How to Validate AI Invoice OCR: Chinese-English Fields, Tax and Duplicate Tests + Free AI Invoice Validation Checklist (Word)
AI can turn a PDF, scan or photo into supplier, invoice number, dates, currency, totals and line items. That does not make it safe to post automatically. OCR may read 90% of characters while missing the one critical decimal, currency, tax, negative amount, credit note—or entering the same invoice twice.
Proper acceptance has three steps: recognise text, extract fields and pass business rules. Google Document AI and Microsoft AI Builder return fields and confidence values, but high confidence is not proof and low confidence is only a routing signal. Combine model scores with arithmetic, vendor master data, purchase orders, receipts, prior invoices and human approval.
Define the schema before buying OCR
For each field define name, data type, required status, accepted format, normalisation, source location, low-confidence treatment and whether automatic write is allowed. Distinguish invoice date, due date and entry date; distinguish subtotal, tax, freight, discount, paid and amount due.
Four validation layers
Validation layer | Example | Automation level | Failure handling |
|---|---|---|---|
1 OCR | Pages, characters, orientation | Search or preview only | Rescan / manual read |
2 Field extraction | Supplier, date, total | Create draft | Low confidence to exception queue |
3 Business rules | Arithmetic, master data, PO, duplicates | Prepare for posting | Block and explain reason |
4 Controlled write | Permissions, approval, idempotency, logs | Automate only after pass | Rollback / human takeover |
Only the fourth layer should be considered for automated posting. Earlier layers can prepare a draft or exception. Payments, bank-detail changes, credit notes and material anomalies should keep stronger approval.
Chinese-English and Hong Kong realities
A supplier may use Chinese, English, abbreviation and group names. Do not rely on exact string equality. Hong Kong has no general VAT, yet local teams receive overseas VAT, GST, sales-tax, withholding or service invoices. Treat tax according to document and jurisdiction, not as zero by default. Retain original currency, FX source, FX date and posting currency.
Hong Kong’s Inland Revenue Department requires sufficient English or Chinese records of business income and expenditure, including invoices and receipts, generally for at least seven years. Automation should not delete the source and retain only extracted values. Link source, processed version, edits, approval and accounting entry.
How to test duplicates
Do not depend on filenames. Use a multi-signal fingerprint: vendor ID, invoice number, invoice date, total, currency, PO, bank details, file hash and similar line items. Exact matches may be blocked; partial matches belong in an exception queue. Test rescans, forwarded PDFs, renamed files, brightness changes and a later credit note.
Hong Kong planning budget and timeline
A pilot for a shared folder, fixed fields and hundreds of invoices per month with human review may start around HK$25,000–60,000 over four to six weeks. Vendor master, PO/receipt matching, duplicate detection, accounting-system writes, permissions and audit may start around HK$80,000–250,000+ over eight to sixteen weeks. OCR/cloud usage, accounting licences, migration, security and professional review are separate. These are Allua planning ranges, not fixed quotations.
Free resource: AI Invoice Validation Checklist
Check fields, totals, duplicates and review rules, with evidence for each validation. Get an editable Word version to adapt and reuse for your team.
Enter your email, choose how you’ll use the kit and confirm your consent to receive it by email.
Prepare normal, error and boundary samples for each item. Record source value, expected normalised value, actual value, confidence, rule result, exception decision and final approval. Example: “HKD 1,234.50 must not become HKD 123,450; a failed total equation must stop posting.”
01 Document type — do not confuse invoice, receipt, quote, credit note or statement.
02 Supplier identity — map Chinese, English, aliases and company number to master data.
03 Invoice number — preserve leading zeros, slashes, hyphens and case rules.
04 Invoice and due dates — do not swap; invalid dates become exceptions.
05 Currency — align symbol and ISO code; do not guess HKD, USD or RMB.
06 Subtotal, discount, freight, tax and total — recalculate with defined rounding tolerance.
07 Tax — identify VAT, GST or other tax by document and jurisdiction; do not assume a Hong Kong rate.
08 Line items — preserve description, quantity, price, discount, rate and amount relationships.
09 PO and receipt — match references; route overage, short receipt or no-PO cases.
10 Bank details — independently verify every new or changed instruction.
11 Duplicate invoice — catch rename, rescan, forward and mild image change.
12 Credit note and negative value — preserve direction and link the original invoice.
13 Low confidence or blank field — never invent a required fact; send for review.
14 Personal data and access — expose only necessary fields; control source, log and export.
15 Write and recovery — preview before posting; avoid duplicate writes; rebuild from source.
Frequently asked questions
Can 99% confidence post automatically?
Not by itself. Set field-specific thresholds and add arithmetic, master-data, PO, duplicate and anomaly rules. Bank, payment and high-value cases deserve stricter controls.
How many invoices should we test?
Coverage of suppliers, layouts, languages, image quality, currencies, tax and exceptions matters more than one sample count. Start with high-frequency suppliers, then long-tail and deliberately damaged examples.
Can the AI output replace the source invoice?
It should not. Keep the original and its traceable relationship so audit, tax, dispute and reprocessing retain evidence.
Validate before writing to accounting
Share de-identified samples, field list, vendor master, PO rules, accounting system and approval roles. Allua can build the acceptance set, exception queue, duplicate detection and safe posting controls. Contact Allua to plan your AI invoice workflow.
Creator partnership invitation
Creators interested in future Allua features or brand partnerships may submit the creator partnership form. Submission does not guarantee selection or collaboration.
Sources and disclaimer
References: Google Cloud Document AI Invoice Parser, evaluation and confidence guidance; Microsoft AI Builder Invoice Processing; Hong Kong IRD record-keeping guidance; Hong Kong PCPD AI guidance. This is not accounting, tax, legal or cybersecurity professional advice. Rules, fields and products change; obtain relevant professional review before production use. Actual cost and timing depend on quotation.
Comments