Writing knowledge for an AI agent that serves customers in many languages: keep one source language, add the terms that do not translate, avoid idiom, and test retrieval in the languages you expect
The instinct is to translate the knowledge base into every language your customers use. The result is three versions that drift apart and an agent that quotes last month's price in Spanish. One plainly written source, plus the handful of terms that do not translate, serves every language better and is one thing to maintain.
Quick answer
Write the knowledge base once, in plain English (or your main language), with one fact per line and no idiom; add a short glossary of terms that must not be translated or that customers use in their own language (your service names, dish names, place names, local payment methods, sizes and units); set formality rules per language in the instructions; and test retrieval by asking the agent your top ten questions in each language you expect. Keep a translated copy only for a language with heavy volume and legal wording that must be exact. ReplyKit replies in the customer's language from that single source.
Plain source language
- One fact per line. "Gel manicure: $45, 60 minutes, includes removal." Translates cleanly and retrieves cleanly.
- No idiom. "Walk-ins welcome" becomes "You can come without an appointment." "Prices are a steal" becomes a number.
- No ambiguity. "Open late" becomes "Open until 9 pm." "Weekends" becomes "Saturday and Sunday."
- Named units and currencies. "cm", "kg", "USD", every time.
- Dates and times with zones, as in multiple time zones.
The glossary
A short section the agent reads: terms to keep untranslated (your brand, product and service names), terms customers use in their language that map to yours ("corte de pelo = haircut"), place names in both forms, local payment methods by name, and sizing conventions ("EU 38 = US 7.5"). This is the single biggest improvement for retrieval in other languages, because a question containing "corte" now matches the line about haircuts. The retrieval mechanics are in how retrieval works.
Formality and tone per language
- In the instructions: "Reply in the customer's language. Use formal address in German, Japanese and Korean unless the customer is informal first. Match Spanish and Portuguese to the customer's register."
- Keep the persona rules language-neutral (short, warm, one question at a time) and let the agent carry them across; the persona work is in persona and tone.
- Ban phrases that translate badly; emoji rules carry across languages.
Testing in each language
Ask the ten most common questions in each language you expect, in the test chat, and check three things: the answer is correct against the knowledge, the language is right, and the agent did not say "not sure" for something the knowledge covers. A "not sure" in Spanish for a question the English knowledge answers usually means a missing glossary term; add it. Retest after changes. The general routine is in testing before going live.
When a translated copy is worth it
Only when a language carries heavy volume and the wording is legally sensitive (terms, medical disclaimers, financial disclosures). Then keep that section translated by a person, clearly labelled with its language, and update both at once. For everything else, one source and a glossary wins. The customer-facing side of this is in enquiries in a language you do not speak.
Frequently asked questions
Should I translate my knowledge base into every language?
No. Write one plain source and a glossary of untranslatable terms; the agent replies in each language from it.
What goes in the glossary?
Brand, product and service names, customer terms in their language mapped to yours, place names, local payment methods, and sizing or unit conventions.
How do I set formality per language?
One line in the instructions naming the languages that need formal address by default.
How do I know retrieval works in another language?
Ask your top ten questions in that language in the test chat; a wrong "not sure" usually means a missing glossary term.
When is a translated copy justified?
For a high-volume language with legally sensitive wording, translated by a person and kept in sync.
One source, every language
Plain knowledge, a glossary, and an agent that replies in the customer's language. 7-day free trial, no card.
Sources and further reading
- Harvard Business Review, The Short Life of Online Sales Leads (Oldroyd, McElheran, Elkington, 2011) · firms that responded within an hour were about seven times more likely to qualify the lead
- WhatsApp Business Messaging Policy · opt-in, opt-out and prohibited-use rules every business must follow
ReplyKit is an independent product and is not affiliated with or endorsed by Meta or WhatsApp. This article is general guidance for running a small business, not legal, financial or regulatory advice.