An invoice arrives from a Thai supplier. The header is in Thai. The line items mix Thai product descriptions with English part numbers. The monetary amounts are in Thai baht with the Thai numeral system. The tax calculation follows Thai VAT rules. The PO reference is an alphanumeric code in the Latin alphabet.
This is a single document. In APAC enterprise operations, every batch of documents looks like this — a polyglot mix of languages, scripts, number systems, and cultural conventions.
Most document AI systems are built for English-first processing. They work well with invoices from American suppliers, British contracts, or Australian compliance documents. When they encounter the multilingual reality of APAC operations, accuracy drops and manual intervention rises.
The Language Challenge Isn't Language
The obvious challenge is character recognition across scripts — Latin, CJK (Chinese, Japanese, Korean), Thai, Arabic (for some APAC markets), Devanagari (for India). Modern OCR handles multiple scripts reasonably well.
The real challenges are less obvious:
Layout variation: Document layouts vary by culture. Japanese documents often use vertical text. Thai documents have different spacing conventions. Chinese documents may mix simplified and traditional characters in different sections. Arabic text runs right-to-left within documents that are otherwise left-to-right.
Name conventions: A Thai company name on a tax invoice follows different formatting rules than a Japanese company name on a 請求書 (invoice). The system needs to understand not just the characters, but the naming convention — where the company name starts and ends, what the legal entity designation looks like, and how to match it against master data.
Number and date formats: Is "03/04/24" March 4th or April 3rd? Is "令和6年" 2024? Is "๒๕๖๗" (Thai numerals for 2567, Buddhist Era) also 2024? APAC date formats are a minefield, and getting them wrong means matching the wrong PO, applying the wrong payment terms, or filing in the wrong reporting period.
Mixed-language documents: The hardest challenge isn't processing a document in one language. It's processing a document that uses three languages simultaneously — which is the norm in APAC trade documentation. A bill of lading might have English field labels, Chinese goods descriptions, Japanese consignee details, and Thai port names.
What We've Learned
After processing millions of documents across APAC languages, here are the patterns that matter:
Language Detection Must Be Granular
Don't classify the entire document as "Thai" or "Japanese." Classify at the field level. The vendor name might be in one language, the product descriptions in another, and the addresses in a third. Field-level language detection enables field-level processing rules.
Cultural Context Matters for Validation
An extraction system that pulls a number from a Japanese invoice needs to know that yen amounts don't have decimal places. A system processing Indonesian invoices needs to understand that NPWP (tax identification numbers) follow a specific 15-digit format. These aren't language rules — they're cultural and regulatory rules that are language-adjacent.
Confidence Scoring Is Critical
For multi-language processing, the system must know what it doesn't know. A confidence score on each extracted field tells downstream processes which values to trust and which to flag for human review. In our deployments, we typically see 95%+ confidence on standardised fields (amounts, dates) and 85-92% on free-text fields (descriptions, addresses) — with the lower-confidence fields routed to human verification.
Training Data Must Be Representative
A document intelligence system trained primarily on English documents will underperform on Thai documents, even after fine-tuning. The training data must include representative examples from each language and document type that the system will encounter in production.
The Competitive Advantage
For enterprises operating across APAC, multi-language document processing capability is a genuine competitive advantage. It means faster supplier onboarding across markets, faster compliance processing, and fewer errors in cross-border operations.
The companies that build this capability aren't just processing documents faster. They're removing a fundamental barrier to regional scaling.