Fields, not free text
Invoice number, issue date, supplier (取引先), line items, subtotal (小計), consumption tax (消費税), and total (合計) — extracted as structured fields with confidence scores.
Solution · Document intelligence
Japanese invoice AI-OCR, built as a production system.
BCAQI Labs builds Japanese invoice AI-OCR pipelines that turn 請求書, quotations, and delivery notes into validated, structured data. We combine OCR with an LLM and business rules to extract fields, normalize 和暦 (Japanese era) dates to ISO, validate the 10% consumption tax, and export JSON/CSV — with confidence flags and a human-review queue for exceptions. It is the class of work a general chat assistant cannot do reliably.
What the pipeline extracts and checks
Every field is typed, validated, and traceable — not just transcribed.
和暦 date normalization
令和・平成・昭和 dates converted to ISO 8601, including 元年 (first-year) handling and full-width digits. Read how we do it →
Business-rule validation
Consumption-tax rate and totals are cross-checked. Anything that fails — a wrong tax rate, a total that does not add up — is flagged, not silently accepted.
How Japanese document intelligence works, step by step
- IngestReceive the document
PDF or scanned image of an invoice, quotation, or form.
- ExtractOCR + LLM field extraction
A layout-aware model reads Japanese text and maps it to named fields.
- Normalize和暦 → ISO, numbers cleaned
Japanese era dates and full-width digits are standardized.
- ValidateBusiness rules applied
Consumption tax and totals cross-checked against expected values.
- RouteHuman-review queue for exceptions
Low-confidence or rule-failing fields wait for a person; nothing auto-posts on a failed rule.
- ExportStructured JSON / CSV + audit log
Clean records for your ERP, with a full trail of what happened.
Why LLM + rules beats plain OCR — and beats ChatGPT
Plain OCR returns characters; it has no idea whether ¥48,200 is the tax or the total, or whether the numbers are consistent. A general assistant like ChatGPT can read one document in a chat window, but it does not integrate with your accounting system, enforce your rules, keep an audit trail, or hold uncertain results for human review. A production pipeline does all four — which is what an enterprise finance team actually needs.
The result: fewer manual keystrokes, a defensible audit log for every posted record, and exceptions surfaced instead of hidden. You can see the exact behaviour in the live demo, running on sample Japanese invoices.
Questions & answers
Japanese invoice AI-OCR — FAQ
How is Japanese invoice AI-OCR different from ordinary OCR?
Ordinary OCR returns text; it does not know which number is the consumption tax or whether the total is correct. A document intelligence pipeline combines OCR with an LLM and business rules to output structured, validated fields — invoice number, issue date (with 和暦 normalized to ISO), supplier, subtotal, 10% consumption tax, and total — and flags anything that fails a rule for human review.
Can it read 和暦 (Japanese era) dates like 令和6年5月12日?
Yes. The pipeline normalizes 和暦 dates (令和, 平成, 昭和) to ISO 8601 — for example 令和6年5月12日 becomes 2024-05-12 — including 元年 (gannen) first-year handling and full-width digits.
Does it validate the Japanese consumption tax?
Yes. The pipeline cross-checks that tax equals the expected rate of the subtotal (for example 10%) and that subtotal plus tax equals the stated total. A mismatch — such as an 8% reduced-rate item — is flagged rather than silently accepted.
What formats can it output?
Structured JSON and CSV, ready to import into an ERP or accounting system. Every record carries confidence scores and an audit log; exceptions are held in a human-review queue before they are committed.
Offshore AI development for Japan → · White-label AI delivery for consultancies & SIs →
Run it on your invoices.
A 4-week PoC turns the demo into a measured result on your own documents.