Import documents brought together at a port review desk

Guide

Invoice OCR testing: check the fields that matter before ERP import

A correct total can hide the wrong supplier reference. Test invoice extraction by field and line item before connecting it to an automated workflow.

Read the guide

Published

Order, delivery, acceptance and invoice records with one discrepancy

A useful OCR test checks whether the right information reaches the right fields and line items. Reading most of a document correctly is not enough if the invoice number is wrong or a quantity belongs to the neighbouring item. Before accounting import, compare the results with a verified reference and examine the errors that pass through your controls.

This protocol is for teams evaluating supplier invoice entry. For a first introduction to the technology, start with the OCR guide.

Define the expected result before running the test

Separate three outcomes: reading text, extracting structured fields and preparing an invoice for import. A tool may achieve the first while leaving substantial work before the third.

List the fields your software requires: supplier, invoice number, date, currency, amounts and purchase order reference. Include line items if your process uses them. Define what counts as correct for each field. A different date format may be acceptable; dropping a leading zero from a reference may change the identity of the record.

Also define the cases that must stay in review: incomplete documents, unknown suppliers, several possible purchase orders or unrecognised units. A missing value must remain missing, even when a model offers a plausible answer.

Build a sample that resembles your inbox

Select authorised documents in the formats your team actually receives. Separate digital PDFs, scans and phone photographs. Include several suppliers and invoices with multiple pages.

Keep a group of ordinary cases and a group of difficult cases. The second tests the limits; its error rate does not automatically represent your everyday workload. Record the number and mix of documents so the results can be interpreted.

Include:

  • a reference containing leading zeros;
  • two lines with similar descriptions;
  • an invoice with separate discounts and charges;
  • a photograph with a poorly legible area;
  • an invoice with a correct total but no purchase order reference.

Keep a test sample separate from the documents used to configure the system. Reusing only the setup examples tells you little about the next supplier's invoices.

Prepare a reference your team has checked

Record the expected values in a review sheet. For critical fields, ask someone familiar with the document to resolve disagreements. If the original is unreadable, mark the value as undeterminable and request another document. A human guess is not a reliable test reference.

For each result, retain the document, page, expected value, proposed value and reason for disagreement. Keep the raw and normalised values separate. This helps distinguish a reading error from an unsuitable conversion rule.

Count errors by field and line item

Google Document AI evaluates extraction by comparing predictions with annotations in a test set, using measures including precision and recall. This distinguishes information extracted incorrectly from expected information that was missed. Source: Document AI evaluation.

For a business review, also report straightforward counts: correct references out of verifiable references, complete lines out of expected lines, records requiring correction and critical errors that were not flagged. One overall score can conceal weak fields.

Fictional example: all ten invoice totals are correct, but two invoice numbers are misread and one quantity is assigned to the wrong item. The result for totals is good. The workflow is not ready to import all ten records without review. This example is not a measurement of Kaliits performance.

Test confidence scores and business controls separately

A confidence score helps direct review. It does not replace comparison with the verified reference. Test thresholds on your sample and count what goes to review, what is blocked and what passes incorrectly.

Include plausible but incorrect values: a valid date taken from the wrong page or a correct amount taken from a subtotal. Format checks do not necessarily detect these errors in meaning.

OCR should not decide whether an invoice has already been recorded or whether a delivery supports the invoiced quantity. Those are separate controls to test in the document workflow.

Finish with a controlled import and a recovery check

Use your software's test environment where available, or review an import file before using it. Check decimal separators, references and returned record identifiers. Correctly extracted data can still be rejected by the destination.

Then simulate an interruption after submission. Before retrying, the workflow should be able to find the created record or request a check. Measure correction and recovery effort as well as reading time.

Compare the proposed approach with the capture and import functions already available in your software. Better use of the existing system may be the right choice.

Make a decision you can explain

The report should identify the sample, weak fields, undetected errors, review workload and cases still excluded. Agree acceptance criteria with finance and operations before choosing the tool.

Kaliits can scope a document entry workflow around your existing systems. Bring an authorised document, the expected fields and the destination software to discuss your case. Describe your invoice OCR test.

All articles

Related reading