Document AI in New Zealand

Check whether the document needs to exist at all

Before automating the reading of supplier invoices, look at eInvoicing. New Zealand and Australia share a framework where invoices arrive as structured data rather than as a PDF that has to be interpreted.

Where a supplier can send one, there is nothing to extract and nothing to get wrong. It will not cover all of your suppliers, and it may cover the largest ones, which is where the volume is.

A developer who mentions this before quoting to build extraction is looking after your money. It rarely removes the whole problem and it often removes the best-behaved third of it.

Do the arithmetic first

This is the AI project where the business case is easiest to prove, so prove it.

• How many documents a month?
• How many minutes does each take a person, including chasing the ones that are wrong?
• What is that person's cost per hour?
• What have errors cost in the last year — duplicate payments, wrong amounts, missed early-payment discounts, a credit note applied twice?

For most New Zealand businesses processing a few hundred documents a month, the answer justifies the project inside a year. The error reduction is frequently worth more than the time saved, and it is the part nobody counts.

Insist on a proof of concept with your own documents

Accuracy claims mean nothing until they have met your paperwork.

Clean, typed invoices in a consistent layout extract very well. Photographs taken on a phone in a truck, twenty-year-old fax templates, scanned dockets with a boot print on them and handwritten quantities do not.

Ask each firm to run a hundred of your real documents — including the ugly ones — and report accuracy per field. Expect to pay a few thousand dollars for it. That money tells you whether the business case is real, and it is the cheapest possible way to find out.

Send them the worst documents, not a tidy sample. The tidy ones were never the problem.

The review queue is the product

The goal is not to remove people. It is to have one person handle fifteen exceptions instead of typing four hundred documents.

So the screen where exceptions land matters more than the extraction:

• Does it show the original document beside the extracted fields?
• Does it highlight where on the page each value came from?
• Can someone correct a field and move on with the keyboard alone?
• Does it flag why the document was queued — low confidence, a failed validation, a supplier not recognised?
• How many can a person clear in an hour?

Ask to be shown this screen from a system they have already delivered. A demonstration of extraction accuracy tells you far less.

Validate against what you already know

The strongest accuracy gains come from checking the extraction against your own data rather than from a better model.

• Does the supplier exist in your system, and does the bank account match?
• Do the line items add to the total, and does the GST come to 15 percent of the right base?
• Does the invoice match an open purchase order, at the right price?
• Have we seen this invoice number from this supplier before?
• Is the date plausible?

Each of those turns a confident wrong answer into a queued exception. That last check — duplicate invoice numbers — pays for a surprising amount of the project on its own.

Corrections should make it better

When someone fixes a field in the review queue, that correction should be worth something. Ask what happens to it:

• Is it recorded so accuracy can be measured over time?
• Does the system learn this supplier's layout, so the same field stops being wrong?
• Can you see whether accuracy is improving or drifting?

A system that makes the same mistake every month on the same supplier's invoice is one somebody will stop trusting.

Keep the originals

IRD requires business records for seven years, and the original document is the record. The extracted data is a convenience.

So the design has to store the original, link it to whatever record it produced, and let someone retrieve it years later. Ask where those files live, whether that location satisfies your retention policy, and — where the documents contain personal information — whether it satisfies your privacy obligations too.

Start with one document type

Pick the highest-volume, most consistent document you have. Usually supplier invoices. Get that working properly, with the review queue and the validation, and run it for two months.

You will learn what your documents are really like, the second document type costs a fraction of the first, and you will have a real accuracy figure rather than a projected one.

Comparing five quotes

Read for:

• Who asked to see your actual documents before quoting?
• Who offered a proof of concept on your own paperwork?
• Who described the review queue rather than only the extraction?
• Who proposed validating against your existing records?
• Who mentioned eInvoicing?

Describe the documents, how many a month, what format they arrive in, what happens to them now, and where the data has to end up. Developers who have shipped this will ask for your ugliest twenty documents before they price anything.