Translating English business content into Japanese is a quality test for any AI translator — and keigo (Japanese honorific language) is where they most often fail. This post catalogs the patterns AI gets wrong, why, and the workflow that catches them before delivery. Aimed at international teams sending Japanese communication to customers, partners, and Japanese staff.
Why keigo is hard to translate
1. It depends on relationships, not just words
The English sentence You should review this has at least three correct Japanese translations depending on who is speaking to whom:
- Manager → report: 「確認してください」 (polite, neutral)
- Report → manager: 「ご確認いただけますでしょうか」 (humble + polite)
- To external client: 「ご確認のほどよろしくお願いいたします」 (highest formality)
Without context, AI defaults to mid-tier "polite form" — too weak for external comms, too formal for internal ones.
2. Verbs, nouns, and prefixes all change together
Keigo isn't a per-word swap; it's a sentence-wide transformation:
- Verb: 見る → ご覧になる (sonkeigo) / 拝見する (kenjogo)
- Noun: 意見 → ご意見 (prefix)
- Receiving verbs: もらう → いただく → 賜る
- Auxiliary verbs: 〜してくれる → 〜してくださる → 〜していただく
AI knows each piece individually but struggles to combine them consistently within one sentence. Result: awkward hybrids like 「ご確認頂戴いたします」.
3. Cultural register matters
Translating Please confirm directly as 「確認してください」 often reads as commanding in Japanese business culture. The cultural-correct version uses cushion language: 「ご確認のほどお願い申し上げます」. That choice is cultural, not grammatical, and AI doesn't always make it.
Five common AI keigo mistakes
1. Double keigo
Wrong:
お見えになられる→ just 「お見えになる」 is correctご拝見いたしました→ just 「拝見いたしました」 is correctおっしゃられる→ just 「おっしゃる」 is correct
AI tries to be polite by stacking honorifics. Meaning gets through but formal Japanese style guides flag this.
2. Confusing sonkeigo with kenjogo
Sonkeigo (respectful) for the other person, kenjogo (humble) for yourself / your side. Common slip:
- Wrong: 「弊社の田中がいらっしゃいます」 (sonkeigo on own side)
- Right: 「弊社の田中が伺います」 (kenjogo on own side)
3. Wrong formality level for the audience
Without telling the AI who the message is for, output lands at mid-tier — wrong on both ends:
- Internal memo with 「謹啓 〜のほどお願い申し上げます」 (overshoot)
- Client email with 「お手数ですが見てください」 (undershoot)
Look for a tool that lets you specify the scene(internal / external / partner / customer).
4. Industry-specific phrasing
IR, medical, and legal Japanese have their own honorific vocabulary:
- IR: 「ご高承のとおり」「平素より格別のお引き立てを賜り」
- Medical: 「ご加療」「ご来院」「ご清祥のこととお慶び申し上げます」
- Legal: 「ご通知申し上げます」「謹んでお伝え申し上げます」
Generic AI output won't have these. Lock them in via a glossary.
5. Direct imperative tone
Casual English imperatives like Send me the file become flat 「ファイルを送ってください」 — but Japanese business comms expect cushion language: 「お手数ですが」「恐れ入りますが」「可能でしたら」. Plain imperatives feel curt.
Four checks to keep keigo translation quality high
1. Provide context up front
- Sender / receiver relationship (internal, external, partner, customer)
- Document type (info, request, apology, announcement, report)
- Industry (finance, medical, legal — for register-specific phrases)
2. Pin a single formality level
Decide on one register for the entire document — don't let the AI drift between polite and formal. Example: "translate this whole document in formal partner-facing keigo."
3. Native review on important docs
AI can be grammatically correct but produce phrases native speakers wouldn't write. Build native review into the workflow for anything externally facing.
4. Lock honorific phrases via glossary
Company-specific honorific terms (titles, signatures, opening salutations) belong in a glossary so they never drift.
How BizHonyaku handles keigo
- Scene selector: internal / external / partner / customer
- Industry glossary presets: IR, legal, medical, HR
- Double-keigo detection: warns on common stacking errors
- Review workflow: native checks slot in cleanly
Summary
Keigo is a problem of relationships and context, not grammar. AI knows the parts but assembling them correctly for the scene still benefits from human review. Combine context-specifying, single-register policy, glossary, and native QA, and you can capture most of AI's speed gains without the awkwardness.
Start with a one-page preview on a real document of yours to see how keigo turns out before committing to volume.