A business has an assistant working well in one language. Adding a second looks like a translation job and is mostly not.
The words are the straightforward part. The decisions are which facts differ in that market, what level of formality to use, which questions still need a person who speaks the language, and what happens when one is not available.
Businesses that treat it as translation end up with fluent, confident and occasionally wrong answers, which is worse than no second language at all.
Step 1: confirm the language is worth supporting
From traffic and enquiry data rather than ambition, which is the decision in deciding which languages to support first.
Step 2: list the questions actually asked
The real set, which is usually a few dozen, rather than everything on the website.
Step 3: identify which facts differ in that market
Payment, delivery, returns, tax, warranty and anything legally required. This is the part that matters and the part translation does not touch.
Step 4: record those facts against the language
So the answer given in that language contains the true local position rather than a fluent rendering of the domestic one.
Step 5: decide the register deliberately
Formal or informal, per context, written down, since this is where a technically correct answer can still read badly.
Step 6: keep brand and product terms untranslated
Consistently, so the same feature is called the same thing everywhere and search still works.
Step 7: have a native speaker review the core answers once
The set is finite. One review resolves almost every tone and phrasing problem permanently.
Step 8: decide what must reach a person
Complaints, negotiations and anything sensitive, with a named route to somebody who speaks the language or a clear stated response time.
Step 9: set up the handover summary
So whoever takes over gets a translated summary of the conversation rather than a raw thread they cannot read.
Step 10: set the response time honestly per language
If a person in that language is only available part of the week, say so rather than implying parity.
Step 11: test with real questions, not translated ones
Ask a native speaker to enquire as a customer would, in their own words, which surfaces gaps that a translated test script never will.
Step 12: keep one source of answers
Updated once and reflected in every language, rather than parallel sets that diverge within months.
Step 13: measure separately from the first language
Enquiries, conversion and what needed a person, which is the arithmetic in measuring whether a second language was worth it.
Step 14: tell visitors it is available
Somebody arriving on an English page has no way of knowing they could be answered in their own language. Detecting the browser language and offering to switch, once, converts people who would otherwise struggle through or leave.
Step 15: add one language properly rather than three badly
An assistant that answers in a second language with the correct local facts, the right register and a working handover serves that market, and one that translates the domestic answers fluently does not. Set it up through install the widget and train your bot, with the approach in multilingual chatbots how ai answers in your customers language and the installation detail in adding an assistant without breaking the theme. The sector view sits under small business.