm your requirements with your accountant.

What usually goes wrong

• Automating before cleaning supplier data. Duplicate supplier records and missing tax numbers make matching unreliable. Clean the supplier list first.
• No purchase orders to match. If most spending has no PO, OCR saves typing but cannot validate much. Decide which spending categories should require a PO.
• Trusting confidence scores blindly. Extraction tools report how sure they are, but a confident misread of a total is still wrong. Key fields such as total, tax and invoice number should always be visible to the reviewer.
• Ignoring the long tail. Handwritten invoices, poor scans and multi-page statements will never extract cleanly. Give them a manual-entry route rather than forcing them through.
• Sending data somewhere you have not approved. Many extraction services are cloud-based. Know where documents are processed and stored, and whether that fits your privacy obligations.

How we build it

We start by collecting a sample of your real invoices and agreeing what "done" looks like. We then choose an extraction approach that suits your volume, your languages and your data-location requirements, and build the capture, matching and approval flow in a separate custom app, so it survives ERPNext upgrades. Every build is tested against your own documents, including the awkward ones, and failures alert a named person rather than failing silently. The same approach extends to receipts, expense claims and other forms; see ERPNext and Prometheus customization.

When to bring in help

If you receive a few dozen invoices a month from a handful of suppliers, a standard add-on or careful manual entry may be enough. Bring in help when volumes are high, when invoices must be matched to purchase orders and receipts, when you need QST and GST/HST handled correctly, or when an auditor will ask how you know each invoice was checked. That is where design, testing and accountability matter more than the OCR engine itself.

Sources and further reading

Product capabilities and guidance change. These are the primary sources this article relies on, checked on the review date above.

1. Input tax credits, Canada Revenue Agency
2. Acceptable format, imaging paper documents and backing up electronic files, Canada Revenue Agency
3. Review of business systems and keeping audit trails of business transactions, Canada Revenue Agency
4. Purchase Invoice, ERPNext documentation (Frappe)

This article is general information, not legal, accounting or security advice for your specific situation. Examples are hypothetical unless stated otherwise.