Can you translate contracts with AI? Short answer — not alone, but yes as part of a workflow. This post catalogs where AI breaks on contracts and gives a practical three-stage process that captures most of the cost savings without raising legal risk.
Five failure modes AI hits on contracts
1. Proper nouns rewritten on the fly
Official company names, product names, and people's names have a registered English spelling. AI happily invents plausible alternatives.
Example: well-known names like Itochu Corporation are fine, but mid-sized company names get rendered as Ito-Chu Co., Ltd. and similar.
Fix: Pass every proper noun through a glossary up front. Pull from your commercial registration, product naming guidelines, and passport transliterations for signatories.
2. Number and date formats drift
「令和8年4月1日」can land as April 1, 2026,1 April 2026, or 2026-04-01 in different places. Worse, "within 30 days of signing" can be ambiguous about whether the signing day counts.
Fix: Lock the date format and the counting rule in a template before translation.
3. Cross-references break
Section references like "per Article 3" do not automatically re-target if article numbering shifts during translation. AI translates the content, not the reference graph.
Fix: Review cross-references after translation. For long contracts, switch to stable IDs (Section 3.1) rather than raw numbering.
4. Legal terminology that sounds close but isn't
- 善管注意義務 →
duty of careis close but not a 1:1 match with common-law concepts. Safer:duty of due care of a prudent manager. - みなす vs 推定する → must stay distinct in English (
be deemedvsbe presumed). AI tends to collapse both intobe considered. - 不可抗力 / force majeure — the English term is standard but the Japanese legal definition and the governing-law interpretation can differ. Keep the original definition clause.
Fix: Maintain a legal glossary and feed it to the translator (human or AI) as non-negotiable.
5. Dropped subjects get guessed wrong
Japanese contracts omit subjects frequently and rely on 甲/乙 markers. AI occasionally swaps them or confuses subject and object.
Fix: Make 甲/乙 explicit in the source before translating. After translation, spot-check every clause for subject consistency.
Three-stage workflow that keeps legal risk down
Stage 1 — Prep
- Classify contracts (NDA, services, sale, employment, license, etc.)
- Build an approved glossary per category.
- Assign a named legal reviewer.
Stage 2 — AI draft
- Feed glossary + template into the AI.
- Review output as a parallel corpus (source ↔ target, clause by clause).
- Auto-flag low-confidence spans — numbers, proper nouns, references.
Stage 3 — Human review
- Legal reviewer confirms clause-level equivalence.
- Checklist sweep: proper nouns, numbers, dates, cross-references.
- Governing law + dispute resolution clauses get extra scrutiny — legal effect must match intent.
Net effect
- 60–75% fewer person-hours — editing beats writing from scratch.
- Consistency — glossary + template prevents drift across the document.
- Auditability — the parallel corpus is its own audit trail.
Contract types where AI alone should never be the final copy
- M&A docs (DAs, SPAs, share transfer agreements)
- Employment contracts (regulatory freshness matters)
- Patent licenses (mixed technical + legal terminology)
- Cross-border contracts with international arbitration clauses
These require a qualified legal professional on the final version, full stop.
Summary
AI is strong at drafting contracts and removing hours from the review loop. It is not a substitute for a legal reviewer on the final copy. With a glossary, a template, and a named reviewer, you can cut translation cost substantially without adding legal risk.
BizHonyaku is built around this workflow: glossaries, parallel view, diffs, and reviewer assignment in one place, so legal ops can adopt the hybrid model out of the box.