Auto-translation — Multilingual conversations
When a visitor writes in a language other than the one the team understands, SilentChat automatically translates the messages via AI. Originals are never overwritten — both agent and visitor can switch between translation and original at any time.
Activation per Conversation
Auto-Translation is activated per conversation individually (Opt-In). This way, it only costs tokens for the chats where it is actually needed.
Inbox → Open conversation → Sidebar „Conversation details" → Switch „Auto-Translation"
Once activated:
- The banner above the history shows „Conversation in {language} — automatically translated"
- Existing visitor and bot messages are translated retroactively (Auto-Backfill in 2–3 seconds)
- New messages are translated live as they arrive — both to the visitor (agent responses) and to the agent (visitor messages + bot responses)
Language Detection
The source language of the conversation is automatically detected with the first visitor contribution and displayed in the sidebar under „Detected source language" (e.g. EN, FR, ES).
If it says UNKNOWN, the detection could not be applied — usually because the first message was too short (e.g. „hi"). As soon as the visitor sends a complete sentence, the detection is made up with the next bot reply and the backfill then runs automatically.
Original Availability
This is a hard guarantee in SilentChat:
messages.contentalways stores the original text — even after translation- Under the eye icon in each bubble, there is a toggle „Show original / Show translation"
- The direction indicator (e.g.
EN → DE) clearly shows which translation is currently rendered - Audits, compliance reviews, and manual error checking always have access to the unchanged original text
Who Sees What?
| Sender | Visitor sees | Agent sees |
|---|---|---|
| Visitor (EN) | Their own EN (original) | DE (translation) |
| Bot (responds in visitor language, EN) | EN (original) | DE (translation) |
| Agent (DE) | EN (translation) | Their own DE (original) |
| System (Handoff banner etc.) | Visitor locale (next-intl) | Agent locale (next-intl) |
System messages (e.g. „An employee has taken over the conversation") use the fixed UI translation — no LLM call needed, no tokens consumed.
Manually Override Source Language
If the automatic detection is incorrect, the source language can be manually set in the sidebar under „Detected source language". Subsequent messages will then use this language as the basis for translation.
Token Consumption
Auto-Translation is billed against the AI monthly budget of the plan — as its own feature translation plus lang_detect, visible separately from the chat and copilot budget.
Order of magnitude per average conversation (5 visitor + 5 bot + 1 detection, one target language, a small IONOS model):
- ~4,000 tokens in total
- Cache is used with each refresh — no double billing
What increases consumption:
- More team members with different languages (each locale = one translate call per bubble)
- Very long bot responses (RAG with many sources)
- Frequent activation in non-multilingual conversations
Safety net:
- Plan limit check before each translation — no unwanted overruns
- In case of errors (e.g. provider timeout), the original is rendered instead of showing an empty bubble
- Super admins see the error in the hover tooltip under the eye icon
Availability by Plan
Auto-Translation is a Pro Plan Feature. Trial and Starter rates do not see the switch. The manual backfill method translate-now is also plan-gated via the ai_chatbot feature, so that free tenants cannot abuse the endpoint.
GDPR / Data Protection
For translation, the message text is sent to the configured AI provider — the same one that already operates the AI chatbot. Therefore:
- No additional sub-processors — anyone who already lists the AI chatbot in the privacy policy is also covered for Auto-Translation
- EU hosting with provider IONOS AI Hub possible (Plan #106), for US providers the standard DPA of the respective provider applies
- Visitor consent for AI chat is sufficient — no separate consent flow for translation
Frequently Asked Questions
What happens if the translation goes wrong? Silent fallback to the original with a subtle subtitle hint. Super admins see the error message as a tooltip under the affected bubble.
Are bot responses written twice? No. The bot responds directly in the visitor language (e.g. EN for EN visitor) and is only translated for other team locales (e.g. DE for German agents). This results in exactly one LLM call per bubble + target language.
Does the bot change language in the middle of the conversation? No. The source language is detected with the first visitor contribution and maintained for the entire conversation. If the visitor changes the language, the sidebar selection must be manually updated.
Can I turn off Auto-Translation in the widget even though it is activated? The visitor sees the same toggle „Show original / Show translation" in the widget as the agent — per bubble individually. The global activation of the conversation only controls whether translations are generated at all.