Selling into three countries used to mean hiring in three languages or accepting that two of those markets got a worse experience. The usual compromise was English-only support and a quiet acceptance of higher churn abroad.
Language models changed the economics, but not evenly. Some parts are genuinely solved and some still need a human. Knowing which is which is what separates a working multilingual operation from an embarrassing one.
Key takeaways
- Detection should key on message content, not phone prefix. Prefixes are wrong constantly.
- Tone drifts across languages unless you specify it per language.
- Product names, legal text and idiom are where machine translation fails hardest.
- A native speaker per market beats a native speaker per language.
Detect from content, not the number
The most common design error is inferring language from the phone prefix. A +49 number is not necessarily a German speaker, travellers carry foreign SIMs, expat populations are large, and in multilingual countries the prefix tells you nothing at all.
Detect from the message. Then, once the customer replies in a given language, persist it for that conversation rather than re-detecting every turn, since short messages like "ok" are ambiguous everywhere.
Always allow an override. Some customers prefer being served in English even when writing in their own language, often for technical vocabulary. A simple "reply in English if you prefer" respected for the rest of the conversation covers this.
Tone drifts across languages
This is the failure teams do not anticipate. An agent configured as warm and informal in English can come across as abrupt in German or excessively formal in Spanish, because register maps differently in each language.
The formal/informal distinction is the clearest case. German Sie versus du, French vous versus tu, Spanish usted versus tú carry social weight with no English equivalent. Defaulting to formal is safe but can feel cold for a consumer brand; defaulting to informal is friendly but risky in some markets and sectors.
Specify register per language rather than translating one tone instruction. Then have a native speaker read twenty real transcripts per language before launch. This catches in an hour what analytics will not surface for months.
Correct translation with wrong register still reads as a bad experience. Customers rarely report it. They just trust you less.
Where machine translation still fails
- Product names. Models translate names that should stay fixed. Pin them explicitly per language.
- Legal and policy text. Returns policies, warranty terms and privacy language need human translation, because a paraphrase can create obligations you did not intend.
- Idiom and humour. Brand voice built on wordplay does not survive. Write the source plainly.
- Units, dates and currency. Formats differ. This is a formatting rule, not a translation problem, and should be handled as one.
- Low-resource languages. Quality varies enormously. Test before assuming coverage.
Which languages justify a human
Not by speaker count. By revenue at risk and by complexity of what customers ask in that market.
| Signal | Automate | Hire a native speaker |
|---|---|---|
| Share of revenue | Under ~5% | Over ~15% |
| Question complexity | Routine, transactional | Consultative, high-value |
| Regulatory exposure | Low | Regulated sector |
| Escalation rate | In line with English | Persistently higher |
The last row is the most useful diagnostic. If one language escalates far more than the others, that is evidence the agent is underperforming there, not that those customers are harder. Investigate before adding headcount.
A native speaker per market generally beats one per language. Someone who knows the market handles the language plus the local expectations around delivery, returns and payment that no translation layer can supply.
Frequently asked questions
How many languages can one agent handle?
Technically many. Practically, limit it to markets you can review, since an unreviewed language is a risk you cannot see.
Should I translate my knowledge base?
Source content in one well-written language, translated at answer time, is usually better than several poorly maintained translations that drift apart.
What about right-to-left languages?
Arabic and Hebrew work well in messaging, but test rendering with mixed content, since embedded order numbers and links are where layout breaks.