Reading a quotation PDF successfully is not the same as producing data that is safe to import into ERP. Before connecting an automated workflow, test critical fields, difficult documents and the cases that require human approval. The aim is to prevent a small extraction error from becoming an incorrect transaction.

Start with a test set and independently checked answers
Include quotations from different suppliers: text-based PDFs, scanned pages, multi-page tables and revised versions. Have a business reviewer establish the correct values from the originals before comparing the AI output. Do not use the model's own output as the answer key.
A small pilot could begin with 20 documents. That is a suggested starting point, not a sample size that proves performance across the whole business. Add missing document types before drawing broader conclusions.
Six failure cases worth testing
- Wrong units: cases, boxes, individual items or labour hours. In a hypothetical example, VND 120,000 for a box of 12 must not silently become VND 120,000 per item. Apply conversions only through an approved business rule.
- Wrong currency or number format: do not infer VND or USD from a supplier's location. Check currency labels, separators and a separate currency field.
- Missing or shifted line items: test page breaks, repeated table headings and descriptions spanning several lines. Compare line counts with the original, not only the total.
- Missing information: flag an empty field for review. Do not invent payment terms or quotation validity dates.
- Wrong version or duplicate import: a quotation number can have a revised document. Retain document identifiers and decide how revisions are handled before creating transactions.
- Failed reconciliation: compare quantities, unit prices, discounts and totals shown in the source. A business reviewer should resolve differences; the workflow should not change numbers merely to make them balance.
A high confidence score is not a business approval
Google Cloud's Document AI evaluation guidance distinguishes precision, recall and confidence thresholds. A confidence score is not a guarantee that a particular transaction is correct. Review results by field and by the impact of a potential error.
For fields that can change the supplier or payable amount, route early pilot results to an approval queue. Automate subsequent steps only after defining acceptance criteria, exception records and a way to stop invalid data from progressing.
Keep an audit trail that supports controlled retries
- The source document and extractor version.
- The extracted value, reviewer-confirmed value and reason for any change.
- Validation results, approver and target transaction identifier.
- Retry status so resubmitting a request does not create another transaction.
Frequently asked questions
Should AI connect directly to the live ERP at the start?
Begin with draft records or a test environment and a reviewer. A successful demonstration on a few documents is not evidence that every control can be removed.
Can the pilot use real quotations?
Use only data authorized for processing in an approved tool. Remove or mask information that is unnecessary for the test and restrict access to the sample set.
Planning a document-extraction pilot? Read the Google Drive handover checklist, explore Viet Nis services or request a consultation. Start with document types, expected volume and the destination system; do not send sensitive documents through a public form. Call +84 888 728 730.


