Language switching

Customers can switch the chat language in the following ways, depending on the channel:

• Autodetect (Chat and Email only): If a customer sends a message with at least three meaningful words and 20 characters in an enabled language, your AI Agent will attempt to detect and switch to that language. Common words like “and” or “the” don’t count toward the three-word minimum.

• Language name (Chat and Email only): If a customer types the exact name of an enabled language (e.g., “French” or “Français”), your AI Agent will switch to that language. This only works with exact matches — typos or phrases like “Do you speak French?” don’t trigger a change.

• Chat settings (Chat only): In Ada’s web chat, customers can click the Settings icon, then select Change language to pick from the languages you’ve enabled in your AI Agent.

For information about how language switching works in Voice, see Enable Voice in multiple languages.

Why your AI Agent might not switch languages

Language detection depends on various factors. Here’s why switching may not happen as expected:

• Minimum character and word requirements for Latin-based languages

  • For messages written entirely in Latin characters, the message must contain at least 3 words (space-separated) and 20 characters to trigger a language switch
  • This requirement applies to all languages except: Greek, Hindi, Japanese, Khmer, Korean, Myanmar, Punjabi, Tamil, Thai, Simplified Chinese, and Traditional Chinese

• Switch-disabled language pairs

  • Some languages are considered too similar by our model, so switching between them is disabled to prevent premature language changes
  • These pairs include:
    • Malay and Indonesian
    • Simplified Chinese and Traditional Chinese
    • Spanish and Catalan
    • Croatian and Serbian

• Switch-disabled languages

  • Once set to these languages, the AI Agent cannot switch out of them through language detection:
    • Haitian Creole
    • Tagalog
    • Bosnian

• Non-Latin writing systems

  • For languages with non-Latin writing systems (like Hindi, Chinese, etc.), we do not support recognizing Latin-based versions of these languages
  • For example, Hindi written in Latin characters (like “kya hal-chal hai”) cannot be detected, while the same phrase in Devanagari script (क्या हाल-चाल है) can be recognized

• Diacritics requirement

  • For languages that use diacritics (like Vietnamese), the model requires proper diacritic usage
  • Messages without diacritics may be misidentified (e.g., “voi” without diacritics might be detected as Italian instead of Vietnamese)

• Low performance languages

  • The model has lower accuracy for these languages:
    • Afrikaans
    • Catalan
    • Croatian
    • Haitian Creole
    • Malay
    • Norwegian
    • Slovenian
    • Urdu

• Unicode-based detection

  • For certain languages with unique writing systems, detection is based on Unicode rather than the model
  • This applies when the entire message uses only one of these writing systems:
    • Greek
    • Devanagari
    • Hiragana
    • Katakana
    • Khmer
    • Hangul
    • Myanmar
    • Gurmukhi
    • Tamil
    • Thai

• Denied words

  • Messages containing certain common words across multiple languages will not trigger a switch
  • These words include: “no”, “meta”, “variable”, “swish”, “support”, “agent”, “face”, “id”, “selfie”, “bet”, “confirmation”, “start”, “vacation”, “hold”, “nya”, “fuel”, “point”, “domain”